Interpretable adaptive artificial intelligence network

Abstract

Computer-implemented methods and systems make a generative AI system more interpretable. A programmed computer system grows a generative AI system by adding one or more interpretable network elements to the generative AI system. Each interpretable network element may be trained to identify two or more sets of interpretable training data items of the generative AI system. After the one or more interpretable network elements are added, training of the generative AI system may be updated using the added one or more interpretable network elements. The programmed computer system may then determine whether sustainable development of the generative AI system is required.

Description

[0001] Cross-referencing related applications

[0002] This patent application claims priority to the following U.S. provisional patent applications, the entire contents of which are incorporated herein by reference: (1) filed July 28, 2023, entitled “Explainable Adaptable Artificial Intelligence Networks”, serial number 63 / 529,563, and (2) filed September 11, 2023, entitled “Explainable Adaptable Artificial Intelligence Networks”, serial number 63 / 537,671.

[0003] Related applications

[0004] This application relates to: U.S. Provisional Patent Application No. 63 / 481,697, filed January 26, 2023, entitled “Training Dynamic Hybrid AI Networks”; International Application No. PCT / US24 / 12671, filed January 24, 2024, entitled “Training Dynamic Hybrid AI Networks”; U.S. Provisional Patent Application No. 63 / 468,145, filed May 22, 2023, entitled “Training Human-Guided Hybrid AI Networks”; and International Application No. PCT / US24 / 30324, filed May 21, 2024, entitled “Training Human-Guided AI Networks”. Background Technology

[0005] Deep neural networks have achieved remarkable success in recent years. However, some fundamental problems remain, such as sensitivity to small adversarial perturbations in data and the difficulty of interpreting the internal nodes of large networks. Sensitivity to small adversarial perturbations can cause deep neural network classifiers to make mistakes that no intelligent entity would make. The difficulty of fully interpreting internal nodes within context may prevent people from fully trusting the decisions and actions of AI systems based on such networks. As society becomes increasingly reliant on AI systems using deep neural networks, the dangers posed by these problems could become very serious.

[0006] While deep learning, using large deep neural networks, is one of the most successful techniques in artificial intelligence, the scale and complexity of such networks can make understanding their inner workings and detecting and diagnosing any problems extremely difficult. Furthermore, the design and training techniques of neural networks make large networks prone to errors that no sensible human would make. Generally, deep neural networks are trained using a process called gradient descent, in which, for each training data item, the computer system applies the chain rule of calculus to backpropagate the derivative of the target, such as a divergence measure to penalize the error. Adversarial attacks can exploit gradient descent to uncover tiny adversarial perturbations that cause deep neural network classifiers to err. Designing a network trained via gradient descent makes it vulnerable to gradient descent-based adversarial attacks and other sources of tiny perturbations.

[0007] Errors caused by such minute perturbations are an example of a system's lack of sensitivity. That is, the system might make a mistake a wise human wouldn't make. More generally, deep neural networks lack common sense. Furthermore, the complexity of large neural networks makes it difficult for humans to understand the details of the training process, let alone contribute common sense. As AI systems become more capable and take on more and more tasks, this lack of common sense will become an increasingly greater danger. Once AI systems take over the task of designing the next generation of AI systems without human understanding and control, introducing sensitivity and common sense will become increasingly difficult. As AI systems control more aspects of human life, the consequences of errors could be catastrophic.

[0008] The difficulty in understanding the internal nodes of neural networks stems primarily from the sheer size and depth of these networks, and the fact that training routines do not guide the representation of internal nodes in a way that is expressible in human language. To a large extent, a lack of sensitivity and overall interpretability is a consequence of the methods used to train deep neural networks. Summary of the Invention

[0009] In a general aspect, this invention proposes the concept of a dynamic hybrid network, a generalization of the neural network concept. The hybrid training approach provides an alternative to training networks solely through gradient descent. The architecture of the hybrid network includes new elements called units and cells, as well as neural network nodes. The training techniques for dynamic hybrid networks support training architectures that are robust to perturbations in the input data. The system supports methods for training multiple elements, such as piecewise constant activation functions, including linear threshold functions. Training supports incremental growth of the network and continuous training during deployment. The configuration of the hybrid network is dynamic and can be changed and customized upon receiving specific input data items. Techniques are included to train the system to avoid classification errors that violate sensitivity, including errors caused by adversarial attacks. The hybrid model and training techniques also help interpret internal elements within the context of surrounding elements and other parts of the network. The system supports supervised training processes through collaboration between human teams and one or more AI systems trained in supervised hybrid network training.

[0010] In another general aspect, the present invention relates to a system and method for adaptively adjusting a large language model (LLM) using human input. The method may include the steps of: (a) generating a plurality of training pairs for a first image generator, wherein each of the plurality of training pairs includes (i) an image and (ii) a corresponding description of the image, wherein generating the plurality of training pairs includes generating the plurality of training pairs using a second image generator; (b) training the first image generator using the plurality of training pairs by a programmed computer system; (c) generating a first detailed description of an image to be generated by the first image generator based on a first cues received from a human by the LLM; (d) generating a first image by the first image generator based on the first detailed description of the image generated by the LLM; (e) receiving a first edit of the first detailed description based on human review of the first image by a programmed computer system; and (f) training the LLM with the first edit as training data by the programmed computer system. Steps (c) through (f) may be repeated multiple times to train the LLM.

[0011] In another general aspect, the present invention relates to computer-implemented systems and methods for generating textual works. In various embodiments, the method may include the step of generating an outline of a textual work based on topical prompts received by a computer system including an LLM, wherein the outline comprises N subtopics, where N > 2. The method may further include: for each of the n=1, ..., n subtopics, the computer system iteratively generates a text paragraph for the nth subtopic using the LLM; solicits user feedback from users of the text paragraph for the nth subtopic; updates the text paragraph for the nth subtopic based on the user feedback (if any); and adaptively trains the LLM using the user feedback (if any).

[0012] In another general aspect, the present invention relates to computer-implemented systems and methods for training target nodes in a neural network to make them more interpretable. In various embodiments, the method includes the step of adding additional nodes to the neural network by a programmed computer system, wherein adding additional nodes includes: initializing the additional nodes to have the same connections and weights as the target node, and wherein the additional nodes are associated with a first designated set of data items. The method also includes the step of training the neural network with the added additional nodes by a programmed computer system, wherein training the neural network includes: applying regularization to the additional nodes to train them to have activation values ​​for each data item in the first designated set, the activation values ​​being better aligned with the data items that are members of the first set. The method also includes the step of creating at least three new test neural networks by a programmed computer system, wherein: at least one of the three new test neural networks includes the target node but not the additional nodes; at least one of the three new test neural networks includes the additional nodes but not the target node; and at least one of the three new test neural networks includes both the target node and the additional nodes. The method further includes the following steps: a programmed computer system computes a regression of the measured performance of each of at least three new test neural networks, depending on whether each new neural network includes (i) a target node but not an additional node; (ii) an additional node but not a target node; and (iii) both a target node and an additional node. The method also includes a step of creating a new neural network by the programmed computer system based on the regression, wherein creating a new neural network includes: determining, based on the regression, whether to include (i) a target node but not an additional node; (ii) an additional node but not a target node; and (iii) both a target node and an additional node in the new neural network. The method further includes a step of training the new neural network by the programmed computer system.

[0013] In another general aspect, the present invention relates to systems and methods for computer implementations of improving the interpretability of neural networks. The neural network may include attention block output nodes, wherein the attention block output nodes compute a weighted correlation between two n-tuples. In various embodiments, the method includes the step of replacing the attention block output nodes with multi-node units by a programmed computer system, wherein the multi-node units include: first and second product nodes, wherein the first product node computes the product of values ​​in a first n-tuple, and the second product node computes the product of values ​​in a second n-tuple; and a summation node that computes a weighted sum of the outputs from the first and second product nodes. The method also includes the step of training the neural network by a programmed computer system using the multi-node units.

[0014] Another general aspect of the present invention relates to a chained sequence of mappings. Suppose a chained sequence has multiple mappings, n = 1, …, n, such that for n < N, the output representation of the nth mapping is the input representation of the (n + 1)th mapping; the output representation of the n > 2 mappings is the same as the output representation of the first mapping; and the input representation of the n < N mappings is the same as the output representation of the Nth mapping. In various embodiments, the method includes the step of training, by a programmed computer system through machine learning, an ensemble of autoencoders, the ensemble of autoencoders including one or more autoencoders, the goal being to generate an example of the output representation of the Nth mapping for an input that is an instance of the input representation of the first mapping. The ensemble of autoencoders includes at least one latent space containing text, and training the ensemble of autoencoders can include receiving edits to the text of at least one latent space containing text from a human and training the ensemble of autoencoders using the edits from the human.

[0015] In another general aspect, the present invention relates to computer-implemented systems and methods for detecting text generation by an LLM. According to various embodiments, the method can include the step of training, by a programmed computer system through machine learning, an LLM to generate text paragraphs with low-probability language units. The method also includes the step of training, by a programmed computer system through machine learning, a detector to detect low-probability language units in input text. The method also includes, after training the detector, the step of detecting, with the detector, text generated by the LLM that includes one or more low-probability language units.

[0016] In another general aspect, the present invention relates to computer-implemented systems and methods for training a set of one or more nodes as a named set discriminator and for training and using a related confidence estimator. In various embodiments, the method includes selecting, by a programmed computer system, a pair of sets of known data items associated with selected nodes of a neural network as a pair of sets to be discriminated. The method also includes the step of creating, by a programmed computer system, a new node for the neural network, where the new node is for discriminating the pair of sets to be discriminated, and where creating the new node for the neural network includes connecting the new node to other nodes in the neural network and training the new node in a different manner from the selected nodes. The method also includes training, by a programmed computer system, a confidence scoring network for each selected node and the new node. The method also includes the step of generating, by a programmed computer system, a single output value from the output values of the selected nodes and the new node, where generating the single output value includes generating the single output value according to a combination rule of the selected nodes and the new node, where the combination rule is selected based on confidence scores from the confidence scoring network. The method also includes training, by a programmed computer system, a network to calculate the single output value.

[0017] In another general aspect, the present invention relates to computer-implemented systems and methods for making generative AI systems (such as LLMs) more interpretable. A method according to such an embodiment may include the steps of: (a) developing a generative AI system by adding one or more interpretable network elements to the generative AI system by a programmed computer system. The generative AI system may include one or more trained machine learning networks such that the generative AI system generates text paragraphs in response to cues and context; and each of the one or more interpretable network elements may be trained to identify two or more sets of interpretable training data items for the generative AI system. The method may further include, in step (b), performing update training on the generative AI system with the added one or more interpretable network elements by the programmed computer system after adding the one or more interpretable network elements. The method may further include, in step (c), after steps (a) and (b), determining by the programmed computer system whether continued development of the generative AI system is required. When step (c) determines that continued development of the generative AI system is required, the programmed computer system may repeat steps (a) through (c). When it is determined in step (c) that the continued development of the generative AI system is not required, the generative AI system can be deployed to generate text paragraphs (e.g., reasoning, i.e., using a trained generative AI model to predict or generate output based on new, unseen cues and context).

[0018] These and other benefits that can be achieved through embodiments of the present invention will become apparent from the following description. Attached Figure Description

[0019] Various embodiments of the present invention will be described with reference to the following figures.

[0020] Figure 1 This is a flowchart illustrating an embodiment of the present invention.

[0021] Figure 2 This is a flowchart illustrating an embodiment of a process for enhancing basic sensitivity in one aspect of the present invention.

[0022] Figure 3A This is a schematic diagram of the hybrid unit in an exemplary embodiment of the present invention.

[0023] Figure 3B This is a schematic diagram of one aspect of the present invention, referred to as active defense.

[0024] Figure 3C This is a schematic diagram of the alternative derivative function used in one aspect of the present invention.

[0025] Figure 4 This is a schematic diagram of the technical hierarchy for improving sensitivity.

[0026] Figure 5 This is a schematic diagram illustrating various aspects of a hybrid training approach organized according to the stages of the training process.

[0027] Figure 6 This is a flowchart illustrating an example of constrained optimization during training.

[0028] Figure 7 This is a flowchart illustrating an embodiment of hidden state space modeling in one aspect of the present invention.

[0029] Figure 8 This is a flowchart of an exemplary embodiment of using a trained hybrid network and fast matching to perform reasonable classification operations.

[0030] Figure 9 This is a schematic diagram of an automatic encoder used in one aspect of the present invention.

[0031] Figure 10 This is a schematic diagram illustrating an embodiment of a robust template model used in one aspect of the present invention.

[0032] Figure 11 A flowchart illustrating exemplary embodiments of training data exclusion and data delegation in various aspects of the present invention.

[0033] Figure 12 This is a flowchart illustrating an embodiment of training an alignment model in one aspect of the present invention.

[0034] Figure 13 This is a flowchart illustrating an embodiment of one aspect of the present invention, referred to as "conditional mixed training".

[0035] Figure 14 This is a schematic diagram illustrating one aspect of the present invention for data space conversion or transformation.

[0036] Figure 15 This is a flowchart illustrating an embodiment of one aspect of the invention, which uses regression of counts in a histogram warehouse.

[0037] Figure 16 This is a schematic diagram of a hybrid network of cells and units.

[0038] Figure 17 This is a schematic diagram of a multiprocessor computer system that can be used to implement various aspects of the present invention.

[0039] Figure 18 This is a flowchart illustrating an example of backpropagation of data in one aspect of the present invention.

[0040] Figure 19 This is a flowchart illustrating an exemplary embodiment of parallel or serial computation in a cell network connected by data communication links.

[0041] Figure 20 This is a flowchart illustrating an example of experiential training.

[0042] Figure 21 These are schematic diagrams illustrating various aspects of the present invention, including one or more hybrid network artificial intelligence systems implemented on computer system 1700 collaborating with one or more human teams on joint tasks.

[0043] Figure 21A This is a schematic diagram of a multilayer feedforward neural network.

[0044] Figure 22 This is a flowchart illustrating an example of training and using an image generation system with human guidance.

[0045] Figure 22A According to various embodiments of the present invention, it is used for Figure 22 The system block diagram in the method.

[0046] Figure 23 This is a flowchart illustrating an example of the process of building and training an interactive, human-guided writer assistant.

[0047] Figure 24 This is a flowchart illustrating an example of a process for training selected nodes to make them more interpretable.

[0048] Figure 25 These are schematic diagrams and flowcharts illustrating an illustrative embodiment of the process of replacing the output node of an attention block with a multi-node unit and training the nodes in that unit in an interpretable manner.

[0049] Figure 26 This is a flowchart of an illustrative embodiment of the process referred to herein as "loop training".

[0050] Figure 27 This is a flowchart illustrating an embodiment of a process for improving the security of a text generation system.

[0051] Figure 28 This is a flowchart illustrating an example of a process for training a group of one or more nodes into a named set discriminator and for training and using a relevant confidence estimator.

[0052] Figure 29 A flowchart of an exemplary embodiment is shown, which is designed to improve performance and interpretability through targeted, systematic network growth.

[0053] Figure 30 It is a system diagram of a distributed system that includes multiple autonomous modular collaborative subsystems.

[0054] Figure 31 This is a flowchart illustrating an example of the process of training a system comprising one or more autonomous modular cooperative subsystems, such as... Figure 30 As shown. In a preferred embodiment, the computer system 1700 can expand the system during initial training and can continue to train and expand while the end user is using the system. During training, the computer system 1700 can expand the system with the goal of making it easier for human users to understand and control.

[0055] Figure 32 This is a flowchart illustrating an embodiment of a process by which a computer system 1700 can efficiently train large language models, including converter models and stochastic models, with an arbitrarily large number of trainable parameters.

[0056] Figure 33 This is a system diagram of an illustrative embodiment of one aspect of the present invention, wherein computer system 1700 collaboratively uses different types of models to efficiently train and rapidly incrementally grow one or more machine learning systems, while improving performance, interpretability, and controllability.

[0057] Figure 34 This is a flowchart illustrating an embodiment of one aspect of the present invention, which relates to data and resources used during the training and use of a user control and computer system 1700 tracking system.

[0058] Figure 35 In some embodiments, computer system 1700 can be used as follows: Figure 30 and Figure 33 The system shown and / or such Figure 31 , Figure 32 , Figure 36 , Figure 37 , Figure 38 and Figure 39 The flowchart illustrates an example of several optional procedures used in the process shown.

[0059] Figure 36 This is a flowchart illustrating an example of a collaborative process using different machine learning systems, such as... Figure 33 As shown, the generation system is a large language model based on a converter.

[0060] Figure 37 This is a flowchart illustrating an illustrative embodiment of the process of constructing a large-scale text generation system based on an integrated hierarchical structure of conditional probability models and joint optimization combinatorial networks. In some embodiments, the computer system 1700 can be implemented on a distributed computer system with multiple local computers. Figure 37 The process is shown.

[0061] Figure 38 This is a flowchart illustrating an embodiment of one aspect of the present invention, through which computer system 1700 can extend the state space of a hidden Markov process modeling sequence of text.

[0062] Figure 39 This is a flowchart of an exemplary embodiment for progressively building and training an arbitrarily large distributed AI system from individual components, each of which meets specific limitations in memory and / or computation.

[0063] Figure 40 This is a flowchart illustrating an example of text generation using a system that includes a stochastic process model.

[0064] Figure 41 This is a flowchart illustrating an embodiment of one aspect of the invention, in which neural networks can be grown incrementally, or hybrid networks can be created, or copies of one or more components can be made to improve network performance or make the network easier to understand and control.

[0065] Figure 42 This is a flowchart of an exemplary embodiment of a computer system 1700, which selects nodes to be segmented based on tests of one or more criteria used to make potential improvements from various reasons and methods of segmenting nodes.

[0066] Figure 43 This is a flowchart illustrating an embodiment of one aspect of the present invention, wherein computer system 1700 can manage the training, saving, and loading of certain types of conditional probability models.

[0067] Figure 44 This is a schematic diagram of an illustrative embodiment of one aspect of the present invention, wherein computer system 1700 can use combined networks, data correlation regularization links, and selective backpropagation to resolve correlation errors in order to jointly optimize the performance of a set of networks and train them to be distinct from each other.

[0068] Figure 45 The flowchart illustrates an embodiment in which computer system 1700 can generate text using a combination of a converter language model and a random model, collaboration between AI language models, and explicit collaborative interaction between a human author and an AI system acting as an author assistant.

[0069] Figure 46 This is a flowchart illustrating an embodiment of one aspect of the present invention, in which, in some embodiments, computer system 1700 can efficiently train a large neural network by first training a smaller neural network.

[0070] Figure 47This is a flowchart illustrating an example of a process by which a computer system 1700 can train a large language model.

[0071] Figure 48 This is a flowchart illustrating an example of a process by which a computer system 1700 can generate text using a pre-trained large language model.

[0072] Figure 49 This is a flowchart illustrating an embodiment of one aspect of the present invention, wherein a computer system 1700 trains a large language model including a Hidden Markov Process Model.

[0073] Figure 50 This is a flowchart illustrating an embodiment of one aspect of the present invention, wherein the computer system 1700 incrementally increases the size of the converter by increasing the number of attention heads in a specified attention layer.

[0074] Figure 51 This is a flowchart illustrating one aspect of an embodiment of the present invention, which uses virtual game theory to train a guardrail for a generative AI system and trains the system to detect guardrail violations.

[0075] Figure 52 This is a flowchart of an illustrative embodiment of the present invention, wherein the computer system 1700 uses a multi-path chain of unidirectional conversion to train the conversion system, wherein each link in the chain converts from the source language to the target language.

[0076] Figure 53 This is a flowchart illustrating an embodiment of one aspect of the present invention, wherein computer system 1700 uses a multi-path chain of pairwise language transformations to compute robust composite transformations.

[0077] Figure 54 This is a flowchart illustrating an embodiment of one aspect of the present invention, wherein computer system 1700 can add nodes with linear threshold activation functions to a neural network or hybrid network and train the nodes using methods other than gradient descent.

[0078] Figure 55 This is a flowchart illustrating an embodiment of one aspect of the present invention, wherein, in some embodiments, computer system 1700 can develop, grow, and train interpretable large-scale language model-generating AI systems.

[0079] Figure 56 This is a flowchart illustrating an example of a process for using an interpretable large-scale language model text generation system in an interactive deployment.

[0080] Figure 57 Examples of converter networks according to various embodiments of the present invention are depicted.

[0081] Figure 58 This is a schematic diagram of an interpretable large-scale language model text generation system according to various embodiments of the present invention.

[0082] The process shown in the figure can be implemented in a multiprocessor computer system 1700, such as... Figure 17 As shown. In a preferred embodiment, the training and development of the system under development can be supervised by a collaborative effort between knowledge engineers and a human team working on the AI ​​system, referred to herein as a Hybrid Network Learning Management System (HNLMS). The AI ​​system in the HNLMS can also be implemented on a computer system such as Computer System 1700. Detailed Implementation

[0083] The following paragraphs provide definitions for the accompanying drawings.

[0084] Neural networks: a type of directed graph consisting of a set of nodes and a set of directed connections between ordered pairs of nodes. Typically, each connection has associated learned parameters called weights. Typically, a computer system 1700 multiplies the output of the source node of a connection by the weights of that connection to compute a value that is then provided as input to the destination node of the connection. Figure 21A A feedforward neural network with multiple hidden layers is shown.

[0085] Most of the discussion in this disclosure likely pertains to acyclic neural networks, whose graph is a directed acyclic graph. However, the computer system 1700 can create multiple copies of a recurrent neural network, where all connections that would create a cycle are redirected to the next copy of the network. In this way, the computer system 1700 can model a recurrent neural network as a large “unfolded” network of acyclic copies of the underlying network, and therefore, for practical purposes, assuming that the graph of the neural network is a directed acyclic graph does not result in a loss of generality.

[0086] The computer system 1700 can also use this unfolding mechanism with hybrid networks. Furthermore, hybrid networks provide additional methods for training models of cyclic processes. For example, the computer system 1700 can model cyclic processes using hidden state-space models within hybrid network units. In hybrid networks, units can be connected using bidirectional data communication links. These data communication link networks may contain cycles.

[0087] Nodes: Nodes in a neural network. In hybrid networks, elements other than internal neural nodes within a unit are called units and cells, not nodes. A node within a unit can receive connections from nodes in other units and can send connections to nodes in other units.

[0088] Unit: A unit is a generalization of a neural network node. A unit may have multiple output values, and multiple connections for each output value. A unit can include multiple nodes and sub-units. A unit can also include special-purpose elements called "cells," which are linked to each other via data communication links rather than network connections. A unit can include a single neural node or a single cell.

[0089] Cell: An element in a hybrid network that can store and transmit the value of a specified variable. Computer system 1700 can store and execute program code associated with the cell when it receives data as network input or data sent from other cells. A cell can be associated with program code that computer system 1700 can execute when computing the network's activation and response to a specified input data item.

[0090] Hybrid networks: Networks composed of cells and connections, rather than neural nodes and connections. Hybrid networks may also include cells and data communication links. The computer system 1700 can change and customize the configuration of dynamic hybrid networks after receiving data items to be classified.

[0091] Components of a neural node: A typical node in a neural network consists of two component operations: affine summation and activation function.

[0092] Affine summation: In the affine summation operation of a neural node, the computer system 1700 calculates a weighted sum of the input values ​​connected to the node plus a node-specific bias term.

[0093] Activation function: In a typical neural node, the computer system 1700 computes a specified function that performs an affine summation. This function is called the node's "activation function." The value of the activation function for a data item d is called the "activation" of the node for data item d. The output value of the node is the output of the activation function for data item d. Examples of activation functions include, but are not limited to, sigmoid, softmax, Tanh, and ReLU (rectified linear unit) activation functions.

[0094] Implicit error: When the computer system 1700 compares the activation of a node with respect to a specified threshold with the sign of the backpropagation derivative of the objective function, the computer system 1700 may determine that it has a standard discriminator activation function (in... Figure 2 An error occurred in an internal node (defined in box 203) on a specific data item.

[0095] Known set: A known set is a set of data items that the computer system 1700 can determine with a specified precision whether any particular data item is in the known set. For example, the training data item set for any output class in a classification system is a known set. Any set of items that the computer system 1700 can detect with a specified precision based on the output values ​​of nodes, cells, units, or networks within a specified range is a known set.

[0096] Named sets: A named set is a known set in a computer system 1700 that has a name that is easily understood by humans. Typically, a collection of data items of any output category is a named set. In some embodiments, humans may provide names for unnamed known sets.

[0097] Network repository: A repository of previously trained nodes, cells, units, and networks, which can be implemented by computer system 1700. In some embodiments, computer system 1700 may place a trained network or a partially trained network into the network repository. In some embodiments, computer system 1700 may place a subnetwork that activates a selected node, cell, or unit into the network repository. In some embodiments, computer system 1700 may share some or all of its network repository contents with other computer systems.

[0098] Knowledge engineering: Developing tools for analyzing data and computing useful functions and properties of data in a specific domain to facilitate the development of machine learning systems that classify data items in that domain.

[0099] Hybrid Network Learning Management System (HNLMS): A system in which a team of one or more people collaborates with one or more AI systems. The human team and the AI ​​systems guide the training of the hybrid network to improve the sensitivity, overall interpretability, and performance of the trained network.

[0100] Detector: A node, cell, or unit with output values, characterized by the computer system 1700 as attempting to have values ​​within a specified interval for data items in the target acceptor set, and values ​​outside the specified interval for data items not in the acceptor set. In some embodiments, the specified interval is a set of values ​​above a specified threshold. In some embodiments, the target acceptor set is known to the computer system 1700, for example, the output node of a classifier used to supervise training data. The actual set of data items within the specified interval may be referred to as the "empirical acceptor set." Where the meaning is clear, the target acceptor set or the empirical acceptor set may simply be referred to as the "acceptor set." In some embodiments, the target acceptor set of a network element is not explicitly specified and is not a known set. In some embodiments where the acceptor set of the detector is not explicitly known, the computer system 1700 may temporarily associate the output values ​​with a known set based on experience.

[0101] Discriminator: A node, cell, or unit with an output value that the computer system 1700 characterizes as attempting to have a value in a first specified interval of data items in a first target accepting set and a second specified interval of data items in a second target accepting set. In some embodiments, the computer system 1700 may not have a target interval for data items not in either accepting set. In some embodiments, a cell may have additional output values ​​to characterize data items not in either target accepting set.

[0102] Recall: In a data retrieval or detection task, the score of a machine learning system is the number of target data items that it correctly retrieves or detects from a specified set of data items, divided by the total number of target data items in the specified set of data items.

[0103] Accuracy: In data retrieval or detection tasks, the score is the number of target data items that a machine learning system correctly retrieves or detects from a specified set of data items, divided by the total number of data items (including incorrect or erroneous items) in the specified set of data items detected or accepted by the machine learning system.

[0104] Association: The association between a specified known or named set and a dataset corresponding to a specified detector node, cell or cell or node activation function interval refers to the specified criteria used to determine the recall and / or precision of a specified detection that meets the specified known or named set.

[0105] Knowledge-sharing links: Knowledge-sharing links are ordered pairs of nodes, a reference node, and a receiving node. Nodes can both be in the same network or in two separate networks. Only the receiving network needs to be in the network currently being trained. If the nodes are in separate networks, they must be able to activate both nodes on the same data item. For example, the two nodes can share a global or local input data space. In some embodiments, the computer system 1700 can compute a mapping from one data space to the other. During training of the network including the receiving node, for a given data item, if the activation of two nodes fails to satisfy a specified relationship, the computer system 1700 can impose a regularization penalty.

[0106] Relationships in Creative Commons Links: A common example of a relationship in Creative Commons links is the "equal to" relationship. For example, in Act... 参考 (Data) and Act 接收 The equality relationship between (data) can be subject to regularization penalties by the computer system 1700.

[0107]

[0108] Here, α is a hyperparameter, controlled by HNLMS, for example. The hyperparameter α is called the "strength" of the knowledge-sharing link. HNLMS can also specify that regularization is applied only to specified data items. A knowledge-sharing link is not a connection. For example, in acyclic networks, a link might run from a higher-level reference node to a lower-level receiving node, which is not allowed in acyclic networks. Other common knowledge-sharing relations include less than, greater than, and not equal to. By convention, in asymmetric relations, the reference node is the first argument.

[0109] Inequality relationships of greater than and less than are useful, for example, when sharing knowledge between two nodes, where one node is associated with a known set that is a subset of the known set associated with the other node. For example, the set of horses is a subset of the set of equines, the set of equines is a subset of the set of mammals, the set of mammals is a subset of the set of animals, and the set of animals is a subset of the set of organisms. In some embodiments, the computer system 1700 may impose a knowledge-sharing link, whereby the activation of a node associated with a superset should be greater than or equal to the activation of a node associated with a subset. For act... 参考 (Data) and Act 接收 For relationships where data is greater than a certain threshold, the computer system 1700 can apply regularization penalties:

[0110]

[0111] Here, α is a hyperparameter, for example, controlled by HNLMS. For instance, in speech recognition, the activation of nodes associated with a set of vowels should be greater than or equal to the activation of nodes associated with a set of higher front vowels. In some embodiments, the computer system 1700 may restrict the execution of regularization to data within a specified interval in the reference nodes.

[0112] In some embodiments, the computer system 1700 may limit the maximum regularization penalty for inequality relations. For example, the computer system 1700 may impose a regularization penalty on inequality relations:

[0113]

[0114] It has a maximum penalty β, where α and β are hyperparameters, for example, controlled by HNLMS.

[0115] In some embodiments, the computer system 1700 may apply a link that is equal to or not equal to a knowledge-sharing link in two directions between a pair of nodes.

[0116] Using links that are equivalent to knowledge sharing in both directions is also called "soft binding" of node pairs. Using links that are not equivalent to knowledge sharing in one or both directions is also called "unbinding" of node pairs. In some embodiments, soft binding and unbinding links can be bidirectional, although unbinding links are asymmetric.

[0117] In some embodiments, the computer system 1700 may apply regularization equal to soft binding and / or not equal to anti-binding to the weight parameters of one or more corresponding connections in a pair of same-origin nodes. However, since the values ​​of the weight parameters are independent of the data, the knowledge-sharing links between the weights are also independent of the data.

[0118] Flat activation interval: An interval within an activation function that satisfies a specified flatness criterion, such as a limit on the magnitude of the function's derivative or a limit on the difference between the function's maximum and minimum values ​​within the interval. The extreme case of a flat activation interval is an interval where the function has a constant value throughout the entire interval.

[0119] Data exclusion: The process of excluding data based on specified criteria when training or deploying units in a hybrid network.

[0120] Data exchange: A component in a network that can selectively pass activation or other incoming variables to a specified subset of one or more destinations. In some embodiments, the specified subset may be an empty set.

[0121] Local data space: In a hybrid network, the n-tuple variables are the input variables for a specified set of units and / or nodes. Variables in the local data space may reside in the inner layers of the network. The local data space can also be called the "local input space" or "local feature space." The local data space may be the encoding of a larger set of variables.

[0122] Decision element: A specified interval within the range of a computable variable f(d) that depends on the network input data d, where the value of f(d) within the specified interval is interpreted as a variable indicating whether the data item d is in the specified set (detection) or not in the specified set (rejection).

[0123] Decision element set: A set of one or more detection decision elements, wherein the specified target detection sets are disjoint. The computer system 1700 may interpret the discriminator as a set of decision elements comprising two intervals, each interval being a decision element detector of one of the discriminator's alternatives. The computer system 1700 may interpret the softmax set as a set of decision elements, wherein each node in the softmax set is a detector of a target set that is disjoint with the other nodes.

[0124] Holistic interpretation: Providing a human-understandable explanation of the relationships between nodes or elements and other nodes and elements, as well as the entire system. Many techniques that improve sensitivity also contribute to holistic interpretability, and vice versa. For example, the association of a node or element with a named set is directly an aspect of holistic interpretability and also contributes to improved sensitivity.

[0125] Alternate derivative function: A specified function used by the computer system 1700 for the activation function input of one or more specified data items, replacing the actual derivative of the activation function. HNLMS can specify the same alternative derivative function for selected nodes of all data items, or can specify different alternative activations for different data items. HNLMS can change the specified alternative derivative function during training.

[0126] Template model: A specified computation designed to assign higher values ​​to data items in a particular target set than to data items not in the target set, while satisfying a specified basic sensitivity criterion. In an illustrative embodiment, the template model includes inputs from a local or global data space, a specified norm in the data space, a specified centroid of the target set within the data space, and an output value that is a function of the distance from the centroid to the input data item, measured by the norm. The template model can be represented in nodes, cells, or cells. Without loss of generality, in an exemplary embodiment, computer system 1700 may represent the template model as a dedicated cell, since a cell paired with a specified node or cell can represent the same computation as the node or cell that includes the computation in that cell.

[0127] Robust template model: A template model designed to meet specified sensitivity standards.

[0128] Various definitions have been provided, and embodiments of the present invention are further described below. Figure 1 This is a flowchart illustrating one aspect of an embodiment of the present invention. Figure 1 In the illustrated embodiment, computer system 1700 constructs and trains a hybrid network. In terms of equivalent computation, the hybrid network comprises a strict subset of neural network categories.

[0129] In a preferred embodiment, the process of building and training the hybrid network is a process of continuous growth and improvement of the system built and trained using multiple training methods. In blocks 101-107, the computer system 1700 modifies and extends the system under development prior to deployment. In blocks 108-114, the computer system 1700 continues to extend and train during and after deployment. In various aspects of the invention, the computer system 1700 may use various processes to improve the sensitivity of the system under development. For ease of discussion, the improvement processes are divided into two levels. Each level is associated with a different criterion for evaluating sensitivity. Generally, the second level of sensitivity involves more complex sensitivity criteria. In some embodiments, the computer system 1700 may use a specific process to improve the sensitivity level, rather than discussing the level of that specific process.

[0130] In box 101, computer system 1700 selects one or more base machine learning systems. In some embodiments, computer system 1700 may select a base machine learning system not represented as a network and use incremental growth to construct a hybrid network. In some embodiments, computer system 1700 may select a partially trained or fully trained traditional neural network as the base system. The following is in conjunction with... Figure 21A A traditional feedforward neural network is described. In some embodiments, the computer system 1700 may select a hybrid network as the base network. Figure 1 As shown in other figures, the computer system 1700 can modify and add to the base system during continuous training.

[0131] In some embodiments, the computer system 1700 can jointly train multiple networks. In some embodiments, the computer system 1700 can jointly train diverse sets of homologous networks, including diverse sets of perceptual hybrid networks and diverse sets of canary networks, as well as optional diverse sets of networks optimized for classification accuracy, regardless of sensitivity, such as combining... Figure 21 and Figure 5 The explanation given in box 516.

[0132] In some embodiments, the computer system 1700 may select a single underlying network.

[0133] If the base network is a traditional neural network, the computer system 1700 can modify and extend the network to make it a hybrid network. In some embodiments, the computer system 1700 may choose an empty network as the starting network and extend a reasonable hybrid network from scratch. In some embodiments, the computer system 1700 may use a non-network or network base system as a reference system for knowledge sharing and / or simulated learning to extend a reasonable hybrid network from scratch. Simulated learning is described in U.S. Patents 11,410,050 and 11,531,900, entitled “Imitation training for machine learning systems with synthetic data generators,” and in published PCT application WO / 2021 / 194516, entitled “Data-related node-to-node knowledge sharing by regularization in deep learning,” the entire contents of all of these patents are incorporated herein by reference.

[0134] In some embodiments, the computer system 1700 may use one or more reference networks as references to known or named sets.

[0135] In some embodiments, the computer system 1700 may use human consultation to associate names with known sets. In some embodiments, when the computer system 1700 associates names with known sets, it can then train one or more detectors for the known sets to better match the detection of named sets. The computer system 1700 may use named set detectors in a reference system to train detectors in the current system through knowledge sharing and / or simulation learning. In simulation learning, elements in the trained system are trained using a local training objective to match the output of specified elements in the reference system. In some embodiments, the computer system 1700 may use one-way or two-way transformations between the data space in the current network and the reference network to apply knowledge sharing and / or simulation learning. Figure 4 Box 414 further discusses human consultation. Combined with... Figure 14 One-way and two-way transformations of the data space were discussed.

[0136] In box 102, in some embodiments, computer system 1700 may optionally acquire and / or build and train one or more systems that are smaller or simpler than the current base system. For example, in some embodiments, computer system 1700 may specify a simpler system to facilitate potential human guidance and consultation, such as in conjunction with... Figure 4 The discussion is framed in box 414. In some embodiments, a human consultant may specify experimental changes to the system. In some embodiments, specifying experimental changes on a simpler system may require less time and effort than on a more complex system. In some embodiments, the computer system 1700 may follow specified design rules to make the simpler system easier for a human consultant to understand and control.

[0137] In some embodiments, the computer system 1700 may specify a simpler network to reduce the computational cost required for training. In some embodiments, the computer system 1700 may specify a simpler network for which it is easier to design and train sensitivity. In some embodiments, the computer system 1700 may specify a simpler network to achieve better overall interpretability.

[0138] In some embodiments, computer system 1700 may work in parallel with one or more simpler systems alongside the current underlying system. In some embodiments, computer system 1700 may temporarily replace the current underlying system with a simpler system.

[0139] In some embodiments, these simpler systems may also be designed to generalize better from a limited amount of training data. The goal of these simpler systems is not to match the classification accuracy of the base system selected in box 101. Instead, the primary goal is to reduce the risk of making imperceptible errors. The vulnerability of a classifier system to imperceptible errors tends to be proportional to the number of input variables, thus the computer system 1700 makes simpler systems with fewer input variables less vulnerable to attack. In some embodiments, the computer system 1700 may use one or more smaller, simpler systems to accelerate the training and use of larger systems.

[0140] In image recognition tasks, a smaller, simpler system example is that computer system 1700 preprocesses images to obtain lower-resolution images. In speech recognition tasks, a smaller, simpler system example is that computer system 1700 uses fewer spectral frequencies and / or computes fewer speech spectral frames per second. In some embodiments, computer system 1700 can reduce the average number of spectral frames per second by using a variable frame rate. For example, if the difference between several consecutive spectral frames is less than a specified amount, computer system 1700 can replace multiple frames with a single frame.

[0141] In smaller, simpler systems, computer system 1700 can use fewer categories in classification tasks. More generally, computer system 1700 can use fewer, larger sets at each level of the ontology. In some embodiments, the larger sets in a simpler system can be the union of sets in the ontology of a more complex system.

[0142] On the other hand, in the case of image recognition, in some embodiments, the computer system 1700 can leverage the availability of higher-resolution images to analyze input data items of a smaller, simpler underlying system. For example, when aligning data items with a volumetric graph, such as in conjunction with... Figure 12 The computer system 1700 discussed herein can use higher-resolution images to verify the preliminary alignment of regions in a low-resolution image with specified portions in volume theory. In this example, the system analyzing the higher-resolution image is... Figure 1 A "simpler" system in the sense of box 102.

[0143] In box 103, computer system 1700 initiates or resumes the ongoing development and improvement process of the current network (i.e., the base system selected in box 101, or optionally combined with a simpler system selected in box 102 if one is selected in box 102). In boxes 101 and / or 102, based on the verification tests in box 106 or box 111, computer system 1700 may have replaced the previous base network with a new base network.

[0144] In some embodiments, the computer system 1700 can use incremental growth ( Figure 5 (504) Improves classification performance and sensitivity through training without using any backpropagation, neither using derivative backpropagation ( Figure 5 506 in the middle), and also does not use backpropagation of labeled data examples ( Figure 5 (510 in the text). For example, in some embodiments, the computer system 1700 may use constraint optimization ( Figure 5 and Figure 6 The 524th generation of the algorithm is used to train each new node incrementally added to the network without backpropagation. As long as any residual errors exist on the training data, the computer system 1700 can use a combination of incremental growth and constrained optimization to reduce the number of errors.

[0145] In some embodiments, the computer system 1700 may add elements to the network as part of various embodiments of hybrid training, such as data delegation (…). Figure 11 and Figure 5 (box 518), splitting one or more nodes ( Figure 5 519), adding additional output values ​​to components, training different sets in the discrimination ( Figure 5 (523) Add a local autoencoder to the network or simply add one or more components for other purposes.

[0146] In some embodiments, the computer system 1700 may add a local autoencoder to the network to support increased sensitivity. Figure 2 , Figure 9 and Figure 10 ), as a local data space ( Figure 3C and Figure 4 and Figure 9 (frame 411), or as Figure 5 Data generator (514).

[0147] In some embodiments, computer system 1700 can create multiple networks from an initial base network and can continue to expand and improve each of the multiple networks. For example, in addition to one or more current base systems, computer system 1700 can also use one or more simpler systems specified in box 102. As another example, computer system 1700 can develop one or more canary networks while developing the current base network. Canary networks are designed to be vulnerable to adversarial attacks and other input disturbances as a means of detecting and diagnosing such disturbances. Canary networks and Figure 4 Let's discuss box 415 together.

[0148] In some embodiments, in block 103, computer system 1700 may build a hybrid network from scratch.

[0149] In boxes 104 and 105, computer system 1700 modifies the base network to improve its sensitivity across a hierarchical structure with two sensitivity levels, employing various training methods and techniques to enhance sensitivity. Each sensitivity level has a different criterion. Computer system 1700 can use different processes, models, and system designs to improve each level. Figure 2 , Figure 4 , Figure 5 The illustrative processes and models for each sensitivity level are discussed in more detail, along with other accompanying figures. However, in some embodiments, the computer system 1700 may also use improved processes or models at levels different from those discussed.

[0150] In some embodiments, in block 104, the computer system 1700 can improve basic sensitivity ( Figure 2 and Figure 4 (Frame 405), actively flatten ( Figure 4 (box 406), perform hybrid training ( Figure 5 and Figure 4(Box 407 in the image) Find the best location for a piece of knowledge in the network. Figure 4 (box 408), and / or perform data-selective training ( Figure 4 (box 409). In some embodiments, the computer system 1700 can Figure 1 Randomization was used in training for boxes 104, 105, 106, 109 and / or 110. Figure 5 (box 520) and randomized activation ( Figure 4 (Block 418) to improve sensitivity, robustness and / or classification performance.

[0151] One aspect of a preferred embodiment of the present invention is a Hybrid Network Learning Management System (HNLMS) comprising a team of human experts and a collaborative association of one or more AI systems to develop tools and models to help computer system 1700 improve the sensitivity, classification performance, and overall interpretability of the developed system. In some embodiments, HNLMS can guide the training of the system under development and can determine its sensitivity.

[0152] The descriptive criterion for the first level of sensitivity of the detector or discriminator is that, for any data item within the empirical and target acceptance sets, for a specified ε, the input... Changes in the norm < ε should not cause a data item to fall out of the empirical acceptance set. In other words, a barely perceptible change in the input data items should not cause the system to commit an error that was not present before the change. Any successful Counter-attacks all violate the first level of sensitivity. Techniques for countering attacks are well-known to those developing deep neural networks.

[0153] A significant subset of the first-level sensitivity is called the "basic sensitivity". Basic sensitivity ( Figure 4 The 405) has standards that can be checked for each node, cell, or internal variable. For basic sensitivity, the computer system 1700 modifies the underlying system to improve the basic sensitivity of each node or cell.

[0154] Informally, sensitivity levels differ in the technology and / or judgment of the development system at each level of HNLMS participation in the development of the 1700 computer system.

[0155] Level 1 technology requires minimal involvement from HNLMS during development. The sensitivity of a system improved through Level 1 sensitivity enhancements is also the easiest to objectively evaluate using a computer system (1700). HNLMS primarily controls the hyperparameters in the sensitivity standard.

[0156] In box 105, in some embodiments, computer system 1700 may modify the current network to improve second-level sensitivity. In the second-level technique, computer system 1700 may utilize further guidance from HNLMS during development and evaluation. Figure 4 (414).

[0157] Such as combination Figure 4 The discussion is in Figure 1 In box 105, in some embodiments, computer system 1700 can analyze and improve decision boundaries ( Figure 4 (the box 410) and / or create and train the local normed space ( Figure 9 and Figure 4 (411)

[0158] In some embodiments, Figure 1 In box 105, computer system 1700 can calculate attributes and other variables that computer system 1700 may store in the cell, such as in combination Figure 4 The discussion is in box 412.

[0159] exist Figure 1 In box 105, computer system 1700 can also construct and train hidden state space models under guidance, for example, HNLMS, such as combining Figure 4 Box 413 and Figure 7 The discussion continues. In some embodiments, the computer system 1700 may also use a hidden state space model in active classification, such as in conjunction with... Figure 1 Box 109 and Figure 4 The discussion is in boxes 403, 416 and 417.

[0160] The computer system 1700 can specify and / or change the state of the hidden state space model and / or associated learning parameters or hyperparameters under the control of, for example, HNLMS. In some embodiments, the computer system 1700 can specify and / or change the state of the hidden state space model and / or associated learning parameters or hyperparameters based on human consultation, such as in combination with... Figure 4 The discussion is in box 414.

[0161] exist Figure 1 In box 105, computer system 1700 can use human consultation to verify the sensitivity of the discriminator and / or classifier decision boundaries, such as by combining Figure 4 The discussion is in box 414.

[0162] To improve sensitivity, Figure 1 In boxes 105 and 106, computer system 1700 can analyze the decision boundary ( Figure 4 410), constructing a local normed space ( Figure 4411 in the middle), calculated properties and cell variables ( Figure 4 412), construct and train the hidden state space model ( Figure 7 and Figure 4 (Box 413), construct and train the active defense structure, optionally using data exchange ( Figure 4 416 and Figure 8 Under 803), perform active alignment ( Figure 12 and Figure 19 as well as Figure 4 Box 417), random activation training ( Figure 4 418), building and training robust template models ( Figure 10 and Figure 4 (box 419) and / or use mixed conditional training ( Figure 13 and Figure 5 (Frame 512).

[0163] In addition to improving sensitivity, some illustrative processes and models can also improve the overall interpretability of nodes and units in the system. Some illustrative processes and models can improve the performance of assigned classification or regression tasks. In one aspect of the invention, the computer system 1700 can reformulate a regression task as a classification task. Without loss of generality, in this disclosure, the term "classifier" is used to refer to a system whose task can be either a classification task or a regression task.

[0164] As mentioned above, the phrase "neural network" is used to refer to a directed network consisting of a set of nodes and a set of directed connections between ordered pairs of nodes. The phrase "neural network" is a generally accepted concept familiar to those skilled in the art of training and using neural networks.

[0165] As mentioned above, the phrase "hybrid network" refers to a generalization of a neural network composed of more complex elements, referred to here as a "unit." A unit may have multiple output values ​​and may include multiple internal nodes and connections, such as... Figure 3A As shown. A unit may also include special elements referred to herein as a "cell". On the other hand, in hybrid networks, a unit may consist of only a single neural node, thus any traditional neural network is also a simple hybrid network.

[0166] Modifications made to the underlying network by computer system 1700 in boxes 104 and 105 may include changing the activation function of one or more selected nodes. Modifications may include transforming one or more nodes into a more complex structure called a “cell.” Modifications may include adding nodes and cells to the network. In some embodiments, modifications may include creating one or more cells and adding them to the network. Computer system 1700 may add cells to cells or to the network outside of any cell.

[0167] Cells in a hybrid network differ from nodes. A cell may include the values ​​of one or more variables that can be computed by the computer system 1700. For example, the computer system 1700 may store in a cell the output value of the network-selected element of the current input data item and / or the output value of the network-selected element of a previous data item. Each cell may include or be associated with any stored program to be executed on the computer system 1700. For example, the computer system 1700 may perform serial computations associated with the cell to compute logical or probabilistic reasoning. A cell may include one or more incoming data communication links and / or one or more outgoing data communication links. Data links differ from neural network connections. Data links only transmit data and do not have associated "weight" parameters. Data links can be unidirectional or bidirectional.

[0168] The data transmitted by computer system 1700 on the data link from the first cell to the second cell may include any value that computer system 1700 can calculate based on the value stored in the first cell.

[0169] The computer system 1700 can also transfer data from a neural node to a cell via a data link. For example, the data transferred from a neural node to a cell may be the input or output of a neural node activation function. In some embodiments, the data transferred from a neural node to a cell may be the value of the backpropagation derivative calculated by the computer system 1700 during gradient computation via backpropagation. In some embodiments, the backpropagation derivative may be derived from an alternative local derivative (…). Figure 5 (509).

[0170] The computer system 1700 can also transmit data from cells to neural nodes via data links. The data transmitted by the computer system 1700 on the data link from cell to neural node can be any value that the computer system 1700 can calculate based on the value stored in the cell. In some embodiments, the computer system 1700 can use the received data value as an additional input connection to the receiving node, with a connection weight of 1.0. In a preferred embodiment, the computer system 1700 does not backpropagate derivatives along the data link from the cell. However, if desired, in some embodiments, the computer system 1700 can achieve a similar effect by creating a second node to receive data from the cell and then connecting the second node to the first node via a neural network connection through which the computer system 1700 can backpropagate derivatives.

[0171] Combination Figure 2 , 4 The details of the processes used in each level (boxes 104 and 105) are discussed in more detail in Figures 5 and others.

[0172] In box 121, in some embodiments, computer system 1700 can train a network to participate in human + AI joint activities, where one or more humans play a sufficient role to contribute a certain amount of common sense. An example of a joint + AI activity is HNLMS. Figure 21 Other joint activities were discussed, including the creation of creative works. Figure 21 Joint educational activities were also discussed.

[0173] In box 106, computer system 1700 trains the modified network and tests the trained network on validation data retained from the training data. Figure 5 An exemplary embodiment of the process of training a hybrid network, referred to as "hybrid training," is discussed in conjunction with the accompanying figures.

[0174] In some embodiments, in block 106, computer system 1700 may perform histogram analysis ( Figure 15 and Figure 5 (box 507), backpropagation derivative () Figure 5 506), create a low-dimensional local data space and construct a low-dimensional model ( Figure 5 517), Performing Data Delegation and Data Exclusion ( Figure 4 Frame 420 Figure 5 Frame 518 and Figure 10 and 11 ), determine local targets ( Figure 5 508), using alternative derivative functions ( Figure 5 509), backpropagation labeled data ( Figure 18 and Figure 5 (box 510), simulate another network ( Figure 5 511), Execution Condition Hybrid Training ( Figure 13 and Figure 5 (box 512), execution experience training () Figure 5 (521) generates more data, optionally under human guidance ( Figure 5 514), constructing a homologous network ( Figure 21 and Figure 5 516), for random training ( Figure 5 (520), based on experience, the weight of individual items is calculated to estimate their reliability ( Figure 5 (522), create different sets to better represent combinations of known sets ( Figure 5 523), and / or use constrained optimization ( Figure 6 and Figure 5 (Box 524).

[0175] If the verification tests performed by computer system 1700 in block 106 meet the specified acceptance criteria, then computer system 1700 replaces the previous base network with the network modified by computer system 1700 in blocks 104 and 105. In some embodiments, computer system 1700 may store the new base network or selected subnetworks in a network repository.

[0176] In some embodiments, the computer system 1700 may compare the performance of the current base system on validation data with the performance of a simpler system. In some embodiments, the computer system 1700 may compare the performance of the current system on data from adversarial attacks on the validation data with the performance of one or more canary systems. In some embodiments, based on the analysis of these comparison results, the computer system 1700 may make experimental changes to the current system and, preferably, retest it on new validation data. In some embodiments, the computer system 1700 may request human consultation, such as in conjunction with... Figure 4 The discussion is in box 414.

[0177] In box 107, computer system 1700 checks the stopping criteria for ongoing modifications and training in the loop from box 101 to box 107. If the stopping criteria are met, computer system 1700 proceeds to box 108. Otherwise, computer system 1700 returns to box 101 to continue modifying the current base network.

[0178] In box 108, computer system 1700 receives items to be classified. In some embodiments, the phrase "items to be classified" may include items for which regression values ​​are to be calculated.

[0179] In block 109, in some embodiments, computer system 1700 may perform a process referred to herein as “active classification” or “actively perceptible classification.” In a preferred embodiment, during active classification, computer system 1700 may modify the network and / or perform computations other than neural network activations upon receiving data items to be classified. Computer system 1700 may customize these additional computations for the received data items.

[0180] Active perceptible classification involves calculating the activation values ​​of neural nodes in a network, a process referred to as "inference" in neural networks. However, in an exemplary embodiment, "active perceptible classification" may include additional processes different from neural network inference.

[0181] In box 109, computer system 1700 can perform diagnostics and defenses on specific data items received in box 108. For example, computer system 1700 can use various unprotected canary networks and various robust networks to classify the received items to analyze dissimilar nodes in the response, such as... Figure 4 The discussion is in box 415.

[0182] When using a hybrid network for active classification in box 109, computer system 1700 can perform serial calculations in the cell after receiving the item to be classified. This capability provides additional functionality to the hybrid network.

[0183] For example, in proactively perceptible classification, the computer system 1700 can modify the hybrid network after receiving items to be classified, as a proactive defense. Figure 4 416 and Figure 8 (803), which enables computer system 1700 to make the network aware of specific items received. In some embodiments, computer system 1700 may construct a hybrid network with data exchanges that efficiently reconnect the hybrid network in a specially designed configuration to avoid imperceptible responses to specific data items received in block 108.

[0184] In some embodiments, in box 109, computer system 1700 may calculate the alignment of the data item to be classified with the model and / or other data examples ( Figure 12 and 19 as well as Figure 4 (Box 417). In some embodiments, the computer system 1700 may use cells in the network to store information used during alignment computation. In some embodiments, the computer system 1700 may use a hidden state space model ( Figure 7 and Figure 4 (See box 413). In some embodiments, the computer system 1700 may retrieve example alignments or other information from a repository when calculating the alignment of data items to be classified. In some embodiments, the computer system 1700 may store the information calculated when aligning data items to be classified for future use.

[0185] As another example, computer system 1700 can use a set of cells to model hidden random processes, such as combining... Figure 7 and Figure 4 As discussed in box 413. For a set of cells in a hybrid network, the computer system 1700 can perform cyclic computations even if the neural node network is acyclic.

[0186] In block 110, in some embodiments, the computer system 1700 can continue training after the machine learning system has been deployed. In some embodiments, the computer system 1700 can continue to acquire new data during system deployment. In some embodiments, the computer system 1700 can acquire data from other already deployed systems. In some embodiments, the computer system 1700 can use data acquired during the development and training of a new system to continue training the deployed system.

[0187] In some embodiments, in block 110, the computer system 1700 may continue to modify and extend the network to improve classification performance, sensitivity, and / or overall interpretability.

[0188] In block 110, in some embodiments, computer system 1700 may use the items received in block 108 for classification to compute incremental training. Since the item is received for classification, unlike the training data, it may not know the correct classification. In this case, in some embodiments, computer system 1700 may perform semi-supervised training, that is, after classifying the received item, computer system 1700 may perform incremental training on the item as if it were training data with classification labels computed during classification.

[0189] However, as is well known to those skilled in the field of semi-supervised training, while semi-supervised training often works well, it can sometimes fail disastrously.

[0190] In a preferred embodiment, in block 110, computer system 1700 may perform additional processes to improve the reliability of semi-supervised training. For example, computer system 1700 may use the data exchange associated with block 109 to construct a virtual ensemble, which not only improves the overall performance of classification but, more specifically, can detect and diagnose that the classification of received items may be unreliable. If computer system 1700 detects that the classification of an item may be unreliable, computer system 1700 may skip that item in the semi-supervised training.

[0191] In some cases, during deployment, computer system 1700 can determine the correct classification from interactions with end users, who can then correct any errors made by the system. In other cases, computer system 1700 may not know the correct answer, but based on user feedback, it may know that the calculated classification is incorrect or unreliable.

[0192] In block 111, in some embodiments, computer system 1700 may perform iterative training using accumulated data acquired through multiple passes of the loop from block 108 to block 112. In some embodiments, computer system 1700 may then validate the performance of the trained system on a set of labeled validation data reserved from the training dataset. If the validation test meets specified acceptance criteria, computer system 1700 may replace the current base network with the newly validated network.

[0193] In box 112, computer system 1700 checks the criteria used to stop or pause the process from box 108 to box 112. If the stopping criteria are met, computer system 1700 continues to box 114. Otherwise, computer system 1700 returns to box 108 to process more items to be sorted.

[0194] In block 113, computer system 1700 can determine whether to add more data to the training data and can determine how much data to select in a specific region. In some embodiments, computer system 1700 can begin training with selected data samples and gradually add more training data as the system grows. In some embodiments where a large amount of data exists, the data may not be uniformly distributed across the regions of interest. In some embodiments, computer system 1700 can selectively add sampled data to regions where sampling is currently sparse. In a preferred embodiment, the computing system can track the relative frequency of sampling and appropriately adjust any estimates of prior or posterior probabilities.

[0195] For example, in some embodiments, the computer system 1700 may use selective sampling in histogram analysis, which will combine Figure 15 This will be discussed. In some embodiments, the computer system 1700 may use selective sampling in any data segmentation process, for example: (1) activating interval data exchange ( Figure 2 (2) Training dependent on intervals (209 and 211) Figure 4 (406, 407, 409, 410 and 416), (3) Node segmentation ( Figure 5 (519), and (4) histogram analysis ( Figure 15 and Figure 5 (507).

[0196] In some embodiments, the computer system 1700 may use selective sampling in other situations where additional data is used, such as (5) backpropagation of data. Figure 18 and Figure 5 (510) (6) Adjust data delegation and exclusion norm ( Figure 4 Frame 420 Figure 5 Frame 518 and Figure 10and 11 (7) Generating data under human guidance ( Figure 5 (514), and (8) randomized training and diagnosis ( Figure 5 (520).

[0197] In box 114, computer system 1700 checks whether to resume the training and growth of the current base network, which was modified in boxes 103 through 110 and validated in boxes 106 and 111. If so, computer system 1700 returns to box 102. Otherwise, computer system 1700 proceeds to box 115.

[0198] In box 115, computer system 1700 checks the stop criteria. If the stop criteria are met, computer system 1700 exits. Figure 1 The process is shown. Otherwise, computer system 1700 returns to box 101.

[0199] In some embodiments, if additional training data has been acquired, the computer system 1700 can resume training of the currently updated base system in box 101. In some embodiments, the computer system 1700 may select one or more new base systems.

[0200] Figure 2 This is a flowchart illustrating an exemplary embodiment of the process for enhancing basic sensitivity in one aspect of the present invention. Figure 4 As shown in boxes 401 and 405, the basic sensitivity is one aspect of the first level of sensitivity. Figure 2 As shown, basic sensitivity has several aspects.

[0201] exist Figure 2 In box 201, computer system 1700 can modify the regression type output to represent it as a perceptible classification type output. The regression type output is a continuous value output from a network or unit, where the output value is a parametric function of the input values, and where the parameters are trained to optimize a specified fit metric between the output of the parametric function and the target values ​​in the training dataset. For example, the regression type could be linear regression, logistic regression, or other suitable regression types.

[0202] In some embodiments, in block 201, the computer system 1700 may replace the continuous-value output with a piecewise constant function. In the typical case where the continuous-valued parameter function is monotonic, the computer system 1700 may replace the parameter function with a step function.

[0203] In some embodiments, in block 201, the computer system 1700 may replace the parameter function with a vector of one or more finite discrete-valued variables. The vector of discrete variables may be referred to as a vector embedding of the continuous-valued function values. The computer system 1700 may compute the vector embedding as a bottleneck layer of an autoencoder. In some embodiments, the computer system 1700 may impose sparsity constraints or regularization on the bottleneck layer. In some embodiments, the computer system 1700 may use an autoencoder with some hybrid parameter control having specified characteristics, such as combining... Figure 9 As discussed. In some embodiments, the computer system 1700 may use such discrete value vector embedding to perform multiple regression on two or more continuous value variables. In some embodiments, the computer system 1700 may use such discrete value vector embedding to perform multiple regression on a continuous value data space.

[0204] Whether it is a piecewise constant function or a vector embedding, the computer system 1700 can train neural networks or hybrid networks to simulate continuous-valued functions or continuous-valued vectors to any desired accuracy, because the computer system 1700 can use continuous-valued functions to compute the target value of an infinite number of input value examples, thus providing an infinite amount of training data.

[0205] However, in some embodiments, the computer system 1700 may limit the number of intervals in the embedded piecewise constant function or discrete vector space in order to better meet sensitivity criteria.

[0206] In box 202, computer system 1700 may replace one or more unbounded variables with bounded variables. For example, computer system 1700 may replace one or more unbounded activation functions with bounded activation functions. In some embodiments, computer system 1700 may simply impose the minimum and maximum values ​​of the activation function's output as constraints. In some embodiments, computer system 1700 may replace the activation function with a new activation function that asymptotically approaches the limit value, which computer system 1700 may change to a step function later in training. In some embodiments, computer system 1700 may restrict values ​​in the global or local data space. In some embodiments, computer system 1700 may restrict values ​​stored in and / or transferred by cells. In some embodiments, computer system 1700 may restrict variable values ​​in the local data space.

[0207] In some embodiments, for a trained or partially trained network, the computer system 1700 may use the minimum and maximum values ​​of node activations observed in the training data to set the limits of the bounded activation function of the nodes, which may leave some extra margin for values ​​that new data may require.

[0208] In some embodiments, the computer system 1700 may implement a semi-automated process and provide human consultation on controlled quantities to specify or verify limits, such as in conjunction with Figure 4 The discussion is in box 414. In some embodiments, the computer system 1700 can use experience-based training (…). Figure 5 The limit is determined by 521).

[0209] In some embodiments, the computer system 1700 may replace nodes or cells with unbounded activation functions with one or two detectors or discriminators, as discussed in conjunction with blocks 211, 212 and 213.

[0210] In box 203, computer system 1700 may replace the activation function of each of one or more nodes that have a non-monotonic activation function with a monotonic activation function or a modified monotonic function. For example, computer system 1700 may specify an activation function that is monotonic over a specified inner interval, rather than monotonic over the entire domain of the activation function.

[0211] In some embodiments, the computer system 1700 may specify a non-monotonic activation function that is monotonic within a specified inner interval, but is modified outside the specified interval. For example, for the activation function that the computer system 1700 represents as a discriminator between sets S1 and S2, the computer system 1700 may specify an activation function having a maximum activation value corresponding to a node in the probability distribution of set S2 and a minimum activation value corresponding to a node in the probability distribution of set S1. The computer system 1700 may specify an activation function that is monotonic within the interval between the minimum and maximum values.

[0212] However, for example, if the nodes of set S1 or set S2 are located inside the data space, the computer system 1700 can specify an activation function that has a local maximum for S2 and a local minimum for S1. In some embodiments, the computer system 1700 can specify an activation function that is equal to or asymptotically approximates a specified out-of-domain background value outside the monotonic interval between the minimum and maximum values, for example, in data exclusion. Figure 2 204 Figure 5 518 and Figure 11 In some embodiments, computer system 1700 may specify an activation function that is monotonic between a background value and a minimum or maximum value. An activation function that is monotonic on the interval between a unique minimum and a unique maximum value and is monotonic outside of that interval is referred to herein as a "standard discriminator function". In a standard discriminator function, the minimum or maximum value may occur at the endpoints (or the limit of infinity), so the monotonic interval can be the entire domain or a half-open interval.

[0213] In some embodiments, in block 203, computer system 1700 may convert the activation of any node that computer system 1700 represents as a discriminator into a standard discriminator function.

[0214] For a node with a standard discriminator function and a specified threshold T between the minimum and maximum values, computer system 1700 can determine whether the node has committed an implicit error on a specific data item.

[0215] Implicit Error: In some embodiments, for a node having a standard discriminator activation function f(x) and a specified discriminator threshold T, where x(d) is a function of input data d, the computer system 1700 may specify an implicit error in the activation value x(d) of the node within the interval between the minimum and maximum values ​​if the sign of (x(d))*(f'(x(d)-T)) is the same as the sign of the backpropagation derivative of the error measurement objective function to be minimized. In some embodiments, the computer system 1700 may reverse the sign test for activations outside the interval between the minimum and maximum values. In some embodiments, the computer system 1700 may not test delegated or excluded data. If the derivative is the objective function to be maximized, the computer system 1700 reverses the sign test.

[0216] In some embodiments, if the magnitude of |T-act(d)| is less than a specified multiple of the magnitude of the backpropagation derivative, the computer system 1700 may determine that the node has made a termination call on data item d, where "act(d)" represents the node's activation value for data d. The termination call can be a termination call with an implicit error or a termination call with an implicit correct answer.

[0217] In some embodiments, the computer system 1700 may add a regularization penalty to the derivative of backpropagation, such as knowledge-sharing regularization, soft binding, and anti-binding, when determining whether a node with a standard discriminative activation function has committed an implicit error. Soft binding is described in U.S. Patent 10,839,294, entitled "Soft-tying nodes of a neural network," and anti-binding is described in U.S. Patent 11,151,455, entitled "Counter-tying nodes of a nodal network." The entire contents of these two patents are incorporated herein by reference. Data-related node-to-node knowledge sharing by regularization is described in published PCT application WO / 2021 / 194516 A1, entitled "Data-related node-to-node knowledge sharing by regularization in deep learning," the entire contents of which are also incorporated herein by reference.

[0218] In some embodiments, if a node is being trained to a known set and the activation x(d) of data item d is on the error side of the discrimination threshold T, the computer system 1700 can determine that an explicit error has occurred in the node. In some embodiments, when such an explicit error criterion is known, the computer system 1700 can use an explicit error criterion instead of an implicit error criterion.

[0219] In some embodiments, the computer system 1700 may ignore relatively small deviations from monotonicity, such as the decrease in Gaussian error linear unit (GELU). Those skilled in the art of neural networks are familiar with the GELU activation function. In some embodiments, the computer system 1700 may use a replacement activation function that is monotonic except for a specified decrease, such as in the GELU function. In some embodiments, for a detector unit, the computer system 1700 may use a center-around function, wherein the function has a decrease in value when activating near but not in the acceptor region. The computer system 1700 may make the function value in this decrease smaller than the function value of activations farther from the acceptor region and the value in the acceptor region.

[0220] In some embodiments, the computer system 1700 may divide the domain of a node having a non-monotonic activation function into alternating intervals of monotonically increasing and monotonically decreasing values. In some embodiments, the computer system 1700 may create a new node for each interval.

[0221] In some embodiments, the computer system 1700 may create nodes for each pair of monotonically increasing intervals, followed by monotonically decreasing intervals, to create one or more nodes with a unimodal activation function. In some embodiments, the computer system 1700 may replace nodes with unimodal activation functions with robust template cells, such as... Figure 10 As shown.

[0222] In some embodiments, the computer system 1700 may replace an activation function with multiple local maxima with multiple robust template units.

[0223] In some embodiments, the computer system 1700 may divide the domain of a discriminator node into a first interval and a second interval, where a local minimum in the activation function in the first interval indicates the detection of a first target set, and a local maximum in the activation function in the second interval indicates the detection of a second target function. In some embodiments, the computer system 1700 may create a first interval for local maximums and a second interval for local minimums. In some embodiments, the computer system 1700 may replace the discriminator node with a unit comprising a detector for the first target set, a detector for the second target set, and elements for calculating a discrimination score based on the scores of the two detectors. In some embodiments, for each target set, the computer system 1700 may train a template model as a detector for the target set.

[0224] In some embodiments, the computer system 1700 can create cells in which nodes with non-monotonic activation functions are replaced by cells with multiple monotonic or unimodal activation functions, separating the computation of the affine sum of the input from the computation of the activation functions, with data exchange in between. In some embodiments, the computer system 1700 can switch any incoming data item to a monotonic or unimodal activation function corresponding to the interval of the incoming data item. Such a structure within the cell is as follows: Figure 3A As shown.

[0225] In some embodiments, the computer system 1700 may replace a node with a non-monotonic activation function with a set of nodes, wherein the activation function of each node is constant outside a specified interval and monotonic or unimodal within the interval. In some embodiments, the computer system 1700 may initialize the incoming connections to each node to replicate the incoming links of the node being replaced. In some embodiments, the computer system 1700 may then train weights on the new connections, which are separate from the weights of the connections to the original nodes. In some embodiments, the computer system 1700 may attach or soft-attach one or more weights to the respective connections.

[0226] In block 204, in some embodiments, computer system 1700 can implement data exclusion and / or data delegation for detector elements and discriminator elements. In some embodiments, computer system 1700 can implement data trimming, limiting detection areas, and / or data exclusion. In some embodiments, computer system 1700 can adjust the limitations of data delegation, data exclusion, and / or trimming based on empirical training. Figure 5 (521).

[0227] In some embodiments, the computer system 1700 can improve the performance of the element by using data delegation to restrict training to an appropriate subset of the training data.

[0228] In basic statistical analysis, data items may be removed from the training data as outliers. In robust statistics, a large portion of the data may be removed from the training parameter probability distribution based on sufficient statistics of the parameters. Typically, when training a neural network, for each training data item, a feedforward computation is performed, which computes the activation of each node in the network, and a backpropagation of the derivative is performed to update each connection to each node.

[0229] However, in large neural networks or large hybrid networks, the situation is more complex. The input of a given data item that one node receives from another node may change as the weights in the network are updated during training. Whether a data item is an outlier of the first node may also change.

[0230] In some embodiments of the present invention, the computer system 1700 can build redundancy in the network so that data items that are no longer outliers of the first node do not necessarily degrade performance.

[0231] In block 205, in some embodiments, computer system 1700 may replace the activation function of one or more selected nodes with an activation function whose value changes less over one or more selected intervals than the replaced activation function. In some embodiments, computer system 1700 may make such a change in the activation function to continue training the selected nodes via backpropagation of the derivative, but later in the training process, the activation function may be changed to a piecewise constant function, as described in conjunction with block 206.

[0232] In block 206, in some embodiments, computer system 1700 may change the activation function of one or more selected nodes to a piecewise constant function. Preferably, computer system 1700 specifies a piecewise constant function that satisfies a specified criterion approximating the selected function being replaced. For example, for each constant interval in the piecewise constant function, computer system 1700 may set the value of the piecewise constant function to the value of the selected function averaged over the interval. In a preferred embodiment, computer system 1700 may replace a monotonic activation function or a monotonic interval in any function with a monotonic step function.

[0233] In some embodiments, the computer system 1700 may make the value of a piecewise constant function within a specified interval a hyperparameter, and the computer system 1700 may change this value during training. In some embodiments, the computer system 1700 may make the value of a piecewise constant activation function a learning parameter, and the computer system 1700 may train it using a hybrid training method, such as combining... Figure 5 The parameters discussed, for example, can be trained using empirical training in a computer system 1700.

[0234] In block 207, in some embodiments, computer system 1700 may specify alternative derivative functions for nodes. Figure 3C An example of an alternative derivative function is shown.

[0235] In box 208, computer system 1700 can replace a selected node with multiple nodes. An example is discussed in conjunction with box 203. Computer system 1700 can replace a node with a non-monotonic activation function, in which each monotonic interval corresponds to a node, with a set of nodes.

[0236] As another example, in box 208, computer system 1700 may replace nodes with two or more nodes or with units comprising two or more nodes. For example, if the range of the activation function of a specified node is associated with a known set, computer system 1700 may create a unit with two or more output values, as well as a node trained to detect data items in the known set and a second node trained to detect data items not in the known set.

[0237] In some embodiments, the computer system 1700 may replace the node that distinguishes the two known sets with two new nodes, or add two new nodes, one of which is trained to detect one of the known sets and the second node is trained to detect the second known set.

[0238] In each of the embodiments in which the computer system 1700 creates two new detector nodes, the computer system 1700 may create a unit that includes both new detector nodes and includes one or both of the two new nodes. The computer system 1700 may create an additional node to detect data items that are not in either of the two known sets, and create a second additional node to directly detect data items in the intersection of the two known sets.

[0239] Please note that nodes trained directly on the task of detecting data items at the intersection of two sets or their complements may not necessarily produce results consistent with those of individual detectors. This is because, generally, each detector has a non-zero error rate, and the error profile may differ depending on the objective. Furthermore, in some embodiments, the computer system 1700 may train new detectors with a different trade-off between precision and recall than known set detectors. In any case, both new detectors provide separate outputs to the unit, indicating directly to higher-level nodes and units in the hybrid network whether data items near the decision boundary between the two detectors are equally good matches for both detectors (referred to here as the "BOTH" detector) or equally poor matches for both detectors (referred to here as the "NEITHER" detector). The computer system 1700 can use the BOTH or NEITHER indication as a useful distinction for higher-level nodes or units receiving connections from the discriminator unit. This information is not available from the output of a single-node discriminator.

[0240] As another example, in box 208, the computer system 1700 can replace a node with two or more nodes or a unit consisting of two nodes, where one new node is trained to detect a known set, and a second new node is trained to recognize a different known set. In some embodiments, the computer system 1700 can add a third node, which includes incoming connections from the two detector nodes, and optional additional incoming connections. The computer system 1700 can train the third node as a discriminator between the two known sets. For example, the activation of the third node can include the difference between the scores of the two detector nodes, or a smooth monotonic function of the difference in scores between the two nodes. The two detector nodes can be newly created nodes, which the computer system 1700 can initialize from two intervals of the replaced node. The computer system 1700 can also train the unit or three-node discriminator to distinguish between the two known sets.

[0241] In box 208, computer system 1700 can also replace nodes with a monotonic activation function and one or more class feature intervals. A "class feature" interval is an interval in which the maximum value is greater than the minimum value; for example, HNLMS has determined that replacing an interval with a constant would cause a performance degradation exceeding a specified amount. A class feature interval can include the entire range of a node, in which case the node can be referred to as a "feature" node.

[0242] In some embodiments, the computer system 1700 may regard extreme values ​​near the end of the class feature interval and / or values ​​outside the extreme values ​​of the class feature interval as detectors.

[0243] In this context, in some embodiments, a computer system 1700 controlled, for example, by HNLMS, can select one or more of several options to process the class feature range:

[0244] (1) The computer system 1700 may replace class feature intervals with units including one or more of the following detectors, which are preferably located in conjunction with... Figure 2 The meaning associated with box 212 is sensitive:

[0245] a. A sensitive detector is used for each extreme value in the class feature interval.

[0246] b. Sensitive detectors are used to detect data items in the “boundary region” where it is unclear which of the two extreme detectors (if any) correctly detected or rejected the data.

[0247] c. Two sensitive detectors are used to distinguish between extreme detector uncertainties for two reasons:

[0248] Both detectors scored above the specified value.

[0249] Neither detector scored above the specified value.

[0250] d. Two or more sensitive detectors are used to detect clusters within the boundary region.

[0251] (2) The computer system 1700 can replace the node with multiple nodes to segment the class feature interval.

[0252] a. The computer system 1700 can create two or more sensitive detectors to detect clusters in a specified interval within an activation function.

[0253] (3) The computer system 1700 can replace nodes with multiple step functions that have different constant intervals, such as Figure 2 Frame 211 Figure 4Box 416 and Figure 8 The discussion is in box 803.

[0254] As another example, in box 208, in some embodiments, the computer system 1700 may replace a single node with multiple nodes to achieve redundancy. In this example, the computer system 1700 may initialize each of the multiple new nodes so that it has the same connections and the same connection weights as the single node being replaced. The computer system 1700 may then train the network including the multiple new nodes, allowing the connection weights entering and leaving each node replica to be offset from each other. In some embodiments, the computer system 1700 may apply regularization, such as unbinding or non-regularized links, to diversify node activations and weight training.

[0255] In box 209, computer system 1700 can use multiple activation functions and data exchange (such as...). Figure 3B The data exchange 325 in the figure replaces a single activation function to select which activation function to use for a specific data item. In some embodiments, the computer system 1700 may create a node for each activation function and create a data exchange, as shown in 342 in Figure 3, to select between two nodes. For example, HNLMS may specify that the computer system 1700 makes such a replacement for any of the following reasons:

[0256] (1) Assign new nodes or activation functions to detect known sets associated with one or more new nodes, in order to simulate the original nodes obtaining data that is not in the associated known sets.

[0257] (2) Delegating one or more problematic data items. The HNLMS or computer system 1700 can delegate specified data to a first node or activation function by controlling data exchange, so that activation from input of the specified data item fails to activate the first node or activation function. In some embodiments, the computer system 1700 or HNLMS can control data exchange to send the data item to a specified second node. In some embodiments, the computer system 1700 or HNLMS can create a new node to receive the data item.

[0258] (3) Exclude data based on basic sensitivity standards:

[0259] a. The computer system 1700 can exclude data based on distances to a specified center point, measured by a specified norm defined in a local data space. For example, HNLMS can specify features for a hybrid parameter-controlled autoencoder to create a local data space.

[0260] (4) For active defense, such as combining Figure 4 Box 416 and Figure 8 The discussion is in box 803.

[0261] In box 210, computer system 1700 can add additional nodes or cells to the network to improve classification performance.

[0262] In some embodiments, the computer system 1700 may add error prediction nodes and error correction nodes to correct one or more explicit or implicit errors. In some embodiments, the computer system 1700 may interpret activation of a first node within a specified interval as accepting or rejecting a received data item belonging to a specified known set. In some embodiments, the computer system 1700 may train a second node to predict whether the first node has made a false positive error, and may train a third node to predict whether the first node has made a false negative error. In some embodiments, the computer system 1700 may create additional nodes or cells called error correction elements that replace changes in the output of the first node when one of the error prediction nodes predicts an error on a received data item. In some embodiments, the computer system 1700 may add the output of the error prediction node as an additional output value to a cell including the first node. Error prediction nodes are also called judgment nodes and are described in published U.S. Patent Application 2022 / 0335296 entitled “Deep learning with judgment,” the entire contents of which are incorporated herein by reference.

[0263] In some embodiments, if a node’s activation value falls within the range that the computer system 1700 interprets as an acceptance or rejection of a received data item in a known set, and the computer system 1700 knows that the acceptance or rejection is false for that received data item, then the computer system 1700 can determine that the node has an explicit error.

[0264] In some embodiments, the computer system 1700 may add one or more nodes to receive data delegations for one or more data items in which an explicit or implicit error has occurred in a node or cell.

[0265] In some embodiments, the computer system 1700 may add one or more nodes to represent clusters in a known or named set. In some embodiments, the computer system 1700 may add one or more nodes to detect clusters in a specified target set. In some embodiments, the computer system 1700 may determine whether cluster modeling is necessary based on analysis of multiple local maxima in a smooth histogram function, such as... Figure 15 As shown in box 1509.

[0266] In some embodiments, the computer system 1700 may add one or more nodes to represent clusters in the complement of the detected set. The complement of the detected set may be more diverse than the detected set. In some embodiments, the computer system 1700 may represent the set of complements by multiple clusters to represent diversity in the data.

[0267] In some embodiments, the computer system 1700 may add one or more nodes to support continuous lifelong learning. For example, the computer system 1700 may add one or more nodes to detect and / or identify new data encountered by the system during continued use.

[0268] Upon receiving items to be classified, the computer system 1700 can add additional nodes to the proactive defense mechanism. Proactive defense combined with... Figure 4 Box 416 and Figure 8 Let’s discuss box 803.

[0269] In block 211, in some embodiments, the computer system 1700 may divide the domain of the activation function into intervals. In some embodiments, the computer system 1700 may replace the activation function with an activation function that satisfies a specified criterion of flatness over each of a specified set of intervals. For example, the computer system 1700 may specify that the difference between the maximum and minimum values ​​of the activation function is less than a specified value. In some embodiments, the computer system 1700 may specify that the activation function is constant within selected intervals. In some embodiments, the computer system 1700 may select all intervals in the activation function partition to satisfy a specific flatness criterion. In some embodiments, the computer system 1700 may specify that the activation function is a piecewise constant.

[0270] In block 211, in some embodiments, computer system 1700 may create two or more partitions of the activation function. In some embodiments, computer system 1700 may define partitions such that the endpoints of some or all intervals in one partition are offset from the endpoints in one or more other partitions. In some embodiments, for each partition, computer system 1700 may specify an activation function that satisfies an interval flatness condition for that partition. In some embodiments, computer system 1700 may create a hybrid node with multiple activation functions, each partition having an activation function and a data exchange, such as... Figure 3B 325. In some embodiments, the computer system 1700 may create multiple nodes, each node having a different one of a plurality of activation functions and having data exchange capabilities, such as... Figure 3B 342 or 362.

[0271] In some embodiments, the computer system 1700 can control data exchange 325, 342, or 362 based on the relative positions of the input values ​​of data exchanged for data items and the start and end points of relevant intervals in the corresponding partitions. In some embodiments, the computer system 1700 can control data exchange as an active defense, such as in conjunction with... Figure 4 Box 416 and Figure 8 The discussion is in box 803.

[0272] In block 212, in some embodiments, the computer system 1700 may replace the detector with a more sensitive detector. In some embodiments, in block 212, the computer system 1700 may replace the selected detector with a piecewise constant function, preferably excluding some data; both of these properties contribute to improved sensitivity.

[0273] In boxes 206 or 212, the computer system 1700 may have replaced the activation function with a piecewise constant function. The piecewise constant function helps the computer system 1700 make the network more intelligent. However, the piecewise constant activation function requires special training techniques, such as alternative derivative functions (…). Figure 2 207 and Figure 5 509), mixed training ( Figure 4 407), selective training ( Figure 4 409), data backpropagation ( Figure 5 510), simulation ( Figure 5 511) and / or mixed-condition training ( Figure 13 and Figure 5 (Frame 512).

[0274] Data exclusion and Figure 11 Related.

[0275] However, for the detector node, the computer system 1700 may take a different approach in some embodiments.

[0276] In some embodiments, the computer system 1700 in block 203 may replace the non-monotonic bounded activation function in block 202 with a bounded monotonic activation function. However, in some embodiments, for a detector node, the computer system 1700 may determine that a non-monotonic activation function with a single node is a more realistic model of a set of target data items.

[0277] In some embodiments, in block 212, computer system 1700 may compute a histogram of the input to the activation function. In some embodiments, computer system 1700 may compute a smoothing function approximation of the histogram. In some embodiments, if a single local maximum exists in the smoothing histogram function, or if a local maximum is larger than the others by at least a specified criterion, computer system 1700 may model the data as a unimodal probability distribution.

[0278] In some embodiments, if multiple local maxima exist in the smoothing histogram function, the computer system 1700 may temporarily divide the domain of the activation function into intervals, each interval containing new nodes, and assign each data item to the corresponding new node for data exchange based on the selected intervals. In some embodiments, the computer system 1700 may use statistical training techniques such as maximum likelihood estimation to train the unimodal parameter probability distributions of the original nodes and each of the multiple new nodes. In some embodiments, the computer system 1700 may train a parameter template model, such as... Figure 10 As shown. In some embodiments, the parametric template model may include parameters equivalent to those of the parametric probability model. In some embodiments, the parametric template model may include additional parameters or hyperparameters, such as restrictions on one or more exclusion norms. In some embodiments, the computer system 1700 may use statistical training methods such as maximum likelihood to estimate the template parameters. In some embodiments, the computer system 1700 may use empirical training to train the template parameters ( Figure 5 (See box 521). In some embodiments, the computer system 1700 may use gradient descent to train some parameters of the template. In some embodiments, some parameters may be specified as hyperparameters controlled by, for example, HNLMS. In some embodiments, the computer system 1700 may specify a local data space for the input values ​​of the detector template. In some embodiments, the computer system 1700 may compute a weighted norm in the local data space.

[0279] In some embodiments, the computer system 1700 may then test a comparison of the performance of a single-node system with that of a multi-node system. In some embodiments, the computer system 1700 may evaluate the performance of the single-node and multi-node systems based on measurements of precision and recall when detecting a specified set of targets, preferably based on data retained from the training data. In some embodiments, the computer system 1700 may evaluate performance based on bias or other precision measures of a system or subsystem including selected elements or their replacements.

[0280] In some embodiments, if one or more detectors in a multi-node version have multiple nodes in their smooth histogram function, the computer system 1700 may repeat the process of partitioning the detector domain.

[0281] In some embodiments, the computer system 1700 may impose data exclusion restrictions on the input and output values ​​of the parametric probability model or template model, such as... Figure 10 The rings 1002, 1003, 1004, and 1010 are shown in the diagram. In some embodiments, the computer system 1700 may use a “center-around” detection score, which is lower for data items that are close to but outside the acceptance distance than for data items that are far from the center point.

[0282] In some embodiments, the computer system 1700 may use a flatter function, such as trimming to a standard deviation or a smaller superGaussian function, for data within the acceptance norm, while using, for example, a function such as... Training with the alternative derivative function of the norm, such as combining Figure 2 The discussion is in box 207. In some embodiments, the computer system 1700 may use a constant acceptance score while training with an alternative derivative function.

[0283] In block 213, in some embodiments, computer system 1700 can create a perceptible discriminator. For example, computer system 1700 can replace the discriminator with two sensitive detectors and replace the combined node with connection weights and activation functions, through which computer system 1700 can calculate an approximation of the difference or ratio of the two detection scores.

[0284] In box 214, computer system 1700 may train nodes or cells to simulate one or more known sets. In some embodiments, computer system 1700 may train nodes such that their activation values ​​are specified to be higher or lower than a specified threshold for data items in the known sets, and that activation values ​​for data items not in the known sets are located on the opposite side of the specified threshold. In some embodiments, for two or more known sets, computer system 1700 may train nodes or cells to have specified values ​​that are higher or lower than the specified threshold for one or more known sets, and located on the opposite side of the specified threshold for one or more other known sets.

[0285] In box 215, computer system 1700 can convert a node into a cell. This cell may have multiple output values. This cell can store one or more values. In some embodiments, computer system 1700 can pass a value to be stored by the cell from the node's activation value. In some embodiments, computer system 1700 can pass a value from another cell to a specific cell for storage in that specific cell. In some embodiments, computer system 1700 can pass values ​​representing node attributes stored in the cell associated with the node. Attributes and Figure 4Box 412 is discussed in connection with this. In some embodiments, the computer system 1700 may store values ​​inferred from the state space probability estimate computed by the computer system 1700 in cells, where a set of cells represents the hidden state space. The hidden state space will be combined with... Figure 7 Let's have a discussion.

[0286] Figure 3A This is a schematic diagram illustrating an example of unit 301. Figure 3A It also includes some external components, including cells 313 and 314, nodes 316, 317 and 318, and an autoencoder bottleneck layer 319 with hybrid parameter control. Figure 3A It also includes components within unit 301, including cell 312, node 315, components of hybrid network nodes (302, 303, 304, 305, 306, and 307), and components of the template model. A unit can include an unlimited number of nodes, cells, template models, and other units.

[0287] exist Figure 3A The explanatory unit also includes a robust template model, which comprises input variable norm cells 309, 310, and 311, bias cell 320, and template summation cell 308. Each of cells 309, 310, and 311 computes a univariate norm of the following form. ,in These can be learning parameters or hyperparameters specified by HNLMS.

[0288] The norm p is a hyperparameter, for example, specified by HNLMS. In some embodiments, the computer system 1700 can estimate it through empirical training. value( Figure 5 (521). For networks not optimized for sensitivity, a typical value for p is 1 or 2. A larger p value is preferred for a flatter response and higher sensitivity. In some embodiments, the computer system 1700 may change the value of p during training, for example, as specified by the system design and / or HNLMS.

[0289] In template summation cell 308, computer system 1700 can calculate...

[0290]

[0291] Where S is, for example, a set of scaling hyperparameters set by HNLMS. In some embodiments, the output of cell 308 may be -g(x) or exp(-g(x)). These can be learning parameters or, for example, hyperparameters specified by HNLMS. In some embodiments, all All are set to 1.0. In some embodiments, the computer system 1700 can train the values ​​through empirical training. and deviation 320 ( Figure 5 (521). In some embodiments, the computer system 1700 can train values ​​using the maximum likelihood of a parametric probability distribution model. The deviation is 320. Figure 3A In the template input, the weights of the input connections are written as... Instead of the more traditional w k To avoid connection weights w with normal nodes connected in component 302 k Confusing. Figure 10 A more detailed diagram of the template is shown, indicating that w k Value (corresponding to) Figure 3A In The value can be estimated as the propagation metric. The reciprocal of.

[0292] The internal components of unit 301 also include internal components of a hybrid network node, including multiple activation functions 305, 306, and 307, data exchange 304, and an element 302 that calculates a weighted sum of input values ​​and biases 303. The input values ​​include the output value of node 316 multiplied by connection weight w1, the output value of node 317 multiplied by connection weight w2, and the output value of node 318 multiplied by connection weight w. k And deviation 303. Solid arrows indicate directional connections, such as connections between nodes in a neural network. Dotted-dash arrows indicate data communication links between cells and between cell 312 and node 315. Data communication links can be unidirectional or bidirectional, such as the link between cell 314 and cell 312.

[0293] In a preferred embodiment, the computer system 1700 may also impose data exclusion restrictions on the template summation variable 308 and / or input variables 309, 310, 311. Figure 11 and Figure 5 (See box 518). For example, in some embodiments, if one or more of variables 309, 310, 311, or 308 exceed a specified limit, the computer system 1700 can impose a data exclusion constraint on a template having output 321 by replacing the output with a background score that is constant. In some embodiments, if the norm of the difference between a data item and a specified central data point of the template exceeds a specified limit for a specified norm in data space 319, the computer system 1700 can implement norm-based data exclusion, replacing output 321 with a specified background value. In some embodiments, the computer system 1700 can impose data exclusion constraints during training and deployment. Hybrid network template units with data exclusion constraints include... Figure 10 As shown.

[0294] Each of the input variables 309, 310, and 311 can have an incoming connection from a node or cell, or as shown in the figure, the input variable can receive an incoming connection from the bottleneck layer of a conventional automatic encoder or a hybrid parameter-controlled automatic encoder 319.

[0295] Figure 3B Three embodiments of data exchange that the computer system 1700 can use in active defense are shown. Figure 4 Box 416 and Figure 8 (frame 803).

[0296] Element 322 is an exemplary embodiment of a hybrid element, comprising two activation functions 323 and 324, having outgoing connections to one or more nodes (such as 327). Element 322 also includes a data exchange 325, which selectively forwards the result of the summing element 326 to one of the activation functions 323 or 324. In an exemplary embodiment of active defense ( Figure 4 Box 416 and Figure 8 (in box 803), the computer system 1700 can control the data exchange 325 to select between activation functions 323 and 324 to reduce the vulnerability of 322 to data that may cause imperceptible errors.

[0297] In some embodiments, an element may have more than two activation functions. In some embodiments, the computer system 1700 may include a probabilistic component in its control of the data exchange 325, wherein the probabilistic component may be selected from two or more activation functions that both satisfy a specified sensitivity criterion. In some embodiments, the computer system 1700 may change the selection probability in the data exchange 325 based on the value of the data item being exchanged.

[0298] Element 331 is an illustrative embodiment of a unit including summing element 334, activation function 333, and data exchange 332. In some embodiments of the illustrative embodiment of element 331, the computer system 1700 may control data exchange 332 as part of a more indirect active defense method than the illustrative example of 322.

[0299] In some embodiments, computer system 1700 can control data exchange 332 to control data delegation. Data delegation will be combined with Figure 5 Frame 518 and Figure 11 Let's have a discussion.

[0300] Component 342 is a pure data exchange device that exchanges data stream 341 between nodes 343 and 344. Figure 3ALike other examples, computer system 1700 can use data exchange 342 in active defense. Figure 4 Box 416 and Figure 8 The difference is that the computer system 1700 can directly control the data exchange 342 without binding the data exchange 342 to a specific node. (Block 803) or control data delegation.

[0301] In some embodiments, the computer system 1700 may use only the data exchange 342 to control data flow. For example, the computer system 1700 may use the data exchange 342 to control the distribution of data in a distributed computing system. As another example, the computer system 1700 may use the data exchange 342 to select specific members of a set to categorize specified data items.

[0302] Since the data exchange 342 is not inside the component, the computer system 1700 can use the exemplary embodiment represented by the data exchange 342 in a component of a conventional neural network or a hybrid network, wherein the component is specified to contain only conventional neural network nodes.

[0303] Figure 3C This is an example diagram of an alternative derivative of an activation function. In some embodiments, the computer system 1700 may use the derivative of a function different from the actual activation function in selected nodes of the network as an alternative derivative. Figure 3C In the illustrative example, the alternative derivative is the function represented by the bold dashed double-dot segments 361, 362, and 363, and it is the derivative of the function represented by the dashed double-dot segments 364, 365, and 366.

[0304] exist Figure 3C In the example, the actual activation function is a piecewise constant function, represented by segments 351, 352, 354, 355, and 356. In some embodiments, the computer system 1700 may use this activation function on nodes that are distinguishing between a known set S1 associated with interval 352 and a known set S2 associated with interval 355. In some embodiments, the computer system 1700 may use a step function, as shown in intervals 353 and 354, to represent a lack of definite decision between 352 and 355. In some embodiments, the computer system 1700 may use more steps in intermediate regions. In some embodiments, the computer system 1700 may use a single intermediate step, or may jump directly from 352 to 355 with a single discontinuity.

[0305] While the illustrative example uses a piecewise constant activation function, in some embodiments, the computer system 1700 may use an alternative derivative function instead of any activation function.

[0306] In some embodiments, the computer system 1700 may use a piecewise constant activation function as the activation function for the detector node. For example, in some embodiments, the computer system 1700 may use an activation function with only three segments 354, 355, and 356 to represent the detector node. As another example, for feature variables with ordered values, the computer system 1700 may use a pure step function, such as segments 352, 353, 354, and 355. In any of these cases, in some embodiments, the computer system 1700 may use an alternative derivative function.

[0307] Figure 4 This is a schematic diagram of sensitivity levels and the hierarchical structure of active perceptible classification. For ease of discussion, Figure 4 The dashed boxes 401, 402, and 403 in the diagram place each illustrative technique into the dashed box that best suits that technique. However, the grouping is not absolute. Many techniques may be useful for multiple dashed boxes.

[0308] Dashed box 401 includes illustrative examples of models and processes related to the first level of sensitivity. The first level of sensitivity is the first line of defense against imperceptible errors in hybrid networks. In some embodiments, computer system 1700 may be able to explicitly test whether the system meets the first level of sensitivity.

[0309] Dashed box 402 includes illustrative examples of models and processes related to the second level of sensitivity.

[0310] Dashed box 403 includes illustrative examples of models and processes related to active classification, including classification during deployment and ongoing lifelong learning.

[0311] In box 405, computer system 1700 can use a relatively simple set of first-level sensitivity techniques, combined with Figure 2 This is discussed in relation to “basic sensitivity” techniques. These basic sensitivity techniques may be based on simple criteria that are based on the following properties: (1) the dimensions of multiple variables, and (2) the derivative of the output function with respect to the input. In some embodiments, the computer system 1700 may assess the relationship between these properties and the degree to which components are susceptible to imperceptible errors. In some embodiments, the computer system 1700 may test whether the basic sensitivity has been violated by using simulated adversarial attacks.

[0312] A classifier system violates sensitivity if a tiny change in the input could change a correct classification to an incorrect one. In some cases, tiny changes may be imperceptible or easily overlooked by a human observer or any intelligent animal.

[0313] For example, in image recognition, if the change in each color component of a pixel is equal to or less than the quantization level, the human observer of the digital image can easily ignore or not perceive the change. The maximum value of the magnitude of change of any input variable is called the change vector. Norm. For The input changes, and the function has continuous derivatives. The biggest change is roughly as follows If the number of input variables N is large, then Small changes in the norm can lead to large changes in the output. This property of multivariate functions in high-dimensional spaces is a major source of imperceptible errors in classifier networks.

[0314] Unfortunately, for classifier systems, the number of input variables is a fixed, specified number. Furthermore, in many classification tasks, including image recognition, N can be very large. On the other hand, the number of input variables for a single component may be specified by the system design and / or HNLMS, and may be much smaller than the total number of input variables for the entire system.

[0315] In terms of basic sensitivity, the computer system 1700 focuses on ensuring that each component meets the specified sensitivity standards.

[0316] An example of the sensitivity standard for a single element is as follows:

[0317] 1) The output derivative should be less than the specified magnitude, except for data items within the specified distance of the decision boundary.

[0318] 2) For any interval of the activation function representing the detection, the difference between the maximum and minimum output values ​​should be less than the specified magnitude.

[0319] 3) The difference between the maximum and minimum output values ​​of all data in the "Remote Area" should be less than the specified value:

[0320] a) "Remote area" refers to a specified area from any point in the area to any point in one or more specified detection areas where the minimum distance is greater than the specified standard.

[0321] b) The detection region can be specified by the interval in the activation function or the norm of a specified point in the template detector.

[0322] In box 405, computer system 1700 can modify the activation functions in nodes, add elements to the network, add several special models, and / or make various other changes to the network to better meet several basic sensitivity criteria that computer system 1700 may automatically check. Combined with Figure 2 An example of the modifications made to computer system 1700 in box 405 is discussed.

[0323] For example, in some embodiments, in Figure 2 In box 202, computer system 1700 can change the unbounded activation function to a bounded activation function to better meet the above criteria (3).

[0324] In some embodiments, Figure 2 In box 204, one of the reasons why computer system 1700 can exclude data is to better meet the above criteria (3).

[0325] In some embodiments, the computer system 1700 may change the activation function so that it... Figure 2 Box 205 has a flatter interval and / or Figure 2 Box 206 contains a piecewise constant interval to better satisfy the above criteria (1) and (2). In some embodiments, the computer system 1700 may use an alternative derivative function to accelerate the training process, especially after the changes made by the computer system 1700 in boxes 205 and 206 that might otherwise slow down or stop training in backpropagation through the modified elements.

[0326] In some embodiments, the computer system 1700 may be modified in blocks 203, 208, 209, 212, and 213 to better meet basic sensitivity criteria, such as the illustrative examples above.

[0327] In box 406, computer system 1700 may select one or more of several methods to improve the sensitivity of a node having an activation function that includes one or more intervals that do not conform to a flatness criterion, i.e., intervals in which the change in activation function value exceeds a specified limit.

[0328] In some embodiments, the computer system 1700 may first divide the domain of the activation function of a node into intervals. For example, HNLMS may specify rules for dividing the activation function into intervals. For example, in some embodiments, the computer system 1700 may attempt to find one or more intervals that satisfy a specified flatness criterion. The computer system 1700 may then divide the domain into alternating flat and non-flat intervals. In some embodiments, the computer system 1700 may arbitrarily divide the domain into intervals.

[0329] Then, the computer system 1700 can select a non-flat interval, which in some embodiments may be the entire domain of the activation function.

[0330] In some embodiments, the computer system 1700 can divide a selected interval into sub-intervals. Then, the computer system 1700 can create a cell with a separate activation function for each sub-interval, the input of the activation function of the original node being used as data exchange. This data exchange structure, which selects activation from multiple activation functions, is as follows: Figure 3A and 3B As shown. In some embodiments, computer system 1700 can use this structure to divide the activation function into alternating monotonically increasing and monotonically decreasing intervals. In some embodiments, block 406, computer system 1700 can use the same structure in a two-layer arrangement, first dividing the domain of the original activation function into alternating flat and non-flat intervals, and then dividing each non-flat interval into multiple sub-intervals. Computer system 1700 can use other embodiments to achieve similar results.

[0331] Once a non-flat interval is divided into subintervals, the computer system 1700 can approximate the activation function in the subintervals using a function that meets the flatness criterion. In some cases, the computer system 1700 can approximate the activation function on the subintervals using a constant.

[0332] In some embodiments, the computer system 1700 may create separate copies of the subnetwork for selected nodes and train the subnetwork separately for each subinterval. In some embodiments, the computer system 1700 may use knowledge-sharing links with equality relationships to normalize the copies of the subnetwork so that they have activation values ​​similar to the original subnetwork. In some embodiments, the computer system 1700 may use knowledge-sharing links with non-equality relationships to create diversity among multiple copies of the subnetwork.

[0333] In some embodiments, under the guidance of HNLMS, the computer system 1700 may analyze selected nodes as discriminators. For example, if the selected node is an output node of a network or a unit with a defined objective, the computer system 1700 may interpret the node as distinguishing data items from data items in another target set. In some embodiments, if the node is already associated with two known sets, the computer system 1700 may characterize the node as distinguishing between the two known sets. In some embodiments where the selected node is trained by the backpropagation derivative, the computer system 1700 may interpret the selected node as distinguishing between data items with negative backpropagation derivatives and data items with positive backpropagation derivatives.

[0334] If the selected node does not have a bounded monotonic activation function, in some embodiments, the computer system 1700 can... Figure 2In steps 201, 202, and 203, nodes are modified to obtain nodes with bounded monotonous activation functions. For a bounded monotonous activation function, correctly distinguishable data items will be activated at the extrema of the activation function's domain, where the activation function is relatively flat because it is bounded. That is, non-flat intervals will lie in the middle region of the activation function's domain. In other words, data items in non-flat regions are data items that have not yet been correctly distinguished in the current training state. Training each sub-interval individually enables the computer system 1700 to successfully distinguish many data items in each sub-interval.

[0335] In some embodiments, such as those guided by HNLMS, the computer system 1700 can take advantage of this opportunity to improve classification performance. For example, in some embodiments, the computer system 1700 can train sub-intervals with the original non-flat activation function until a stopping criterion is met, and then change the activation function of the sub-intervals to be flatter, while approximating the original activation function.

[0336] In some embodiments, the computer system 1700 may partition the domain of the activation function of a selected node in a variety of different ways. For example, the computer system 1700 may first partition the domain into intervals, and then perform a secondary partitioning of the domain, wherein, except for open intervals at extrema, the boundary of each interval in the second partition is located at the center of the interval in the first partition. In some embodiments, the computer system 1700 may create two or more ways to partition the domain into intervals. In some embodiments, the computer system 1700 may also partition each non-flat interval into sub-intervals in a variety of ways. Two easily confused data items within the same sub-interval in one partition may be within different sub-intervals in another partition. Therefore, cells in different partitions may differ in which easily confused data pairs become distinguishable.

[0337] In some embodiments, the computer system 1700 can use this diversity to improve classification performance, even more so than that achieved using a single partition. In some embodiments, the computer system 1700 can test each partition and select the partition with the best performance. In some embodiments, the computer system 1700 can use a collection of networks with different partitions, such as an integration.

[0338] In some embodiments, the computer system 1700 may use a network set with different partitions for diagnosis and detection, such as Figure 4 As shown in box 415. In some embodiments, the computer system 1700 can use a network set with different partitions for proactive defense, such as combining Figure 4 Box 416 and Figure 8 As explained in box 803.

[0339] In some embodiments, in block 406, computer system 1700 may use different methods to justify non-flat intervals as alternatives to or supplementary sub-intervals. In some embodiments, computer system 1700 may verify that outgoing connections from non-flat nodes or intervals are only connected to robust template models. In some embodiments, computer system 1700 may impose data exclusion restrictions on nodes or cells receiving connections from non-flat nodes or intervals.

[0340] In box 407, in some embodiments, computer system 1700 can perform hybrid training. That is, computer system 1700 can use multiple training techniques, not just gradient descent calculated via derivative backpropagation. Figure 5 Many examples of hybrid training techniques were discussed.

[0341] In box 408, in some embodiments, computer system 1700 coordinates with HNLMS to find the optimal location in the network to integrate selected "knowledge fragments". The selected knowledge fragments may come from external sources or may be knowledge represented in or accompanying network cells and / or nodes. In some embodiments, the knowledge fragment may be in a network repository.

[0342] An example of a "knowledge fragment" is knowing which data items are members of a known set. By definition, a set of data items is a known set only if the computer system 1700 has a way of determining whether a specified data item is in the set. While any subset of the training data items is a known set, preferably, in block 408, the computer system 1700 is able to determine whether a data item not in the training data is in the known set. For example, any set of datasets defined as being accepted by a specified detector node or unit is a known set, and the computer system 1700 can determine whether a specified data item is in the known set by computing the activation of a subnetwork of the detector and observing the output of the detector. In some embodiments, a "knowledge fragment" may relate to two sets distinguished by a discriminator. Without loss of generality, some illustrative examples can be discussed with respect to detector elements. However, in some embodiments, the computer system 1700 may use substantially the same process for discriminator elements.

[0343] In some embodiments, in block 408, for a given knowledge fragment, computer system 1700 may test selected candidate locations in the network to see if integrating the knowledge fragment into the selected network location can improve classification performance, sensitivity, and / or overall interpretability.

[0344] If the knowledge fragment is a detection of a known set, the computer system 1700 can integrate the knowledge fragment in any of a variety of ways. In some embodiments, the computer system 1700 can connect the detector to one or more nodes or cells in the candidate locations.

[0345] In some embodiments, computer system 1700 may create new nodes or units in the existing base network and train the new nodes or units to simulate the detector. During simulation training, the new nodes or units are trained to match the detector output for all specified data items. The specified data items do not need to be labeled. The specified data items do not even need to be real data items. They can be generated or synthesized data items. Computer system 1700 may train new nodes to match the detector output for synthesizing generated data. Computer system 1700 is not limited to using existing subnetworks of candidate locations with new nodes. In some embodiments, utilizing an unlimited amount of potential training data for simulation, computer system 1700 may train entirely new subsystems.

[0346] In some embodiments, the computer system 1700 may test the performance, sensitivity, and / or overall interpretability of each selected candidate location. For example, guided by HNLMS, the computer system 1700 may then select one or more candidate locations and integrate the knowledge fragment into those locations.

[0347] In some embodiments, the computer system 1700 may screen potential candidate locations. For example, in some embodiments, the computer system 1700 may calculate the correlation between the detector's output and the backpropagation derivative of the global or local target of the potential candidate node. This correlation represents the degree to which incremental training updates can improve the target, and this degree is averaged over a dataset measuring the correlation. A high correlation will indicate a good candidate location. If the potential candidate node backpropagates data examples rather than derivatives, the computer system 1700 may calculate the consistency between the backpropagated data examples and the detector's detection and rejection sets. In some embodiments, for example, based on an analysis of the requirements for candidate locations estimated by the computer system 1700 under the guidance of HNLMS, the computer system 1700 may restrict the consistency metric to recall or precision.

[0348] In block 409, in some embodiments, computer system 1700 may selectively train only a subset of elements in the network being trained and / or selectively train elements only on a specified subset of data. Computer system 1700 can use selective training to accelerate or better control hybrid training, which can be applied to any sensitivity level. Figure 4 In this context, selective training is arbitrarily placed in box 401.

[0349] In some embodiments, the computer system 1700 may selectively train an element that distinguishes two related known sets only on data items in the union of the two known sets.

[0350] In some embodiments, the computer system 1700 may selectively train the decision element only on data close to the decision boundary. In some embodiments, the computer system 1700 may change the selection of training data items when the location of the decision boundary changes during training.

[0351] In some embodiments, the computer system 1700 can apply selective training by selecting a subset of elements to be trained for one or more specified data items.

[0352] The selective subset of training elements complements two features of sensitivity-based hybrid learning. In some embodiments, the first feature of hybrid training is that the computer system 1700 continuously modifies the network during training and, in some embodiments, during deployment. In a preferred embodiment, when the computer system 1700 modifies the trained network, it may temporarily focus the training on the modified elements and other elements most affected by the modified elements.

[0353] The second characteristic of hybrid training is that the learning process can be actively controlled, for example, by HNLMS. For instance, an AI system or human team within HNLMS can instruct computer system 1700 to focus on training specific elements. Furthermore, HNLMS can actively monitor the training process and prioritize training on the elements that most need improvement.

[0354] In an illustrative embodiment, the computer system 1700 may maintain a list of elements being trained.

[0355] Due to errors or termination calls to explicit or implicit local or global targets, computer system 1700 may proactively add the element being trained to a list, or add data items to the element's data item list. In some embodiments, errors or termination calls may be directed to data for which the element previously did not have errors or termination calls. In some embodiments, errors or termination calls may be directed to new real data or newly generated or simulated data. Errors or termination calls may occur on data items modified by simulated attacks or other interference.

[0356] In some embodiments, the computer system 1700 may remove elements from the list based on specified criteria.

[0357] In some embodiments, the computer system 1700 may add newly created or modified elements to the list being actively trained.

[0358] In some embodiments, for new or modified elements, computer system 1700 may temporarily suspend training of elements connected to the new or modified elements, as instructed by, for example, HNLMS. In other embodiments, computer system 1700 may activate element training via connections from new or modified elements.

[0359] In box 410, the computer system 1700 can test the sensitivity of the decision boundary, and if necessary, the computer system 1700 can modify the network to move the position of the decision boundary, thereby improving the sensitivity of the decision. For the discussion of box 410, the "decision boundary" is a set of points in the local or global data space where the activation of the discriminators for two target sets is at a specified threshold. Preferably, each target set is a known set. The discriminator can be a node or unit trained as a discriminator, or a new node or unit created by the computer system 1700 by combining the scores of two trained detectors.

[0360] For box 410, the desired objective is to make any data point in the selected normed local or global data space located on or near the decision boundary reasonable to a human observer, serving as a data example on the boundary. A human observer might agree that a data point is reasonable because (1) it matches fairly well with both target sets. In some embodiments, a human observer might agree that a data point is reasonably located on the boundary because (2) it matches poorly with either target set and should not be accepted as an example of either. For the purposes of box 410, in some embodiments, data points that connect data points satisfying reasonableness condition (1) with data points satisfying reasonableness condition (2) to form a smooth surface can also be considered reasonable.

[0361] In some embodiments, in block 410, computer system 1700 may construct and train a conventional neural network with outputs differentiable relative to input values ​​from a global or local data space to simulate a hybrid network discriminator that computer system 1700 is testing and improving decision boundaries. Computer system 1700 may train a neural network or hybrid network to simulate another network using generated or simulated data as well as unlabeled real data. Using the necessary data, computer system 1700 may train the simulated network using necessary unlabeled or generated data until the limits of the simulated network's capabilities are reached. In some embodiments, computer system 1700 may train a simulated neural network having nodes corresponding to each node in the hybrid network being trained, each node in the neural network being trained to simulate the corresponding node in the hybrid network as closely as possible. In a preferred embodiment, computer system 1700 uses at least the same local or global input space as the discriminator being simulated and trains the nodes in the neural network to simulate the discriminator as closely as possible. This simulation cannot be perfect. For example, a simulated neural network with differentiable activation functions can at most approximate the activations of nodes with discontinuous activation functions, and vice versa.

[0362] In some embodiments, the computer system 1700 can minimize by backpropagating to the input data value d. The goal is to find data points on the decision boundary of a traditional neural network with differentiable outputs, where T is the discrimination threshold of the decision boundary and act(x(d)) is the activation of the discriminator node for data item d.

[0363] For this objective, each point on the decision boundary of the simulated neural network will have a zero value. Through many different random starts, the computer system 1700 can find multiple points on the decision boundary of the simulated hybrid network. In some embodiments, the computer system 1700 can locally estimate the tangent hyperplane of the decision boundary of the simulated neural network by fitting a multiple linear regression model to example points on the decision boundary. In some embodiments, the computer system 1700 can then compute an orthogonal line to the estimated decision boundary. In some embodiments, the computer system 1700 can then search along this orthogonal line, for example, by using a binary search, to find points in the data space located on the decision boundary of the hybrid network.

[0364] In some embodiments, the computer system 1700 can test reasonableness by testing consistency. That is, the computer system 1700 can train a diverse set of networks. The computer system 1700 can then measure how much the location of the decision boundary changes from one network to another. If there are significant differences between the networks, the computer system 1700 can use this as a diagnosis that at least some networks have not found reasonable decision boundaries.

[0365] In some embodiments, the computer system 1700 can train a "BOTH" detector and / or a "NEITHER" detector for data points on or near the decision boundary of a hybrid network and / or an analog neural network. The computer system 1700 can train and... Figure 2 Box 208 is associated with the BOTH and / or NEITHER detectors. In some embodiments, the computer system 1700 can assign a constant background score as a unit output value to all data items detected by the NEITHER detector.

[0366] In some embodiments, the computer system 1700 may train a discriminator between sets “BOTH” and “NEITHER”, as well as a detector for each set. In some embodiments, if the discriminator variable associated with the decision boundary includes input from the detector for each alternative, the computer system 1700 may use two detectors above a specified detection threshold as initial indications of data items in the “BOTH” set. In some embodiments, the computer system 1700 may use two detectors below a specified detection threshold as initial indications of data items in the “NEITHER” set. In some embodiments, if the discriminator elements do not yet include such a detector or input from such a detector, the computer system 1700 may train a detector for each discriminated set.

[0367] In some embodiments, the computer system 1700 may use additional indications to distinguish between the "BOTH" set and the "NEITHER" set. For example, in some embodiments, the computer system 1700 may compute a histogram of the data from the union of two sets on or near the decision boundary. The computer system 1700 may then determine whether the histogram is unimodal or bimodal, as in combination with... Figure 15 The discussion is in box 1509. In some embodiments, the computer system 1700 may compute such a histogram of data projected onto lines orthogonal to the hyperplane of the estimated decision boundary. In some embodiments, the computer system 1700 may compute the projection of multiple orthogonal lines onto the orthogonal lines.

[0368] As a second example, the computer system 1700 can calculate the magnitude of the discrimination score derivative of a point on the decision boundary through the data input being evaluated, along a line orthogonal to the decision boundary. A low magnitude of this derivative indicates that the data point is in the "NEITHER" set. A high magnitude indicates that the data point is in the "BOTH" set.

[0369] In some embodiments, the computer system 1700 and HNLMS can create one or more new features to distinguish data items detected by the BOTH detector. For example, in some embodiments, the computer system 1700 can create new features to distinguish the two sets through standard training of the discriminator nodes. In some embodiments, the computer system 1700 can train additional new nodes in the subnet for the new discriminator nodes. As another example, the computer system 1700 can use constrained optimization to train a new discriminator (…). Figure 5 (524).

[0370] In some embodiments, the computer system 1700 may use fractional theory knowledge to refine decision boundaries. In an illustrative embodiment, the computer system 1700 may calculate the alignment of portions of an image with fractional representations of objects in the image or objects assumed to exist in the image. Figure 4 Boxes 413 and 415 and Figure 7 , 12 13 and 19 further discuss the alignment of parts with the specified dichotomy.

[0371] For example, in some embodiments, the computer system 1700 may sample a pair of data items near a decision boundary, the pair of data items coming from each of two known sets whose partition theories contain one or more shared components. In some embodiments, a pair of data items from the same category or named set may share the same partition theory. For any shared partition theory component, the computer system 1700 may align each data item with its partition theory and store the alignment information in a cell in the unit that detects a specified portion of each image, thereby at least partially aligning the two images to each other. Even if the partition theories are different, the computer system 1700 may create and train detectors and / or feature variables to distinguish one or more pairs of aligned portions.

[0372] In some embodiments, the computer system 1700 may project a selected set of data items onto a line computed by the computer system 1700 as orthogonal to the estimated decision boundary of the simulated neural network and / or the estimated decision boundary of the hybrid network. In some embodiments, the computer system 1700 may constrain the selected data items to a specified distance within the orthogonal line. In some embodiments, the computer system 1700 may generate additional data items for each of two distinguished sets. In some embodiments, the computer system 1700 may generate additional data items by random perturbation and / or adversarial attacks on each selected data item. In some embodiments, preferably, the computer system 1700 may augment each selected data item with the same number of generated items. In some embodiments, the computer system 1700 may use a pair of generators to generate additional data items, one generator being trained to generate examples of one of the identified known sets and a second generator being trained to generate examples of a second known set. Generally, the computer system 1700 may use any method to create a proportional number of additional examples of each known set near the decision boundary.

[0373] In some embodiments, the computer system 1700 may then estimate the probability density function for each of the two sets to be identified. In some embodiments, the computer system 1700 may compute a histogram of data item counts based on the position of each selected data item projected onto a line orthogonal to the decision boundary. In some embodiments, the computer system 1700 may estimate a regression function of the difference or ratio of the two estimated density functions. In some embodiments, the computer system 1700 may estimate a Bayesian minimum error split point for the two estimated probability density functions or a smoothed estimate obtained from a smooth approximation of the regression estimate or histogram counts. In some embodiments, the computer system 1700 may use this estimated Bayesian minimum error point as a point on the updated decision boundary.

[0374] In block 411, in some embodiments, computer system 1700 may create a locally normed space. In some embodiments, computer system 1700 may use a neural network autoencoder or an autoencoder with hybrid network parameter control having specified characteristics to create the locally normed space. Figure 9In box 411, the specified features can be engineered features specified and / or computed by, for example, HNLMS. As is well known to those skilled in the art of neural networks, an autoencoder is a network trained to encode each input data item using a restricted encoding, called the "bottleneck" layer of the autoencoder, for example, a vector with a specified finite dimension, for each data example, and then, for a specified training dataset, to produce an output that matches the input as closely as possible. For example, for a local autoencoder, a computer system 1700 or HNLMS can specify a set of nodes as the input data space. For example, the input space of an autoencoder can be a set of nodes connected to nodes or units, such as detector nodes or units or discriminator nodes or units. As another example, the input space can be the union of elements connected to a pair of detectors, a classifier, or a set of two or more detectors. The input space can be the union of input variables or the union of elements connected to a group of decision elements.

[0375] Combination Figure 9 An automatic encoder with hybrid parameter control having specified characteristics is discussed.

[0376] In some embodiments, the computer system 1700 may introduce a local normed space to limit the effective dimensions of the input to one or more detectors and / or discriminators, thereby helping to improve the sensitivity of the detectors and / or discriminators. For example, in some embodiments, the computer system 1700 uses a local normed space in block 410.

[0377] In block 412, in some embodiments, computer system 1700 can manipulate data and perform sequential computations in ways that conventional neural networks cannot represent. In some embodiments, each cell has local memory. In some embodiments, computer system 1700 can perform sequential computations associated with cells before, during, and / or after the activation of computational units and nodes.

[0378] For example, in some embodiments, computer system 1700 may use these cells to calculate attributes and characteristics, as described in the following paragraphs.

[0379] In some embodiments, computer system 1700 may use these cells to implement specialized code developed specifically for domains deploying hybrid networks. This specialized code may represent a process known as "knowledge engineering." In some embodiments, computer system 1700 may use these cells for logical reasoning (…). Figure 21 (2102). In some embodiments, the computer system 1700 may use these cells to represent probabilistic networks, such as hidden Markov processes or dynamic Bayesian networks for probabilistic inference. Figure 21(2102). In some embodiments, the computer system 1700 may use these cells to represent cellular automata. (Combined with...) Figure 19 and 21 These uses of cells for sequential calculations after receiving data items to be categorized are discussed.

[0380] In some embodiments, computer system 1700 may perform knowledge-engineered sequential computations on data stored in cells or in input or output data. For example, if computer system 1700 has generated text, images, or videos, in some embodiments, computer system 1700 may compare the proposed generated output with training data to verify that the proposed output is not sufficiently close to any project data to infringe copyright.

[0381] As another example, in some embodiments, computer system 1700 performs logical or set-theoretic computations on inputs, outputs, and / or data computed within the network. For example, in a text generator, in some embodiments, computer system 1700 may test the logical consistency of the output. For example, computer system 1700 may have program code representing syllogisms such as "If A implies B is true, A is true, then B is true" and "If A is true and B contradicts A, then B is not true." In some embodiments, computer system 1700 may have ontology-based logic, such as "If A is a kind of B, and there exists an instance of A with attribute C, then there exists an instance of B with attribute C."

[0382] As a concrete example of violating ontological usage, a state-of-the-art text generator repeatedly asserts that "a perceptron cannot represent an XOR function," while simultaneously acknowledging that "a basic perceptron can represent an XOR function," and even providing an algorithm to train a basic perceptron to represent an XOR function. This behavior is neither illogical nor perceptible.

[0383] The statement that "a perceptron cannot represent the XOR function" is incorrect, yet widely cited online. Text generators are trained on online text, but one can also cite verbatim from an out-of-print book where Frank Rosenblatt introduced perceptrons and proved that even a basic perceptron can be trained to represent any Boolean function, including the XOR function. Without explicit logical analysis, it's difficult to make a neural network with trillions of learned parameters forget something, even if it logically contradicts other things it has learned. In various embodiments, computer system 1700 can overcome this difficulty by explicitly applying logical reasoning in cells within computations that are separate from and / or cover operations within nodes.

[0384] In some embodiments, the computer system 1700 may store known values, referred to as "attributes," associated with a specified element as variables in a cell. In some embodiments, the computer system 1700 may determine whether to store the attribute associated with the element based on the activation value of the element of the current data item. For example, in some embodiments, for a detector or discriminator element, the computer system 1700 may store only the attribute if the activation value is within a specified interval, such as a detection acceptance interval.

[0385] An example of an attribute is the position of a node in an image within a convolutional network. Another example of an attribute is the orientation of a detected object, such as the rotation angle of a line segment. Other attributes of an object include size, color, and texture. In models based on hierarchical knowledge structures such as volume theory or ontology, elements may have attributes inherited from other elements in the hybrid network. In some embodiments, cells can be programmed to pass attributes via data communication links between cells and between cells and nodes. In some embodiments, the computer system 1700 can control attribute communication based on node activation values ​​and attribute values ​​of the current data item.

[0386] In some embodiments, the computer system 1700 may implement software to compute attributes or features specified by a human team, such as HNLMS. In some embodiments, the computer system 1700 may store the values ​​of human-specified features in cells within specified cells. An example of human-specified features is the estimated frequency of formants in speech analysis. The estimation of formant frequencies is well known to those skilled in the art of speech signal processing. Another example of human-specified features is the explicit detection of edges in an image via a high-pass filter. While convolutional neural networks can detect edges in an image, edge detection in a convolutional neural network is mixed with all other activations of the network nodes. In some embodiments, the computer system 1700 may explicitly label detected edges as edges. Edge detection in an image is well known to those skilled in the art of image digital signal processing. In some embodiments, the computer system 1700 may use the detected edges in a split-theory. In some embodiments, the computer system 1700 may use the detected edges to align an image with a model or another image.

[0387] In some embodiments, the computer system 1700 may specifically design and train new features to improve the discrimination between two known sets. In some embodiments, the computer system 1700 may use such features as specified features in a hybrid parametric control autoencoder with specified features, wherein the bottleneck layer includes new features in variables within a locally normed space. In some embodiments, as part of HNLMS, the computer system 1700 may develop new features to discriminate real or generated data item instances near the decision boundary between two known sets, such as... Figure 4 As described in box 410. For example, computer system 1700 can create and train new features to identify data items from two known sets detected by the BOTH detector, as described in connection with box 410.

[0388] In some embodiments, the computer system 1700 can automatically create new features by training new discriminator nodes to improve the discriminative ability of existing discriminator nodes or units for a specified set of target pairs. In some embodiments, the computer system 1700 can train new feature nodes or units on a selected dataset. In some embodiments, the computer system 1700 can select errors and termination calls of the existing discriminator as training data for the new features. In some embodiments, the computer system 1700 can select data items near the decision boundary of the existing discriminator as training data for the new features.

[0389] In some embodiments, the computer system 1700 may create and train one or more candidate new features, and then test the system's performance in an autoencoder with hybrid parameter control of the specified features, using one or more selected candidate new features added to the specified features. In some embodiments, the computer system 1700 may test the system's comparative sensitivity using the selection of new features and classification performance. For example, in some embodiments, the computer system 1700 may perform one or more simulated adversarial attacks on the system and measure the success rate of the adversarial attacks.

[0390] For example, in some embodiments, the computer system 1700 may sample a pair of data items near a decision boundary, selecting one from each known set. The computer system 1700 may then align each data item with a split theory and store the alignment information in a unit that detects a specified portion of each image, thereby aligning the two images to each other. The computer system 1700 may then create and train detectors and / or feature variables to distinguish the two aligned portions from each other.

[0391] In some embodiments, the computer system 1700 may use attributes as features. In some embodiments, a node may have a known latent attribute implemented for a given data item if the node's activation is within a specified range when the specified data item is used as input to a global or local data space. For example, a node may have a latent positional attribute that is activated when the node's activation value is above a specified threshold.

[0392] For example, in convolutional networks designed for image recognition, each low-level node typically receives activation connections for only a small number of pixels located at or near a specified position in the image. Similarly, in speech recognition systems, nodes receive a series of input vectors, each from a finite time interval. Furthermore, nodes in speech recognition systems may receive values ​​for only a single frequency or a finite frequency range. The position of the input received by a node in convolutional image recognition is a constant and does not change with the input data item. However, in some embodiments, the computer system 1700 may store the position of the detector node activated above a specified detection threshold in the position attribute cell as an attribute of the current data item. Similarly, in a speech recognition system, the computer system 1700 may store the time and frequency positions of the detector node activated above a specified detection threshold in the time-frequency attribute cell.

[0393] In some embodiments, when one or more nodes associated with an attribute cell are activated above a specified minimum threshold, the computer system 1700 may set the attribute value in the cell to a known attribute of the associated cell with the highest activation level. Such an attribute is not explicitly represented in node activation and therefore cannot be used by higher-level nodes via network connections. However, depending on the system design or HNLMS specifications, the computer system 1700 may, for example, store the attribute in the cell and create data links from that cell to other cells and / or other nodes in the network. In a network representing a body theory, at higher-level nodes or cells, the computer system 1700 may match two or more attributes (e.g., the position of a related part in the body theory) to a trained model of the relative values ​​of attributes in a specified object image. In some embodiments, the computer system 1700 may scale the position values ​​of object components based on the size of the object seen in the image.

[0394] In some embodiments, the computer system 1700 may use cells to store state information in state-space modeling. Figure 7 and Figure 4 (See box 413). In some embodiments, the computer system 1700 can perform state-space analysis on data items, thereby altering the behavior of the system after receiving and classifying data items.

[0395] In some embodiments, the computer system 1700 may use cells to calculate the active alignment of data items. Figure 12 and Figure 4 (Box 417) changes the system's behavior after receiving and classifying data items.

[0396] Changing the system's behavior after receiving data items can help computer systems 1700 make them more robust against attacks and other interference that could lead to imperceptible errors.

[0397] When using cells for active alignment and / or other analyses related to fractional theory and other human knowledge representations, the computer system 1700 can make the system easier to understand and can facilitate interaction with HNLMS and other human consultations.

[0398] For example, in some embodiments, as part of HNLMS, computer system 1700 can train a model of attribute combinations while training the weights and biases of network connections.

[0399] In box 413, computer system 1700 can construct one or more hidden state space models.

[0400] In classification tasks where input data variables can be organized by time and / or spatial location, computer system 1700 can add cells to a network connected to a structure representing the relative positional geometry of the input variables. More generally, computer system 1700 can construct structures between cells in the network to represent any adjacency graph between input variables. In some embodiments, at higher layers of a hybrid network, computer system 1700 can construct adjacency graphs between higher-level sets of cells. In each cell, at each layer, computer system 1700 can store the value of one or more hidden variables. In some embodiments, higher-level cells may have the same adjacency graph as lower-level cells but with different or additional hidden variables.

[0401] In some embodiments, in box 413, computer system 1700 may implement probabilistic inference or dynamic Bayesian networks in the cells of the network. Figure 21 (2102 in the middle).

[0402] Hidden state space model combined Figure 7 To explain.

[0403] In block 414, computer system 1700 can manage options for human consultation in many aspects of the invention. In a preferred embodiment, computer system 1700 can manage human consultation to maximize the amount of improvement required per unit of human time and labor. In some embodiments, computer system 1700 can semi-automate a process that would otherwise require human knowledge engineering with specialized expertise, and the workload increases with the size and complexity of the network. Other aspects of communication between computer system 1700 and one or more individuals will be combined. Figure 21 Let's have a discussion.

[0404] Several examples of aspects of the invention exist in which the computer system 1700 can efficiently and effectively manage human consultation. In some embodiments, the computer system 1700 can provide information to human team members of HNLMS and / or users of the system so that humans can initiate human consultation processes.

[0405] An example of human consultation that the computer system 1700 or a human can initiate is the naming of known sets. The computer system 1700 can request a human to provide a human-understandable name for a known set, and the computer system 1700 can provide examples of such a set. In a preferred embodiment, the computer system 1700 can manage the efficiency of this process by only asking for the names of known sets associated with elements that play an important role in the hybrid network, which has been trained to meet specified criteria. In some embodiments, a human can voluntarily name any known set or any variable at any time, at the discretion of the voluntarily naming human. For example, a human can voluntarily provide a name if a human consultant believes that a name will enable the computer system 1700 to guide training to learn concepts that can better generalize to new data. A human can also voluntarily provide a name as long as the human believes that the provided name will effectively improve the overall interpretability of the hybrid network.

[0406] In some embodiments, when associating sets with elements being actively trained, the computer system 1700 may preferentially associate elements with named sets rather than with unnamed known sets. This preference may help meet human expectations that naming sets will help improve the network's generalization performance. This preference will also improve the overall interpretability of any elements associated with named sets.

[0407] For example, human team members in computer systems 1700 or HNLMS can initiate human consultation when defining the initial state space of the hidden state space model, such as... Figure 7As shown. In some embodiments, the computer system 1700 can largely automate future changes in the state space. However, further human consultation can be initiated by either the computer system 1700 or a human, provided that the consultation appears efficient, valuable, and effective.

[0408] In a preferred embodiment, the computer system 1700 can provide data and displays to help humans follow and understand the training process and system being trained. For example, in Figure 15 and Figure 5 In the histogram analysis in box 507, computer system 1700 can generate a histogram graph.

[0409] In some embodiments, the computer system 1700 can provide data from any comparative evaluation that can make significant improvements or show performance degradation exceeding a specified standard.

[0410] Humans can provide fractional theories and other human knowledge representations, and / or supervise computer systems 1700 to select human knowledge representations from publicly available sources.

[0411] Humans can supervise any changes in the hybrid network that alter the tradeoff between classification performance and sensitivity by more than a specified amount.

[0412] In some embodiments, the computer system 1700 can provide data to inform humans about changes in classification and sensitivity, although consultation may not be required.

[0413] In some embodiments, humans can provide guidance during hybrid training when deciding when to use alternatives to backpropagate derivatives. Preferably, to reduce human intervention, this human guidance applies a single decision to a large portion of the hybrid network, such as one or more complete layers, rather than a single element. In some embodiments, if a single element is critical to overall performance based on specified criteria, the computer system 1700 can enable human intervention in that element. This implementation may include the computer system 1700 collecting data and presenting it in a manner that enables efficient and effective human understanding. In some embodiments, if a single element is critical to one or more critical data items based on specified criteria, the computer system 1700 can enable human intervention in that element.

[0414] In some embodiments of continuous lifelong learning, computer system 1700 can continuously test the performance of new versions of the system on old tasks and prepare a report for humans on the performance degradation of old tasks.

[0415] In some embodiments, the computer system 1700 may seek human consultation to verify the sensitivity of the decision boundary in the discriminator. If the human consultant disagrees that examples of data items on or near the decision boundary are appropriately described as close to the boundary, it indicates that the system has failed to meet the second level of sensitivity, and the computer system 1700 should take remedial action. In some embodiments, the computer system 1700 may take remedial action by delegating and / or excluding data items. For example, the computer system 1700 may identify other data items to delegate by empirically training data weights and delegating data items with negative weights, such as... Figure 11 As discussed in 1123-1126. If human consultation indicates that one of the alternatives is not a match, the computer system 1700 can take remedial measures through data exclusion.

[0416] In a preferred embodiment, the computer system 1700 may seek this form of human consultation only for a small fraction of examples that are less than a specified consultation volume standard.

[0417] In box 415, in some embodiments, computer system 1700 may perform diagnosis and detection of instances of sensitivity violations. In some embodiments, computer system 1700 may use a tool called a "canary network." A canary network is a network designed and trained to readily alter its classification output due to adversarial attacks and other minor variations in the input. In some embodiments, computer system 1700 may train diverse sets of canary networks and robust sets of networks. In some embodiments, computer system 1700 may diversify its networks by using anti-binding to train multiple networks with the same or similar architectures. The use of anti-binding to increase diversity in a set of networks is described in patent 11,151,455 entitled "Counter-tying nodes of anodal network," the entire contents of which are incorporated herein by reference.

[0418] Given a classification task, computer system 1700 can create a canary network by training a conventional neural network on the classification task, avoiding any methods that would make the neural network resistant to adversarial attacks. For example, in some embodiments, computer system 1700 can avoid training the canary neural network with random perturbations or simulated adversarial attacks. In some preferred embodiments, computer system 1700 can also avoid... Figure 1 , 2Any steps to improve network sensitivity discussed in connection with figures 1, 2, 3, 4, 5, and others. Furthermore, in some embodiments, the computer system 1700 may perform the reverse of some of the recommended steps associated with these figures. For example, in some embodiments, the computer system 1700 may replace a bounded activation function (if any) with an unbounded activation function, rather than replacing an unbounded activation function with a bounded activation function. In some embodiments, the computer system 1700 may increase the slope and / or length of the non-flat interval of the activation function. Preferably, the computer system 1700 will select variations that will increase the canary network's vulnerability to input variations while minimizing the impact of variations on classification performance. In some embodiments, the computer system 1700 may retrain the canary network to achieve optimal performance on clean data while allowing it to fail on disturbed data.

[0419] In some embodiments, the computer system 1700 can be combined with Figure 1 , 2 Use the methods recommended in Figures 3, 4, 5, and others to create one or more robust networks.

[0420] Based on one or more examples of canary networks and one or more examples of robust networks, in some embodiments, computer system 1700 can create an arbitrarily large set of different networks by continuing or resuming training on multiple copies of a base network and debinding corresponding nodes between selected pairs in any two copies of the same base network. In some embodiments, computer system 1700 can debind a pair of nodes by creating bidirectional knowledge-sharing links with unequal relationships. By enforcing link regularization by selecting different subsets of nodes in different network pairs and / or selecting different subsets of the training dataset, computer system 1700 can produce a wide variety of differences between network pairs in different network sets.

[0421] Once the System-1700 has trained various canary networks and robust networks, it can use these different networks to diagnose any data items presented for classification. Any adversarial attacks or other perturbations to the input data are more likely to change the answer of the canary network than the robust network.

[0422] In some embodiments, the computer system 1700 can test the null hypothesis that there is no difference between the responses of the canary network and the robust network. The computer system 1700 can continue testing with new choices of one or more canary networks and one or more robust networks until the null hypothesis is rejected or the stopping criteria are met.

[0423] In other embodiments, the computer system 1700 may test the differences between the responses of the canary network and the robust network in other ways. In some embodiments, to further confirm that normal input has been interfered with, the computer system 1700 may perform a non-targeted adversarial attack. That is, the computer system 1700 may simulate an adversarial attack on the data item to be identified and present the data altered by the simulated adversarial attack to one or more canary networks. Preferably, in the non-targeted attack, the computer system 1700 may simulate a form of adversarial attack that attempts to make the canary network lower the score of the current answer, without targeting any new answer. If a new answer appears multiple times in multiple simulated non-targeted attacks, it indicates that the multiple answers can be easily accessed through small changes in the input. If the multiple answers from the non-targeted simulated attack are consistent with multiple responses from the robust network, this is strong evidence that the presented data item has been altered by an adversarial attack or other interference, and that the multiple answers are the correct answers to the original, undisturbed input.

[0424] In block 416, in some embodiments, computer system 1700 can implement proactive defense against imperceptible errors. In some embodiments of proactive defense, computer system 1700 can control one or more units via data exchange, such as... Figure 3B As shown. In some embodiments, active defense and Figure 8 Use it in conjunction with box 803.

[0425] In some embodiments, in order to achieve active defense, the computer system 1700 can train two or more activation functions for nodes that have discontinuities offset from each other and intervals with higher-order derivatives, these intervals being separated by intervals with zero derivatives and / or intervals where the difference between the maximum and minimum values ​​is less than a specified value and the magnitude of the derivative is less than a specified value.

[0426] In some embodiments, the computer system 1700 may implement one or more data-dependent data exchanges. In some embodiments, the computer system 1700 may specify a set of activation functions and data-dependent data exchanges such that, for an input data value d, under the control of the computer system 1700, the data exchange uses d as the input to an activation function for which the input lies within a relatively flat interval and does not approach a specified amount at the nearest end of the flat interval. In other words, the computer system 1700 can control the data exchange such that small changes in the input do not cause changes in the output exceeding a specified amount. In some embodiments, all relatively flat intervals in all activation functions have constant values, so that no change in the output occurs for any small change in the element input.

[0427] In block 417, in some embodiments, computer system 1700 can perform data item-specific active alignment. That is, computer system 1700 can compute the alignment of a data item after receiving it for classification. In some embodiments, computer system 1700 can perform local classification within a hybrid network, and the classification can be a specified known set rather than a final classification category.

[0428] In data-item-specific active alignment, in some embodiments, computer system 1700 may compute values ​​of variables in a set of cells that specify the alignment of the cells with a human knowledge representation (such as a split-theory). In an image recognition task, each alignment unit may be associated with a specific location in an image received for classification. Thus, in such an embodiment, computer system 1700 is computing the alignment between the received image and a split-theory model.

[0429] In some embodiments, the computer system 1700 may have trained an enhanced body theory model that also models the relative positions of the parts in the body theory.

[0430] Combination Figure 12 The process of training the split-theory alignment model is discussed.

[0431] In some embodiments, the computer system 1700 may align data items with human knowledge representation types other than scalar theory. For example, in tasks involving words, such as speech recognition, handwriting recognition, translation, or text understanding, the computer system 1700 may align observed or hypothesized words with parsing in a specified grammar. In some embodiments, the computer system 1700 may align words with a semantic web.

[0432] In some embodiments of image or video alignment, computer system 1700 may first create a lower-resolution representation of the video image for rapid preliminary analysis, thereby accelerating the analysis of the original image or video. The computation of a low-resolution representation of a high-resolution image or video is well known to those skilled in the art of image processing.

[0433] In some embodiments, computer system 1700 may perform classification of low-resolution images or videos. In some embodiments, computer system 1700 may use the classification of low-resolution data items to construct a list of best-scoring categories or known sets. In some embodiments, computer system 1700 may use the list of best-scoring categories or named sets to partially limit the possible classifications of higher-resolution data items. In some embodiments, computer system 1700 may add candidates to the list if the goodness-of-fit alignment of the shard theory is greater than a specified standard deviation. In some embodiments, computer system 1700 may determine a specified criterion in each target category or named set based on a measure of the goodness-of-fit in previous alignments of instances of the category or named set.

[0434] In some embodiments, the computer system 1700 may align a low-resolution image or video with cells in a simpler hybrid network trained on the low-resolution image or video. In a preferred embodiment, the computer system 1700 may design the body-theoretic alignment cells in the low-resolution model to be homologous to a specified subset of body-theoretic alignment cells in the high-resolution model. In this embodiment, the computer system 1700 may use the alignment of the low-resolution image to initialize a coarse alignment of the high-resolution image. In some embodiments, the computer system 1700 may then refine the alignment of the high-resolution image by filling in the alignments of cells that are not yet aligned. In some embodiments, the computer system 1700 may iteratively improve the alignment by changing the alignment of one or more cells to better match the cell alignments that are close to the changing cell in the body-theoretic adjacency graph. In some embodiments, the computer system 1700 may stop the alignment computation if no changes are made during the iteration of incremental improvement, or if the computer system 1700 detects a recurring loop. In some embodiments, the computer system 1700 may stop the iterative alignment process if certain other specified stopping criteria are met. For example, in some embodiments, if the only change still in progress is too small to be detected by the computer system 1700’s previously trained detection criteria, i.e., detecting changes that are too small to affect the classification by more than a specified small error rate, the computer system 1700 may stop the alignment process.

[0435] In some embodiments, the computer system 1700 may update the split-theory alignment model. In some embodiments, the computer system 1700 may store data and analysis in a repository.

[0436] In box 418, computer system 1700 can implement random activations during training and inference, including inference during deployment. In random activations, the activation values ​​of one or more elements in a hybrid network may differ when the same input data items are presented repeatedly.

[0437] In some embodiments, in block 418, computer system 1700 may use one or more of six types of randomization or noise: (1) adding noise to the output of one or more components and / or other variables, (2) simulation errors in one or more components, (3) probability swapping of data exchange destinations, (4) probability swapping of partitioned activation function intervals, (5) random dropout, and / or (6) simulated adversarial attacks on network inputs and / or one or more local data spaces. In a preferred embodiment, computer system 1700 may use the same type of randomization or noise in randomized training and diagnostics. Figure 5 (520). In some embodiments, the computer system 1700 may use a higher degree of randomization and / or noise during training compared to inference during deployment.

[0438] In some embodiments, in block 418, computer system 1700 can use with Figure 5 The techniques associated with box 520 are explained to implement any of the six types of randomization or noise. In some embodiments, when generating randomization and / or noise during inference during deployment, the computer system 1700 may use control hyperparameters that produce less variation than those used in training and / or diagnostics.

[0439] In some embodiments, in block 418, for classifying received data items, the computer system 1700 may generate randomization and noise multiple times for the network or selected unit. In some embodiments, the computer system 1700 may combine multiple sets of output values ​​in multiple randomization processes, as if they were output values ​​from a virtual integration. In such embodiments, the computer system 1700 may empirically train the hyperparameters controlling the randomization.

[0440] In some embodiments, computer system 1700 may use random activation to help create diverse sets of canary networks and / or robust sets of networks. Figure 4 (415). In some embodiments, the computer system 1700 may use unbinding and / or non-equivalent knowledge-sharing regularized links to further increase diversity. In some embodiments, the computer system 1700 may use soft binding and / or equivalent knowledge-sharing regularized links to mitigate differences between different sets of networks so that corresponding elements in each network maintain responsive relationships except for differences in random activation.

[0441] In some embodiments, during deployment, after receiving the data item to be classified, the computer system 1700 may randomly select a subset of various canary networks and a subset of robust networks using a selection probability distribution. The computer system 1700 does not specify the selection probability distribution until after receiving the data item to be classified.

[0442] In box 419, in some embodiments, computer system 1700 can build and train robust template models, such as Figure 10 As shown.

[0443] In box 420, in some embodiments, computer system 1700 may replace values ​​of x that are less than a specified threshold T1 and / or values ​​of x that are greater than a specified value T2 with a constant background score in the activation function f(x).

[0444] In some embodiments, in box 420, if the input value The value satisfies the specified value for more than a specified number of k input values. of Then, the computer system 1700 in the robust template model can replace the output value of the template model with a specified constant background score. In some embodiments, if If the value exceeds the specified value, the computer system 1700 can replace the template model's output value with the background score. (Combined) Figure 10 Robust template models were discussed.

[0445] In box 421, in some embodiments, computer system 1700 may jointly build and train one or more generators and / or classifiers with one or more teams of humans, such as in combination. Figure 21 As described above. In some embodiments shown in the figure, human involvement in training and development may be greater than... Figures 1 to 5 The description is more extensive. Figure 21 In joint development, one or more humans can directly control the training process. Figure 21 In the generator discussed, computer system 1700 can implement an interface allowing one or more humans to directly control the details of the generation. In collaborative development, unlike semi-automated knowledge engineering which minimizes human labor, more human involvement can be used to associate names with a wider set of known but unnamed features and unnamed characteristics. Additional names make the network more interpretable, thus providing more human guidance during training. Additional named features also make the generator's control more human-centric. In some embodiments, in box 412, computer system 1700 can implement logical and / or probabilistic reasoning within the network's cells, such as combining... Figure 21 The discussion is in box 2102.

[0446] In some embodiments, joint development and human guidance can be used in conjunction with collaborative generators, for example, for... Figure 5 Additional training data is generated in box 514.

[0447] In block 422, in some embodiments, computer system 1700 may train an adversarial generator and a real-number and non-real-number discriminator. In some embodiments, computer system 1700 may also train one or more cooperative generators. In some embodiments, computer system 1700 may train generators such as Figure 21 As described in boxes 2109, 2110, and 2111. In some embodiments, the computer system 1700 may use training of real and unreal discriminators and adversarial and cooperative generators based on variable resolution game theory, as described in international application PCT / US23 / 64296 entitled “Generation and discrimination training as a variable resolution game,” the entire contents of which are incorporated herein by reference, for generation and discrimination training / discrimination games.

[0448] Figure 5 This is a schematic diagram illustrating various aspects of hybrid training. Figure 5 In this diagram, topics are grouped according to the stages of the training process, as shown in the dashed boxes: 501 represents the initial training, 502 represents the main hybrid training stage, and 503 represents continuous training during lifelong learning and deployment (i.e., deploying the model to perform its training tasks). However, many concepts and techniques apply to multiple stages.

[0449] Figure 5 This is not a flowchart. The order of the boxes is not implied. The computer system 1700 may apply concepts and techniques in the various boxes in any order, rather than being roughly grouped into stages represented by dashed boxes 501, 502, and 503. In some embodiments, the computer system 1700 may impose some constraints on the order of application of technical prerequisites in some details. In some embodiments, all concepts and techniques work together, and the computer system 1700 may develop them collaboratively.

[0450] In some embodiments, to begin training, in box 504, the computer system 1700 may select a base network and progressively modify the network to improve it, such as by combining... Figure 1 Boxes 101 and 103 and Figure 1 and Figure 2 The other boxes in the table discuss this. The selected base network can be a neural network or a hybrid network.

[0451] In some embodiments, in block 504, computer system 1700 repeatedly seeks opportunities to improve sensitivity, overall interpretability, classification performance, and / or cost / performance. In some embodiments, computer system 1700 may repeatedly test the system on validation data in addition to training data.

[0452] In some embodiments, the computer system 1700 may extend the network from scratch. In some embodiments, at block 505, the computer system 1700 may grow the neural network from scratch and then convert the neural network into a hybrid network. In some embodiments, the computer system 1700 may build the hybrid network directly from scratch.

[0453] In box 506, in some embodiments, the computer system 1700 may use backpropagation with the derivative of gradient descent to train the connection weights and node biases of one or more elements. Gradient descent backpropagation is a standard method for training neural networks. However, for a reasonable network, not all training is done via gradient descent.

[0454] However, in some embodiments, training may be partially based on gradient descent even after initial training. However, in a preferred embodiment, training is not based solely on gradient descent. In a preferred embodiment, computer system 1700 uses a hybrid training approach to improve the sensitivity of the network being built and trained.

[0455] In block 507, in some embodiments, computer system 1700 performs histogram analysis. Block 507 uses an initial training group for two reasons: (1) histogram analysis is a fundamental technique that does not require other techniques as prerequisites, and (2) histogram analysis is also a widely useful technique that can serve as a preliminary step for other techniques. On the other hand, histogram analysis can also be used during main training (502) and / or continued training (503). For example, in any case where computer system 1700 uses human consultation, computer system 1700 can use histogram analysis to facilitate human consultation (…). Figure 4 (414). Histogram analysis will combine Figure 15 , Figure 2 Boxes 202 and 204 and Figure 4 Box 405 will be discussed further.

[0456] In box 507, computer system 1700 can compute histograms of one, two, or more variables. Variables can be continuous real numbers or discrete variables with values ​​belonging to a specified finite set. The specified finite set can represent a finite number of categories, a set of known sets, or a set of possible states of a hidden state-space model.

[0457] For a neural network node, the computer system 1700 can use the value of an affine sum or the output value of the node's activation function as a variable to compute a histogram. The computer system 1700 can also use values ​​received from any connection within the node as histogram variables.

[0458] For hybrid networks, the computer system 1700 can also use the values ​​provided by the cells as histogram variables.

[0459] The Computer System 1700 can also use the values ​​of the backpropagation derivatives as histogram variables. The derivative can be the derivative of the classification target or other specific functions. In hybrid networks, the Computer System 1700 can use the derivatives of local targets as histogram variables, such as... Figure 5 As shown in box 508. The computer system 1700 can also use local substitution derivative functions as histogram variables, such as... Figure 5 As shown in box 509.

[0460] In some embodiments, the computer system 1700 may perform regression on histogram counts to test a known set of data to determine whether any of the known sets meet a specified criterion associated with a specified variable. For example, in some embodiments, if the magnitude of the regression coefficient is greater than a specified value, the computer system 1700 may temporarily associate the variable with a known set. Figure 15 The use of histogram counting regression was further discussed.

[0461] In some embodiments, the computer system 1700 may select an interval of a specified variable to represent the detection of a known set, the selection being based on histogram counting.

[0462] In some embodiments, the computer system 1700 may select a variable as a discriminator between two known sets. In some embodiments, the computer system 1700 may use histogram counting to determine an initial threshold to be used with the discriminator. In some embodiments, the computer system 1700 may perform a comparative performance test to empirically adjust the threshold used with the discriminator. In some embodiments, the computer system 1700 may use data retained from the training data to perform such a comparative performance test. In some embodiments, the computer system 1700 may continue to use data collected from one or more deployed systems to empirically adjust the threshold.

[0463] When testing a selected variable as a detector of a known set or as a discriminator of two known sets, the computer system 1700 can find multiple known sets where the variable, as a detector or discriminator, satisfies a specified criterion. In this case, in some embodiments, the computer system 1700 can create multiple copies of the variable and the subnetwork that causes the variable. The computer system 1700 can then train each copy and its subnetwork on different tasks in the detector and / or discriminator tasks.

[0464] In some embodiments, the computer system 1700 may use histogram counts to determine the boundaries of data-related variables. Variables may be the output values ​​of nodes, cells, or units. In some embodiments, the computer system 1700 may limit the maximum and / or minimum values ​​of specified variables to better ensure the sensitivity of nodes or units that directly or indirectly receive input values ​​as functions of the variables. In some embodiments, the computer system 1700 may limit the minimum and / or maximum values ​​of variables based on the extreme values ​​of the variables observed on a specified dataset, such as a training dataset. In some embodiments, the computer system 1700 may set the limit values ​​of the variables to the extreme values ​​plus a specified margin. In some embodiments, for some variables, the margin may be zero or negative, thereby reducing the observation range. In some embodiments, the computer system 1700 may later adjust the limits of the variables.

[0465] In some embodiments, when computer system 1700 is Figure 2 When replacing an unbounded activation function in box 202, computer system 1700 can use histogram counting to determine the boundaries used in the new activation function.

[0466] In some embodiments, the computer system 1700 may use histogram counting to determine the template model. The initial value of the parameter. In some embodiments, the computer system 1700 can... The parameter estimate is the mean, median, or pattern of a set of data items. In some embodiments, the computer system 1700 can make any of these estimates based on a histogram.

[0467] In some embodiments, if the selected variable is associated with one or more known sets, the computer system 1700 may restrict the data selected for the histogram to data from the union of the associated known sets. The computer system 1700 may then initially set exclusion restrictions for the selected variable based on the histogram, such as... Figure 15 As shown.

[0468] In some embodiments, the computer system 1700 may use histogram counting when setting decision thresholds, such as... Figure 15 As described in frame 1506.

[0469] In some embodiments, the computer system 1700 may compute a joint histogram of two or more variables. In some embodiments, the computer system 1700 may use fewer and longer intervals for each variable in a multivariate joint histogram than in a single-variable histogram.

[0470] In some embodiments, if the computer system 1700 detects significant correlations or significant clustering in a low-dimensional histogram, the computer system 1700 may perform additional low-dimensional analysis. Figure 5 (Frame 517).

[0471] In block 508, in some embodiments, computer system 1700 can determine implicit local targets of nodes.

[0472] In some embodiments, computer system 1700 can determine implicit or explicit local targets based on the association between one or more intervals of a node activation function and a known set. For example, computer system 1700 can set a specified point in the interval as the target of a data item in the known set.

[0473] In some embodiments, computer system 1700 may determine implicit local targets based on the sign of the backpropagation derivative of a data item. For example, computer system 1700 or HNLMS may specify a pair of values, such as {0,1} or {1,1}, where the lower value is the target of any data item with a negative backpropagation derivative, and the higher value is the target of any data item with a positive backpropagation derivative. In some embodiments, computer system 1700 may use the lower bound of the activation function as the lower bound value and the upper bound of the activation function as the upper bound value.

[0474] In some embodiments, the computer system 1700 may use one or more intermediate values ​​as the target of a data item having a backpropagation derivative that is less than a specified absolute value.

[0475] In some embodiments, the computer system 1700 may use the determination of the presence or absence of implicit errors to convert the backpropagation of the derivative into the backpropagation of the data. Figure 18 and Figure 6 as well as Figure 5 (Frame 510).

[0476] In some embodiments, computer system 1700 may determine whether a node has committed an implicit error in determining the degree to which the node's activation within a specified interval is consistent with its membership in a known or named set. In some embodiments, computer system 1700 may use one or more corrected outputs to correct implicit errors in determining whether to associate a node with a known or named set. When computer system 1700 establishes a new association or changes an existing association, in some embodiments, computer system 1700 may subsequently use the new or modified association to determine explicit errors in the node's association.

[0477] Combination Figure 2 Box 203 discusses the use of implicit errors in training error prediction nodes.

[0478] In box 509, in some embodiments, the computer system 1700 can create alternative derivative functions for nodes. In some embodiments, the computer system 1700 can use alternative derivative functions to enable or accelerate node training with one or more intervals of relatively low magnitude derivatives, such as... Figure 3C As shown. In some embodiments, the computer system 1700 may select a basis to substitute the derivative function, and then the computer system 1700 multiplies it by the backpropagation derivative value or the sign of the backpropagation derivative value.

[0479] In some embodiments, the computer system 1700 may use continuous alternative derivative function activation during a portion of training, such as early training before a criterion is met, and discontinuous alternative derivative function activation after the criterion is met. In a preferred embodiment of this type of alternative derivative function, the computer system 1700 may multiply the basis alternative derivative function by the value of the backpropagation derivative or the sign of the value of the backpropagation derivative. In some embodiments, the computer system 1700 may multiply the basis alternative derivative function by the backpropagation derivative or multiply by the sign of the backpropagation derivative only for the input value x, for specified thresholds T1 and T2, T1 < x < T2. In some embodiments, the computer system 1700 may set a constant background score and apply this background score to all data having values ​​outside a specified interval, regardless of the sign or magnitude of the backpropagation derivative. In some embodiments, the computer system 1700 controlled by HNLMS may customize the criterion for the variation of the alternative derivative function for individual nodes. For example, HNLMS may calculate a customized criterion for a node based on measurements collected during node training.

[0480] In some embodiments, the computer system 1700 may design alternative derivative functions to push the activation from discontinuities or higher-order derivatives in the activation function toward the center of a relatively flat interval, such as... Figure 3C Functions 361, 362, and 363 are shown in the figure. In some embodiments, the computer system 1700 may delay the use of this alternative activation function until the training criteria are met.

[0481] In some embodiments, the computer system 1700 may use an alternative derivative function, wherein the value of the alternative derivative function is always positive for input values ​​less than a specified threshold T1, and / or always negative for input values ​​greater than a specified threshold T2. In some embodiments, the computer system 1700 may use such an alternative derivative function on nodes, which multiplies the basis alternative derivative function by the backpropagation value of x in the interval T1 < x < T2, such as... Figure 3C The intervals 353 and 354 are shown.

[0482] In block 510, in some embodiments, for a particular node, computer system 1700 may backpropagate labeled data examples instead of derivatives. In some embodiments, computer system 1700 may continue to backpropagate derivatives on pre-existing incoming connections while backpropagating the labeled data examples to one or more new elements.

[0483] In some embodiments, for a selected element with a standard discriminator activation function, computer system 1700 may determine whether an implicit error has occurred in the element for each data item in a specified set. In some embodiments, computer system 1700 may use this information to backpropagate data items with corrected implicit errors.

[0484] In some embodiments, computer system 1700 may backpropagate these labeled data items to one or more new elements while optionally continuing to backpropagate derivatives to its pre-existing incoming connections.

[0485] By correcting implicit errors, computer system 1700 may be able to train new elements with information that cannot be obtained through conventional backpropagation. An illustrative example of such a training procedure is described below, as Figure 18 shown. In the description, set S1 is the set associated with lower values x < X1 in the activation function input, and set S2 is the set associated with higher values x > X2, where X1 < X2. S1 or S2 may be associated with the maximum output value of the activation function, and the internal interval of the activation function is correspondingly monotonically increasing or decreasing.

[0486] In some embodiments, computer system 1700 may use the following procedure, as Figure 18 shown:

[0487] (1801) Obtain training data items.

[0488] (1802) Compute the activations of the network; call the activation value x.

[0489] (1803) Backpropagate the derivatives.

[0490] (1804) Associate X1 with set S1 and X2 with set S2.

[0491] (1805) If x < X1, select label S1 and go to (1809).

[0492] (1806) If x > X2, then select label S2 and go to (1809).

[0493] (1807) If the derivative backpropagated to the discriminator element is negative, select S1 and go to (1809).

[0494] (1808) Choose S2;

[0495] (1809) Backpropagate the current data item with label S1 or S2 to one or more of the following

[0496] a. A pair of detector elements;

[0497] b. Linear separator ( Figure 6 );

[0498] c. Output the subnet trained directly from labels S1 and S2;

[0499] d. Backpropagate the label to the error predictor and determine if there is an implicit error.

[0500] (1810) Save the corrected labels of S1 and S2 for all training data.

[0501] (1811) Repeat steps (1801) to (1811) until the stopping criterion is met.

[0502] (1812) Train new elements using the correction labels of S1 and S2.

[0503] In some embodiments, the computer system 1700 may delay saving the correction labels of S1 and S2 until the training of the selected standard discriminator elements is sufficiently stable, such that the correction sets S1 and S2 no longer change beyond the specified standard.

[0504] In some embodiments, the computer system 1700 may create new cells, using constraint optimization to identify S1 and S2 with correction labels, such as in combination with Figure 5 As discussed in box 524. Based on the solution of the constrained optimization, computer 1700 can create a linear thresholding function as a new element. In some embodiments, computer system 1700 can freeze a copy of the subnetwork such that the performance of the linear thresholding function does not degrade as the network is modified through further training. In some embodiments, if further training causes the selected standard discriminator element to produce new errors, computer system 1700 can train another linear thresholding function. If computer system 1700 removes the selected standard discriminator element from the network at or before training stops, there will be no path for backpropagation of the non-zero derivative of the specified function output through the one or more linear thresholding functions used by computer system 1700 to replace the selected standard discriminator element.

[0505] In some embodiments, the computer system 1700 may create one or more new units, each unit comprising a pair of detectors trained on sets S1 and S2, and an associated discriminator node. In some embodiments, the computer system 1700 may specify the associated discriminator to compute the output difference of the two detectors or a similar combination function, without requiring any training on the connection weights. In some embodiments, the computer system 1700 may use a piecewise constant function as the activation function for the discriminator. In some embodiments, the computer system 1700 makes the activation function a standard discriminator function. In some embodiments, the computer system 1700 may train two or more new units to diversify their S1 and S2 detectors. In some embodiments, the computer system 1700 may also train various canary network sets as S1 and S2 detectors and / or various discriminator sets.

[0506] In some embodiments, computer system 1700 may connect one or more new elements created in block 1809 to elements in higher layers of the network, up to and including the network's output. In some embodiments, computer system 1700 may train higher-layer subnetworks by backpropagating the output target derivative without backpropagating the derivative to or through the new elements. Furthermore, in some embodiments, once the new elements are trained, computer system 1700 may discard the original standard discriminator elements. In this embodiment, there is no corresponding derivative backpropagation for network activation of new data items through one or more new elements, thereby enhancing protection against adversarial attacks.

[0507] If the selected standard discriminator elements still have implicit errors when the network training converges, the computer system 1700 can improve the network's performance by replacing the standard discriminator elements with one or more new elements trained to the calibration sets S1 and S2. Furthermore, since the new elements are trained with data explicitly labeled S1 or S2, they may be easier to interpret than typical internal nodes of a deep network.

[0508] In block 511, in some embodiments, the computer system 1700 may partially or approximately train a second network to mimic a semi-homogeneous first network, wherein each designated node in the second network is associated with a node to be mimicked in the first network. In some embodiments, the computer system 1700 may use the output activation values ​​of nodes in the first network as targets for the activation values ​​of one or more designated nodes in the second network. In some embodiments, the computer system 1700 will use equivalent knowledge-sharing links to train designated nodes in the second network to better align with their counterparts in the first network.

[0509] In some embodiments, the design of the first network may be less sensitive than the design of the second network. In some embodiments, the first network may be a neural network, and the second network may be a hybrid network. On the other hand, in some embodiments, the second network may be less sensitive than the first network. For example, the first network may be trained as a perceptual hybrid network, while the second network may be a canary network (…). Figure 4 (415). In each case, the computer system 1700 can relax the simulation when the activation in the first network approaches the discontinuity or higher derivative point of the activation function of the node in the first network.

[0510] In some embodiments, the simulation may be limited to specified data items. For example, in some embodiments, the second network may be a new element of an ensemble that is being trained to be diverse on a specified subset of the data, but consistent on a specific disjoint subset, and in some embodiments, neutral on a third subset.

[0511] In block 512, in some embodiments, computer system 1700 implements conditional blending training. In conditional blending training, computer system 1700 can customize blending training techniques, such as applying the technique only to selected data items and / or only to selected cells or nodes.

[0512] For example, in block 512, in some embodiments, computer system 1700 may implement conditional flattening. In some embodiments, computer system 1700 may use data exchange (such as...) Figure 3B 325) implements conditional flattening tailored to each selected data item. In some embodiments, after training amounts specified, for example, by HNLMS, the computer system 1700 may start with a partially trained selected node having an activation function y=act1(x) divided into disjoint intervals such that act1(x) is non-flat with respect to one or more intervals. The computer system 1700 may then copy act1(x) to act1A(x) ( Figure 3B (323 in the original text), which may make some intervals less flat. The computer system 1700 can then copy act1(x) to act1B(x), making some or all intervals flatter. In some embodiments, the computer system 1700 can make act1B(x) ( Figure 3B 324) becomes a piecewise constant function. Then, computer system 1700 can add... Figure 3B Data exchange 325, to form Figure 3B Unit 322.

[0513] In some embodiments, the computer system 1700 may conditionally apply and Figure 5Any technology associated with boxes 508, 509, 510, 511 and / or 512.

[0514] In some embodiments, the computer system 1700 may apply any one of training techniques 513, 514, and / or 516 as continuous training after system deployment. In some embodiments, the computer system 1700 may apply one or more of these techniques during the main training phase prior to deployment.

[0515] Mixed-condition training will combine Figure 13 Further discussion.

[0516] In block 513, computer system 1700 can apply continuous learning during deployment, meaning that computer system 1700 can actively update learning parameters using acquired data during operational use. In some embodiments, computer system 1700 can continue to add elements to the network.

[0517] In some embodiments, the computer system 1700 can continue to test the performance of previously trained and validated data. In some embodiments, the computer system 1700 can apply equivalent knowledge-sharing links from earlier versions of the network to specified nodes in revised versions of the network to maintain the performance of specified data items.

[0518] In a preferred embodiment, the computer system 1700 can repeatedly test the system's performance on data reserved for verification testing. Preferably, the computer system 1700 adds new data to the verification data according to a specified schedule.

[0519] In some embodiments, the computer system 1700 may use one or fewer learning iterations to train a new template model to match new data.

[0520] For example, in some embodiments, computer system 1700 can, as Figure 10 The new template shown The value is set to the value in a single example or the average of multiple examples. In some embodiments, the computer system 1700 can... The value is set to the value specified by the hyperparameter. In some embodiments, the computer system 1700 may adjust the hyperparameter to a specified tradeoff between precision and recall. Such a template is called a single-shot or few-shot template. In some embodiments, the computer system 1700 may continue training the single-shot or few-shot template when additional data is acquired.

[0521] In some embodiments, the computer system 1700 may compute the alignment between the current data item and a split-theory model or other human knowledge model represented in a graphical structure. Training the alignment model will combine... Figure 12 and Figure 4Let’s discuss box 417.

[0522] Continuous learning during deployment will be combined Figure 8 Further discussion.

[0523] In box 514, computer system 1700 may generate additional data examples. For example, in some embodiments, computer system 1700 may use a mixture of generator models as described in U.S. Patent 11,354,578, entitled “Mixture of generator models,” the entire contents of which are incorporated herein by reference. As another example, computer system 1700 may use a Stochastic categorical autoencoder (SCAN) as described in U.S. Patents 10,679,129 and 11,461,661, entitled “Stochastic categorical autoencoder network,” the entire contents of which are incorporated herein by reference. In some embodiments, computer system 1700 may develop a SCAN with a hybrid autoencoder having parameter control, such as Figure 9 As shown. In some embodiments, the computer system 1700 can train a generator system or a mixture of SCANs with backpropagation from a joint objective to produce data classified as real numbers by a synthetic discriminator.

[0524] In some embodiments, the computer system 1700 may generate additional data examples as part of a combined human + AI creative activity, such as Figure 21 and Figure 4 As described in box 421.

[0525] In some embodiments, the computer system 1700 may generate data from some other form of collaborative generator, where the phrase "collaborative generator" is used in contrast to generative adversarial generators (GANs). Unlike GANs, the computer system 1700 may train the collaborative generator based on examples of real data. In some embodiments, the computer system 1700 may train the generator to generate real data using one or more real and synthetic discriminators. In some embodiments, the computer system 1700 may train the real and synthetic discriminator as a discriminator in a GAN and then use that discriminator with one or more collaborative generators. In some embodiments, the computer system 1700 may co-train the real and synthetic discriminator as a hybrid network, co-train it with one or more hybrid classifier networks, and share known sets and human knowledge representations (e.g., body theory). In some embodiments, the computer system 1700 may use one-way or two-way knowledge-sharing links between the classifier hybrid network and the real and synthetic discriminator. In some embodiments, the computer system 1700 may also share human knowledge representations with one or more collaborative generators.

[0526] In some embodiments, the computer system 1700 may use a conventional automatic encoder with a random bottleneck layer or a parameter-controlled automatic encoder with a random layer. Figure 9 Generate additional data examples.

[0527] In box 516, computer system 1700 can jointly train a set of partially or completely homologous networks. In a set of partially homologous networks, each specified node in the network is structurally homologous to a corresponding node in one or more other networks. In a set of completely homologous networks, each node in each network is structurally homologous to a corresponding node in each homologous network.

[0528] Computer system 1700 can perform initial training ( Figure 5 501) and / or master training ( Figure 5 502) and continued training ( Figure 5 During the 503 period, joint training using the same source network was performed.

[0529] In some embodiments, the computer system 1700 can reduce the computational cost of training multiple networks by using joint training of homologous networks. For example, in some embodiments, the computer system 1700 can use standard training on a single network or a selected subset of a set of networks. The computer system 1700 can then reduce the computational cost by using intensity hyperparameters. For high values, the remaining network is trained using links equal to knowledge-sharing on a specified subset of nodes. In some embodiments, the computer system 1700 may also train the diversity of the network using links that are not equal to knowledge-sharing for selected nodes and / or selected data items.

[0530] In some embodiments, the activation functions in a specified set of nodes in one or more source networks may have activation functions different from those of source nodes in other networks. For example, one network may have continuous activation functions for nodes, while a second network may have piecewise constant activation functions for source nodes.

[0531] In some embodiments, the computer system 1700 can create diversity by unbinding a selected set of nodes in specified pairs of a network set. In some embodiments, the computer system 1700 can create diversity by having one or more non-similar nodes in each network.

[0532] In some embodiments, the computer system 1700 may obtain one or more pre-trained networks, such as conventional neural networks that have not undergone sensitivity training. In some embodiments, the computer system 1700 may then use links equivalent to knowledge sharing to train homologous conventional or hybrid networks as a supplement to or alternative to gradient descent training. In some embodiments, the computer system 1700 may reduce the intensity hyperparameter in a later stage of training the homologous conventional or hybrid networks. .

[0533] As another example, computer system 1700 can use joint training to share knowledge among a set of distributed systems. For instance, during continuous learning while deploying a set of distributed, homogeneous networks, a particular distributed network might encounter new data items that lead to misclassification. In some embodiments, computer system 1700 can train the particular distributed network to correctly classify the new data items. In a preferred embodiment, computer system 1700 can restrict variations in the particular distributed network to a selected set of nodes. In some embodiments, computer system 1700 can then use knowledge-sharing links to train other networks to mimic the selected nodes of the particular distributed network.

[0534] Although corresponding nodes in a set of homologous networks are structurally homologous, computer system 1700 can jointly train a set of different homologous networks by applying equal-regularized links only on selected subsets of the data and non-equal-knowledge-sharing links on selected subsets of the data.

[0535] For example, when the activation of a robust node in a data item is closer to the discontinuity of the activation function in a robust network than a specified value, the computer system 1700 can jointly train one or more robust networks and one or more canary networks without forcing an equal knowledge-sharing link between robust nodes and canary nodes.

[0536] When jointly training a set of different, homogeneous networks, in some embodiments, the computer system 1700 may select a subset of nodes and / or a subset of data on which equal-knowledge-share links are not enforced. In some embodiments, the computer system 1700 may select a subset of nodes and / or a subset of data and enforce non-equal-knowledge-share links on the selected nodes and selected data. In some embodiments, the computer system 1700 may select a different subset of data for each selected node.

[0537] In some embodiments, computer system 1700 may use equal to and / or not equal to knowledge-sharing links on unlabeled data to train a set of homologous networks.

[0538] exist Figure 5 In box 517, computer system 1700 can perform analysis on two or more variables. Each variable can be the output value of a node, cell, or unit, the input of an activation function, or one of the input values ​​of a node or template. The set of two or more variables can be a subset of the variables in the local data space.

[0539] In some embodiments, in block 517, computer system 1700 may compute the correlation of all pairs of variables in a specified variable set. In some embodiments, computer system 1700 may compute a covariance matrix of a set of variables. In some embodiments, the specified variable set may be a set of values ​​of incoming connections of elements. In some embodiments, the variable set may be the union of the sets of incoming values ​​of a specified set of elements. In some embodiments, the specified set of elements may be two or more detectors for disjoint sets. In some embodiments, computer system 1700 may compute the correlation or covariance evaluated only on a specified subset of training data. For example, in some embodiments, computer system 1700 may compute the correlation or covariance only on the data to be discriminated by a specified element. For example, for a discriminator of two known sets, in some embodiments, computer system 1700 may compute the correlation or covariance only on the data in the union of the two known sets. In some embodiments, computer system 1700 may compute the correlation or covariance only on the data to be classified by a specified unit or subnetwork.

[0540] In some embodiments, the set of elements may be two detectors whose outputs are the inputs to the combination node. In some embodiments, the combination node may be a discriminator. In some embodiments, the computer system 1700 may train the combination node to approximate some logical function of its input, such as (A AND B), (A OR B), (A = B), (A ≠ B), or (A means B).

[0541] In some embodiments, the computer system 1700 may multiply a set of variables by a matrix to eliminate one or more pairwise correlations. In some embodiments, the computer system 1700 may specify a linear order of variables and may multiply variables by a matrix to eliminate correlations between pairs of variables that are adjacent in the linear order. For example, in a spectrum, the computer system 1700 may multiply the spectrum by a matrix to eliminate pairwise correlations in spectral amplitudes at adjacent frequencies.

[0542] In some embodiments, the computer system 1700 may multiply the set of variables by the inverse of the estimated covariance matrix.

[0543] In some embodiments, the computer system 1700 may replace the original variables with a set of variables obtained by multiplying by the decorrelation matrix or the estimated inverse covariance matrix. In some embodiments, the computer system 1700 may replicate the set of nodes receiving the original variables and connect the transformed variables to new nodes while retaining the original nodes with untransformed variables. In some embodiments, the computer system 1700 may temporarily create two networks, one without a specified variable transformation and the other with a specified transformation. In some embodiments, the computer system 1700 may compare the performance of the two networks and select the network with better performance. In some embodiments, the computer system 1700 may keep the two networks as integrated members.

[0544] In some embodiments, in block 517, computer system 1700 can perform cluster analysis on a specified dataset using a specified set of variables. In some embodiments, computer system 1700 can perform cluster analysis using a set of variables for which the computer system... Figure 5 Clustering of data was detected in histogram analysis performed by computer system 1700 in box 507.

[0545] In some embodiments, in block 517, the computer system 1700 may train a discriminator or classifier in a data space for which the computer system 1700 has detected a nonlinear decision boundary between two or more known sets. In some embodiments, the computer system 1700 may detect such a nonlinear decision boundary through multivariate histogram analysis, such as by combining... Figure 15 and Figure 5 The discussion is in box 507.

[0546] In box 518, in some embodiments, computer system 1700 may determine control parameters for excluding or delegating data from the training and / or inference of selected elements. Data exclusion and delegation will be combined... Figure 11 Let's have a discussion.

[0547] In box 519, computer system 1700 can add new nodes and / or new connections to the network, such as in combination with Figure 2 As discussed in box 208. In some embodiments, the computer system 1700 may create new nodes to achieve node partitioning, wherein a node is replaced by a set of two or more nodes.

[0548] In box 519, in some embodiments, the computer system 1700 may create one or more copies of an element and then train these copies to be different from the original element and from each other. In some embodiments, the computer system 1700 may train each copy on a different dataset, or each copy may be trained with data weights having different weights. In some embodiments, the computer system 1700 may utilize data exchange to distribute data. In some embodiments, the computer system 1700 may implement different data weights via a digital multiplier in learning parameter updates. In some embodiments, the computer system 1700 may implement data selection and weighting by specifying data relevance probabilities in probabilistic data exchange. Data weighting is described in U.S. Patent 11,010,671, entitled "Iterative training of a nodal network with data influence weights," the entire contents of which are incorporated herein by reference.

[0549] In some embodiments, the computer system 1700 may segment nodes to create nodes for receiving data delegations, such as... Figure 11 As shown.

[0550] In block 520, in some embodiments, the computer system 1700 may use randomized training and diagnostics. In some embodiments, the computer system 1700 may use randomized training to make the system more robust to external noise (such as noise in the input data) and internal noise (such as noise and / or errors generated by individual components in the network). In some embodiments, the computer system 1700 may use randomized training to support randomized activation (…). Figure 4(418) to improve sensitivity. In some embodiments, the computer system 1700 may use random training and random activation to improve classification performance, for example, by training and using virtual random ensembles. In some embodiments, in block 520, the computer system 1700 may use one or more types of randomization and / or noise to better understand the interdependencies of elements in the network and diagnose potential vulnerabilities.

[0551] In some embodiments, in block 520, computer system 1700 may use one or more of six types of randomization or noise: (1) adding noise to the output of one or more components and / or other variables, (2) simulation errors in one or more components, (3) probability switching of the destination of data exchange, (4) probability switching of the partitioned activation function interval, (5) random drop, and / or (6) simulated adversarial attacks on network inputs and / or one or more local data spaces. In some embodiments, computer system 1700 may use a higher degree of randomization and / or noise during training than during inference during deployment.

[0552] In box 520, for the noise type (1) described above, in some embodiments, when a specified input data item is provided to the network global input space or any selected local data space, the computer system 1700 may apply a technique referred to herein as “additive noise activation” to one or more variables during the calculation of the activation of the hybrid network. In some embodiments, the computer system 1700 may apply noise activation to the output values ​​of one or more nodes, cells, or cells. The underlying variable to which noise is added is called the “underlying activation variable.” The random variable to which the amount of noise is added during a particular activation calculation is called the “additive random noise variable.”

[0553] In some embodiments, the computer system 1700 may use noisy activation during training, diagnostic processes, and / or classification reasoning. The computer system 1700 may use noisy activation during initial training ( Figure 5 (dashed box 501) and / or main training ( Figure 5 During the period indicated by the dashed box 502, additive activation with noise is used. When a data item d for training or classification is received, the computer system 1700 determines the value of each additive random noise variable as a new random sample.

[0554] The probability distribution of the additive random noise variable that specifies the noise activation variable can be any type of probability distribution. For example, it can be a Gaussian distribution, a trimmed Gaussian distribution, or a uniform distribution.

[0555] The type of probability distribution can be specified, for example, by system design, HNLMS, or can be selected by computer system 1700 through empirical testing of two or more specified options of the distribution. In some embodiments, computer system 1700 can use different types of probability distributions for different noisy activation variables.

[0556] Without loss of generality, the mean of an additive random noise variable can be set to zero, since any non-zero mean is equivalent to only a change in the underlying activation variable.

[0557] For each additive random noise variable, the computer system 1700 can specify one or more variables or hyperparameters to control the diffusion of the random sample population. For example, for a Gaussian distribution, the computer system 1700 can specify the standard deviation. For a uniform distribution, the computer system 1700 can specify the length of the interval centered at zero. For a pruned Gaussian distribution, the computer system 1700 can specify the standard deviation to be pruned and the number of standard deviations.

[0558] In some embodiments, the computer system 1700 can empirically estimate the values ​​of one or more diffusion parameters of one or more additive random noise variables through empirical training, such as by combining Figure 5 The discussion is in box 521.

[0559] In some embodiments, for simulated error type (2), for an element associated with one or more known sets, the computer system 1700 can simulate an error in a data item in a known set by randomly selecting an alternative activation value in an interval unrelated to the known set. For a data item not in an interval associated with a named set, the computer system 1700 can randomly select an alternative activation value in an interval associated with a known set different from the named set.

[0560] In some embodiments, for randomization type (3), activation interval switching, or type (4) data exchange destination switching, the computing system 1700 can generate discrete-valued random variables to select the activation interval or the destination of the data exchange. The probability distribution of the discrete-valued random variables can be specified by parameters or hyperparameters, such as by HNLMS, or the computer system 1700 can determine it through empirical training. Figure 5 (521).

[0561] In some embodiments, for the random drop type (5), the computer system 1700 may determine whether to randomly drop selected elements of a particular data item with a probability specified by a hyperparameter. In some embodiments, the activation value used in the case of dropping may be specified as zero or may be specified by a hyperparameter. In some embodiments, the element may have an element-specific alternative activation value in the case of dropping.

[0562] In some embodiments, for the random type (6), the computer system 1700 may randomly select whether to use a simulated adversarial attack on a specified element of a specified data item with a probability specified by a hyperparameter. In some embodiments, for example, the system design and / or HNLMS may specify multiple adversarial attack methods. In such embodiments, the computer system 1700 may randomly select which adversarial attack method to use on a specific element of a specific data item.

[0563] In some embodiments, at block 520, computer system 1700 may use randomization and noise to understand and diagnose interactions between elements in a network. For example, computer system 1700 may add noise and / or change the output of a first designated element to discover and / or evaluate the impact of these changes in the output of the first designated element on a second designated element. In some embodiments, the first designated element does not need to be directly connected to the second designated element. The second designated element may be any element in the network that is directly or indirectly affected by changes in the output of the first designated element.

[0564] In some embodiments, at block 520, the computer system 1700 may determine the magnitude of additive noise, the probability of one or more other variations, and / or the strength of simulated adversarial attacks based on the values ​​of a set of hyperparameters. In some embodiments, the computer system 1700 may use separate randomization hyperparameters for each noise or randomization type for each element in the network.

[0565] In some embodiments, at block 520, the computer system 1700 may use a greater degree of noise and randomization during training than during inference during deployment. In some embodiments, at block 520, the computer system 1700 may estimate the optimal values ​​of the randomized hyperparameters during training by using empirical training with the randomized hyperparameters, such as in conjunction with… Figure 5 The discussion is in box 521.

[0566] In some embodiments, as a diagnostic procedure, the computer system 1700 may choose to study the effects of randomization and noise of other variables on a specified set of important components or variables. For example, in some embodiments, the computer system 1700 may choose to study the effects of internal variable randomization on network output nodes. In some embodiments, the computer system 1700 may choose to study the effects of randomization of other variables on the output values ​​of one or more units. In some embodiments, the computer system 1700 may choose to study the effects of randomization of other variables on the values ​​of one or more variables in one or more local data spaces.

[0567] In some embodiments, when studying the effects on a specified set of important variables, the computer system 1700 can calculate the effects of multiple randomizations, randomly varying the value of each randomization hyperparameter within a specified range. In some embodiments, the computer system 1700 can measure the effects of noise activation or measure each ordered pair including a noise variable and an affected variable.

[0568] To improve efficiency, in some embodiments, the computer system 1700 may first select an important variable to measure its impact, and then select a set of noise activation variables specific to the affected important variable, instead of analyzing each ordered pair of selected important variables and noise variables, as described below. In some embodiments, the computer system 1700 may reverse the order, first selecting a noise variable, and then selecting a set of important variables affected by the selected noise variable, as described in later paragraphs.

[0569] In some embodiments, as a diagnostic procedure, the computer system 1700 may select one of a set of important variables to measure the effect of noisy activation. In some embodiments, the computer system 1700 may select each important variable sequentially. The computer system 1700 may then compute multiple randomizations and calculate the regression correlation between the change in the selected important variable and the degree of change in one or more variables altered in the randomization. In some embodiments, the computer system 1700 may use a greater degree of randomization and noise during the diagnostic process than during training.

[0570] In some embodiments, for a specified important variable, the computer system 1700 may select one or more noise variables, wherein the randomization of the noise variables has a greater impact on the specified important variable than a specified criterion. In some embodiments, the computer system 1700 may use a specified criterion that preferentially selects noise variables that are less directly related to the important variable than noise variables that are more directly related to the important variable. In some embodiments, the computer system 1700 may make additional modifications to further improve the sensitivity and robustness of one or more selected noise variables.

[0571] In some embodiments, when diagnosing an error or termination of call to one of the critical variables, the computer system 1700 may examine relevant noise variables to determine whether an error or disturbance in one of the relevant noise variables may have caused or significantly contributed to the error or termination of call to the critical variable. If so, the computer system 1700 may take corrective measures to improve the accuracy and / or robustness of the noise variables.

[0572] In some embodiments, the computer system 1700 may select one or more candidate noise variables and calculate the randomization and impact of noise in the noise variables on other variables in the network. In some embodiments, the computer system 1700 may select a set of one or more other variables that are significantly affected by the selected candidate noise variables based on specified criteria. In some embodiments, the computer system 1700 may add the selected candidate noise variables and the selected significantly affected variables to a set of associated significant variable and noise variable pairs.

[0573] In some embodiments, the computer system 1700 may use the relationship between a noise variable and one or more related important variables to help interpret the noise variable. In some embodiments, the computer system 1700 may use the relationship between an important variable and one or more noise variables to help interpret the important variable.

[0574] For example, computer system 1700 can determine whether the set of data items with activation values ​​within a specified activation interval in one member of a pair of variables is equal to or unequal to a specified degree of proximity to the set of data items with activation values ​​within a specified interval in the other member of the pair. If so, in some embodiments, computer system 1700 can create knowledge-sharing links between the specified activation intervals in one or both directions.

[0575] In some embodiments, if an interval in an important or noisy variable is associated with a known or named set, the computer system 1700 may examine it to determine whether the known or named set is likely associated with a pair of affected or important variables.

[0576] In some embodiments, computer system 1700 may use pairings of critical and noise variables to diagnose the causes and potential solutions for erroneous or terminated calls to individual data items. For example, computer system 1700 may attempt to determine changes that computer system 1700 may be able to make in the network design and / or the learning parameters of one or more noise variables to correct erroneous or terminated calls to critical variables on individual data items. In some embodiments, computer system 1700 may generate simulated adversarial attacks and / or random perturbations in the network input space and / or local data space to create examples of erroneous or terminated calls to one or more critical variables.

[0577] In block 521, in some embodiments, the computer system 1700 may empirically estimate the optimal values ​​of one or more hyperparameters. In some embodiments, the computer system 1700 may empirically estimate the values ​​of one or more learning parameters. In some embodiments, the computer system 1700 may use the empirical estimation of the learning parameters as an alternative to gradient descent training and / or data backpropagation training. In some embodiments, the computer system 1700 may alternate between the empirical estimation of the learning parameters and one or more other methods of training the learning parameters. In some embodiments, the computer system 1700 may alternate between training the learning parameters by empirical estimation and / or by another training method, and further alternate with hyperparameters controlled by, for example, HNLMS. In some embodiments, the computer system 1700 may empirically estimate the performance of the hyperparameters as information provided to the HNLMS for controlling the hyperparameters.

[0578] like Figure 5 As described in the discussion of box 520, computer system 1700 can empirically estimate the values ​​of one or more diffusion parameters of one or more additive random noise variables. Another example of how computer system 1700 can empirically estimate parameters are the endpoints of acceptance or rejection intervals in detector or discriminator nodes or cells. As another example, computer system 1700 can empirically estimate the background score of any detector or discriminator variable. More generally, computer system 1700 can empirically estimate the values ​​of any constant interval of a variable. Furthermore, computer system 1700 can empirically estimate the maximum and minimum values ​​of any specified relatively flat interval. As another example, in some embodiments, computer system 1700 can empirically estimate the norm or other constraints of the acceptance region of a template model. In some embodiments, computer system 1700 can empirically estimate the norm of data exclusion for detector or discriminator elements. In some embodiments, computer system 1700 can estimate one or more norms of data exclusion for robust template models.

[0579] In some embodiments, the computer system 1700 may empirically estimate multiple parameters simultaneously. For example, in some embodiments, the computer system 1700 may empirically estimate the diffusion parameters of one or more diffusion parameters of an additive random noise variable. In some embodiments, the computer system 1700 may empirically estimate one or more parameters associated with one or more constant or relatively flat intervals.

[0580] In some embodiments, the computer system 1700 may empirically estimate one or more parameters characterizing the location and orientation of the decision boundary.

[0581] In some embodiments, the computer system 1700 can simultaneously evaluate a plurality of quantifiable targets or a specified combination of a plurality of quantifiable targets.

[0582] Without restrictions, illustrative examples of quantifiable objectives that the computer system 1700 can use in empirical learning of classification tasks include: (1) classification performance, (2) sensitivity, and (3) overall interpretability.

[0583] In an unrestricted context, illustrative examples of quantifiable objectives that the computer system 1700 may use in the experiential learning of generation tasks include: (1) the recall of generated named set examples, (2) the precision of generated named set examples, (3) the performance of one or more previously trained real and synthetic discriminators for cooperative or adversarial generators, (4) the performance of a classifier trained on new data using supplementary data generated by the generator, and (5) the sensitivity of a classifier trained on supplementary data generated by the generator.

[0584] In some embodiments, the computer system 1700 may compute a function of two or more quantifiable objectives as a new quantifiable objective. For example, the computer system 1700 may compute a weighted average of classification performance, sensitivity, and overall interpretability, which represents a trade-off between objectives.

[0585] In some embodiments, the computer system 1700 can evaluate classification performance by running multiple trials with noise activation and / or random noise added to the input variables.

[0586] In some embodiments, the computer system 1700 can evaluate sensitivity by running multiple trials using simulated adversarial attacks and / or noise activation.

[0587] In the illustrative embodiment, the computer system 1700 can simultaneously optimize multiple parameters and / or hyperparameters based on experience, such as combining... Figure 20 The subject of discussion.

[0588] In some embodiments, during training or ongoing learning after deployment, the computer system 1700 may repeat empirical optimization of one or more parameters based on criteria controlling the repetition frequency. In some embodiments, the computer system 1700 may repeat empirical estimates more frequently based on observations of system operation. For example, if a measurement of one or more quantifiable targets decreases during continued use or training, the computer system 1700 may repeat the empirical estimate. In some embodiments, the computer system 1700 may repeat the empirical estimate if ongoing training on new data examples has changed the values ​​of the learned parameters beyond a specified criterion.

[0589] In block 522, in some embodiments, computer system 1700 may replace a selected node with a set of three or more nodes. More specifically, computer system 1700 may replace a node with a unit or a set of nodes, including (1) a first new node created from the selected node and a copy of a connection to the selected node with positive weights, (2) a second new node created from the selected node and a copy of a connection to the selected node with negative weights, and (3) a third new node having a copy of the connections from the first and second new nodes and the outgoing connections of the selected node. In some embodiments, computer system 1700 may copy connections with weight magnitudes less than a specified value to the first and second new nodes of the selected node.

[0590] In some embodiments, the computer system 1700 may create more new nodes and divide incoming connections into more groups.

[0591] In some embodiments, the computer system 1700 may interpret each source node sending a connection to a selected node as a detector of data items that generate higher activation values. Therefore, the computer system 1700 may interpret an incoming connection with positive weights as evidence that the connected source node has a set of data items with high activation values. In some embodiments, the computer system 1700 may interpret nodes with a mixture of negative and positive weights as a means of distinguishing between the set of data items detected by the consensus of source nodes with positive weights and the set of data items detected by the consensus of source nodes with negative weights.

[0592] In continuous training where the symbols of the incoming connections do not change significantly, the computer system 1700 will tend to train backpropagation from selected nodes to enable the source nodes to learn to better match this interpretation.

[0593] In some embodiments, computer system 1700 may create new units including new nodes. Each of a pair of new nodes may have a subset of the incoming connections of the original node and an outgoing connection to a third node. In some embodiments, computer system 1700 may select only connections with weights greater than a specified threshold T1 for the first node in the pair. Computer system 1700 may select only connections with weights less than a threshold T2 as incoming connections to the second node in the pair. In some embodiments, T1 ≤ 0 ≤ T2. In some embodiments, computer system 1700 may reverse the weight signs on all incoming connections to the second node in the pair. In these embodiments, for the second node in the pair, computer system 1700 may replace the activation function of the original node with an activation function equal to a constant minus the original activation function. In some embodiments, computer system 1700 may limit the magnitudes of T1 and T2 to less than a specified amount. In such embodiments, a majority of the incoming weights of each of the pair of new nodes will be positive. In some embodiments, T1 = T2 = 0.

[0594] In some embodiments, the computer system 1700 may interpret each node in the new pair as a detector representing a detection with a higher activation function value.

[0595] The new third node may have an activation function representing some form of difference, such as

[0596]

[0597] In some embodiments, the computer system 1700 may associate a third node as a discriminator between two sets, the discriminator modeling the two sets as disjoint.

[0598] In some embodiments, during continued training, if the updated weight of a connection exceeds one of a threshold T1 or T2, the computer system 1700 may add or remove an incoming connection.

[0599] In some embodiments, for two known sets A and B, computer system 1700 may associate one of a pair of new nodes with the data set in A, rather than with the data set in B, and associate the other node of the pair of new nodes with the data set in B, rather than with the data set in A. In some embodiments, computer system 1700 may train additional nodes associated with the intersection of A and B and / or train additional nodes associated with data sets not in A and B.

[0600] If the original node is associated with a detector of a known set, in some embodiments, the computer system 1700 may temporarily associate the first node in the pair with a detector of the known set and the second node in the pair with a detector of a supplementary subset of the known set.

[0601] If the original node is a discriminator between two known sets, in some embodiments, the computer system 1700 may associate each node in the node pair with a detector from one of the known sets. In this association, each detector has an incoming connection with predominantly positive weights.

[0602] In some embodiments, the computer system 1700 may train nodes using weight decay. That is, at each weight update, the computer system 1700 may multiply the revised weights by a specified constant r < 1. The process of weight decay is well known to those skilled in the art of training neural networks. In some embodiments, if the magnitude of the weights is less than a specified magnitude, and this has been the case for a specified number of iterations, the computer system 1700 may prune the connections.

[0603] In some embodiments, the computer system 1700 may use a template model to replace one or more new detector nodes.

[0604] In block 523, in some embodiments, the computer system 1700 may select a set of two or more decision elements. In some embodiments, for each selected decision element, the computer system 1700 may create a new decision element that is initialized to replicate the selected decision element. In some embodiments, the computer system 1700 may connect each replicated element to an incoming connection that replicates the selected decision element, and initialize the connection weights to be the same.

[0605] In some embodiments, the computer system 1700 may then form a set of decision elements including copies of the selected decision elements. In some embodiments, the computer system 1700 may add one or more decision elements that represent the intersection of the target sets of the original selected decision elements and the complementarity of the target sets. In some embodiments, the computer system 1700 may then form softmax relations on the extended set of copy detectors. The computer system 1700 may then train the system to associate the extended set of copy detectors with disjoint sets.

[0606] In some embodiments, the computer system 1700 may replace one or more disjoint set detectors with a template model and continue training using softmax relations.

[0607] In box 524, in some embodiments, the computer system 1700 may use constrained optimization to train the weights of a linear threshold function, such as by combining Figure 6 The discussion focuses on training the weights of the linear threshold function. After training, the computer system 1700 can then backpropagate to the nodes connected to the linear threshold nodes using backpropagation of derivatives, backpropagation of labeled data examples, or both, or neither. This is done as the increment grows ( Figure 1 103 and Figure 5 (504) In some embodiments, computer system 1700 can build and train the entire network without using any backpropagation.

[0608] Figure 6 This is a flowchart illustrating an example of constrained optimization during training.

[0609] In box 601, computer system 1700 obtains or selects a network.

[0610] In box 602, in some embodiments, computer system 1700 may switch activations and / or make other modifications to the selected network, such as in combination with Figure 2 The subject of discussion.

[0611] In block 603, in some embodiments, computer system 1700 selects an authentication task. For example, computer system 1700 may select elements that include a standard discriminator activation function. In some embodiments, computer system 1700 may select a target set or a known set of detector elements and specify the authentication task as performing authentication between the selected set and its complementary set. In some embodiments, computer system 1700 may select a task that authenticates two known sets.

[0612] In block 604, in some embodiments, computer system 1700 may select a set of data items with target values ​​for the task selected in block 603. For example, in some embodiments, computer system 1700 may select only data items for which the selected node produces an implicit error. In some embodiments, computer system 1700 may select data items for which the selected node has terminated the call. In some embodiments, computer system 1700 may avoid selecting data items that exceed specified exclusion limits. In some embodiments, computer system 1700 may avoid selecting data items that have already been delegated from the selected node. In some embodiments, computer system 1700 may avoid selecting data items that are correctly classified by the network, even if they are incorrect for the selected node.

[0613] In block 605, in some embodiments, computer system 1700 can determine whether the implicit objective of a node is linearly separable by finding weights that minimize T2-T1, provided that for any data item with a lower objective value, the input of the activation function is less than or equal to T2, and for any data item with a higher objective value, the input of the activation function is greater than or equal to T1. For example, if the input of the activation function is a weighted affine sum of the values ​​of the incoming connections from the node, computer system 1700 can find the optimal weights through linear programming. In some embodiments, computer system 1700 can select a nonlinear objective function for optimization in block 605. In this case, computer system 1700 can find the weights through nonlinear programming with linear constraints. Linear and nonlinear programming with linear constraints are well known to those skilled in the art of mathematical programming.

[0614] In some embodiments, the computer system 1700 can use incremental growth ( Figure 1 103 and Figure 5 (504) to construct a hybrid network without any backpropagation, neither the derivative backpropagation ( Figure 6 (612), and there is no backpropagation of data examples ( Figure 6 613 and Figure 5 (510). For example, in some embodiments, the computer system 1700 may repeatedly discard the target ( Figure 6 (607).

[0615] In some embodiments, in block 605, computer system 1700 may create a new element having an activation function, such as a linear threshold function or other monotonic function, the weights of which and a discrimination threshold are calculated in block 605.

[0616] In box 606, computer system 1700 checks whether the minimum value of T2-T1 is less than or equal to 0. If so, the selected data item is linearly separable. In this case, computer system 1700 proceeds to box 609. Otherwise, computer system 1700 proceeds to box 607.

[0617] In block 607, in some embodiments, computer system 1700 may determine whether to discard some selected targets, and if so, which targets to discard. In some embodiments, computer system 1700 may choose to continue without discarding any selected targets.

[0618] In some embodiments, the decision of whether to discard selected data items of a node may involve cost / performance trade-offs. In some embodiments, the computer system 1700 may make the decision based on fixed criteria specified in the system design. In some embodiments, HNLMS may perform cost / performance analysis on specific cases of selected nodes or cells. In some embodiments, for example, the computer system 1700 and HNLMS may test performance trade-offs, preferably on data other than the training data.

[0619] In block 608, computer system 1700 decides whether to repeat constraint optimization after discarding some target data items. If yes, computer system 1700 returns to block 605. Otherwise, computer system 1700 continues to block 609. In some embodiments, computer system 1700 may repeatedly discard target data items until the number of errors is reduced. Unless there are two identical data items, one of which is an error and the other is not, computer system 1700 can eventually reduce the number of errors as long as there are remaining errors, because a set of two distinct data items is always linearly separable.

[0620] In block 609, in some embodiments, computer system 1700 may check the performance of the selected element on unselected data items in block 604, if any. The performance of the selected element on these unselected data items may have changed because the weights of the incoming connections may have changed.

[0621] In block 610, in some embodiments, computer system 1700 may determine whether to select additional data items for the element selected or created in block 603.

[0622] In some embodiments, the decision of whether to select alternative data items for a node may involve cost / performance trade-offs. In some embodiments, the computer system 1700 may make a decision based on fixed criteria specified by the system design. In some embodiments, HNLMS may perform cost / performance analysis on a specified condition for a selected node or cell. In some embodiments, for example, the computer system 1700 and HNLMS may test performance trade-offs, preferably on data other than the training data.

[0623] In box 611, computer system 1700 selects whether to backpropagate the data example, derivative, or both, or neither. If computer system 1700 decides to backpropagate the data example, it proceeds to box 613. If computer system 1700 decides to backpropagate the derivative, it proceeds to box 611. If computer system 1700 decides to backpropagate both, it can proceed to boxes 612 and 613 in parallel. If computer system 1700 decides not to backpropagate either, computer system 1700 proceeds directly to box 614. For example, if computer system 1700 determines to make and freeze a copy of the subnetwork with a new linear thresholding function, computer system 1700 can choose not to backpropagate either of the above. If, in addition to the linear thresholding function, each discriminator trained on the task selected in box 603 is eventually discarded from the network and replaced by one or more linear thresholding functions with frozen subnetworks, as... Figure 5 As shown in box 510, the finally trained network will not have a path for backpropagating the derivatives of the selected discrimination task to the input variables, which prevents adversaries from using backpropagation of gradients to compute adversarial attacks. In some embodiments, the computer system 1700 can use this strategy without limitation for multiple discrimination tasks.

[0624] In block 613, in some embodiments, computer system 1700 may backpropagate data examples. In some embodiments, computer system 1700 may backpropagate only errors and terminate the call. In some embodiments, for backpropagation purposes, computer system 1700 may use data items as a criterion for terminating the call, which accepts more data items than the criteria selected as data items in block 604.

[0625] In block 612, in some embodiments, computer system 1700 may use, for example... Figure 3C The alternative derivative function shown is used for backpropagation of the derivative.

[0626] In block 614, in some embodiments, computer system 1700 may determine whether to select an additional authentication task based on a specified stopping criterion.

[0627] Figure 7This is a flowchart illustrating an embodiment of one aspect of hidden state space modeling in one aspect of the present invention. Note that the meaning of the word "hidden" in the phrase "hidden state space model" is very different from the phrases "hidden layer" or "hidden node" in the discussion of hierarchical neural networks. In the discussion of hierarchical neural networks, all layers except the output layer and its nodes can be called "hidden." Input values ​​are also not considered "hidden." However, the values ​​of state variables in a hidden state space model are hidden much deeper. In a hidden state space model, the activations of all nodes are considered observables. In some embodiments, some values ​​stored in cells can also be considered observables. However, in a hidden state space model in a hybrid network, state variables are not considered observables, although estimates of their values ​​may be stored in cells.

[0628] In some embodiments, the computer system 1700 can model hidden state variables as unobserved random variables. In some embodiments, the computer system 1700 can model observable variables as random variables whose values ​​depend on the unobserved hidden state variables. Based on the values ​​of the observed variables, the computer system 1700 can estimate the hidden variables by applying Bayesian rules.

[0629] In block 701, in some embodiments, computer system 1700 may specify a cell space that includes hidden state variables. For example, for an image, in some embodiments, computer system 1700 may specify a two-dimensional rectangular cell grid. The hidden state variables may then represent an interpretation of a local region in the image. Alternatively, in some embodiments, computer system 1700 may specify a planar two-dimensional hexagonal tiling or other tiling. In some embodiments, the hidden state space may represent a conditional random field.

[0630] For data represented as a sequence, in some embodiments, computer system 1700 may define a one-dimensional cell sequence. The hidden state-space variables can then represent the state of a time-varying process over a specified period. In some embodiments, the hidden state space can represent a Hidden Markov Process.

[0631] In some embodiments, the computer system 1700 may specify an adjacency graph, that is, a graph in which each cell is connected to its neighboring cells, such as four neighbors in a rectangular grid (or eight neighbors if angular neighbors are counted) or six neighbors in a hexagonal grid. In a sequence of cells, the computer system 1700 may connect each cell to the preceding and following cells in the cell sequence.

[0632] In some embodiments, the computer system 1700 can represent the relationships between adjacent parts in a split theory as an adjacency graph. In some embodiments, the computer system 1700 can determine the mapping from elements in the split theory to cells in the hybrid network through an alignment process, particularly for each input data item ( Figure 12 ).

[0633] In block 702, in some embodiments, computer system 1700 may specify one or more hidden state variables. In some embodiments, a hidden state variable may be a variable having values ​​selected from a finite set. In some embodiments, a hidden state variable may be a continuous value variable.

[0634] In some embodiments, the computer system 1700 may represent the hidden state using an n-tuple variable.

[0635] In block 703, in some embodiments, computer system 1700 may obtain a model of the relationship between hidden state variables and observable variables. In some embodiments, this relationship may represent any numerical relationship. In some embodiments, the model may represent the conditional probabilities of observed variables within and around grid points of hidden state cells, conditional on the values ​​of the hidden state variables. In some embodiments, the model may represent the relationship between state variables in adjacent cells of an adjacency graph. For example, the graph may be an adjacency graph of the parts of a split-theory model of a detected hypothetical object.

[0636] In box 704, in some embodiments, computer system 1700 can obtain a model of the co-occurrence of specified state pairs in adjacent cells. For example, computer system 1700 can represent the probability of a particular hidden state variable as the probability of the hidden state variable values ​​at adjacent positions in the adjacency graph.

[0637] In some embodiments, the computer system 1700 may use learning parameters to train an abstract model of the degree of association between state values ​​in adjacent cells of an adjacency graph; these parameters are not necessarily trained to model conditional probabilities. In some embodiments, the computer system 1700 may train directional learning parameters between state values ​​in ordered pairs of adjacent cells. In some embodiments, the computer system 1700 may train association parameters in each direction. In some embodiments, the computer system 1700 may train non-directional associations between learning parameters for unordered pairs of adjacent cells.

[0638] In block 705, in some embodiments, computer system 1700 may select one or more paths in the state space for evaluation. For example, in a layer of a convolutional neural network, computer system 1700 may select a cell path corresponding to a grid point path in an image. In a sequence model, computer system 1700 may select a forward sequence or a reverse sequence. More generally, in some embodiments, computer system 1700 may select any path through an adjacency graph.

[0639] In box 706, in some embodiments, computer system 1700 may calculate the probability of a state given observed text. In some embodiments, computer system 1700 may update the learning parameters of an abstract model of the degree of association between ordered or unordered state value pairs of adjacent cells in an adjacency graph.

[0640] In block 707, in some embodiments, given an estimated distribution of the hidden state space variables, computer system 1700 can update the model of the observed variables.

[0641] In box 708, in some embodiments, computer system 1700 may update the model of the conditional probability model of state values ​​in adjacent cells, or update the model of the abstract model of the directed or undirected association of state values ​​in adjacent cells.

[0642] In block 709, in some embodiments, computer system 1700 determines whether to select a new path through the graphics based on specified criteria. If yes, computer system 1700 returns to block 705. Otherwise, computer system 1700 continues to block 710.

[0643] In box 710, in some embodiments, computer system 1700 determines, based on specified criteria, whether to train different models for the association between the observed variable and the state values ​​in adjacent cells. If yes, computer system 1700 returns to box 703. Otherwise, computer system 1700 continues to box 711.

[0644] In block 711, in some embodiments, computer system 1700 may determine whether to perform analysis of different state-space formulas. If yes, computer system 1700 returns to block 701. Otherwise, computer system 1700 will proceed according to... Figure 7 The process shown is complete.

[0645] Figure 8This is a flowchart of an exemplary embodiment of performing a perceptual classification operation using a trained hybrid network and fast matching. The exemplary embodiment includes defenses against potential interference in the data. The exemplary embodiment also includes methods for reducing the computational cost required for classification. In one aspect of the invention, the exemplary embodiment also provides continuous training performed concurrently with continuous training during the use of fast matching and inference.

[0646] In box 801, computer system 1700 obtains the trained system.

[0647] In box 802, computer system 1700 receives data items to be classified.

[0648] In block 803, in some embodiments, computer system 1700 may use sensitivity data exchange to implement active defense against disturbed data, as discussed in conjunction with block 416. In active defense, the network includes one or more data exchanges through which computer system 1700 selects among multiple activation functions or multiple nodes such that the selected activation of the data item received in block 802 is located in a relatively flat area and not near the boundary of that area.

[0649] In block 804, in some embodiments, computer system 1700 may perform rapid preliminary classification. In some embodiments, computer system 1700 may use a lower resolution image or other simplified representations of the received data items for classification to compute the preliminary classification. In some embodiments, computer system 1700 may use a simpler model instead of a fully hybrid network or instead of some cells.

[0650] In some embodiments, the computer system 1700 may perform a pre-computed classification table lookup on the low-order representation of the cell input.

[0651] In some embodiments, the computer system 1700 may perform bottom-up component detection. In some embodiments, the computer system 1700 may use a simplified network to perform bottom-up component detection. In bottom-up component detection, the computer system 1700 may first perform classification and detection of smaller units, such as smaller objects or parts of objects in an image, or short sound segments in speech or other audio. In bottom-up component detection, the computer system 1700 may then classify a selected subset of larger units based on the identity of the smaller unit with the best score.

[0652] In some embodiments, the computer system 1700 may perform hypothetical pruning on some larger units based on their scores relative to the best-scoring unit at one stage of bottom-up component detection.

[0653] In some embodiments, the computer system 1700 may create a short list of optimal score alternative classifications for one or more cells or the entire classification network. In some embodiments, the computer system 1700 may then skip some calculations for hypotheses not on the calculated short list. In some embodiments, the computer system 1700 may replace hypotheses not on the short list with specified backoff scores.

[0654] In some embodiments, computer system 1700 may coordinate bottom-up component detection and alignment with the adjacency graph, as described in conjunction box 805.

[0655] In box 805, in some embodiments, computer system 1700 can perform fast classification based on alignment with the adjacency graph. Combined with Figure 12 Training based on adjacency graph alignment is discussed.

[0656] As an example of initial alignment for classification via a fully hybrid network, computer system 1700 can detect some parts of the object's periphery. Computer system 1700 can then align the detected parts and other peripheral elements with the object's structural components. Computer system 1700 can then align and classify the parts within the structural components. In some embodiments, computer system 1700 can coordinate this alignment-based rapid classification with bottom-up component detection, as discussed in conjunction with box 804.

[0657] Within block 806, computer system 1700 can perform other sequential processing within the cell. For example, computer system 1700 can compute the hidden state-space model, such as in conjunction with... Figure 7 As another example, the computer system 1700 can depict line segments, curves, and / or contours by sequentially connecting pairs of related or similar chains of adjacent elements. The computer system 1700 can use this sequential processing for tasks such as: (1) determining whether two local regions are connected, (2) finding the contours around an object, (3) finding the boundaries separating two regions, or (4) solving a maze.

[0658] In block 807, in some embodiments, computer system 1700 may perform checks on preliminary results.

[0659] In some embodiments, the computer system 1700 may validate classification results based on results obtained through other means. For example, the computer system 1700 may compare the results of the current preliminary match with results obtained from other preliminary matches.

[0660] In some embodiments, in an image recognition task, if the current preliminary match uses a low-resolution representation of the image, the computer system 1700 can compare the result of the current preliminary match with a classification result using a higher-resolution image. In some embodiments, the computer system 1700 can accelerate the classification of the higher-resolution image by pruning computation based on the preliminary match result.

[0661] In some embodiments, the computer system 1700 can verify preliminary results by comparing them to a higher resolution image at key points in the body theory of a short list of initially matched best candidate classifications. For example, the computer system 1700 can verify classifications along portions of the aligned body theory perimeter.

[0662] In some embodiments, the computer system 1700 may compute backpropagation based on the output activation of each candidate category on the short list of initial matches. In some embodiments, the computer system 1700 may use a network other than the one used in the initial matching to compute this backpropagation, and / or may compute backpropagation based on a higher resolution image. In some embodiments, the computer system 1700 may then examine each node in the network to see if the node is incorrect relative to the implicit local object, such as... Figure 5 As described in box 508. In some embodiments, the computer system 1700 can increase the short list of answers from the initial match by adding candidate answers obtained by changing the activation of selected nodes that are close to a threshold, which will change the erroneous or termination call to the implicit local target.

[0663] In some embodiments, the computer system 1700 may use different knowledge sources or different input data sources to verify the preliminary matching results against the results obtained from classification. For example, in the classification of speech or other audio, the computer system 1700 may use different signal processing of the audio signal to verify the preliminary classification results. As another example, in speech recognition or handwriting recognition, the computer system 1700 may compare the results obtained from recognizing phonemes or letters with the results obtained using a word sequence language model.

[0664] In some embodiments, the computer system 1700 may validate the results of an initial match using a parameter generator. In some embodiments, the computer system 1700 may adjust the parameters of the parameter generator to suit the observed input data, but is constrained by the generator's parameters conforming to one of the constraints on the short list of candidate answers that are initially matched. In some embodiments, the computer system 1700 may select the answer whose output of the parameter generator best matches the input data of the classifier. In some embodiments, the computer system 1700 may compare the output of the parameter generator with the input to trim or add candidate answers to the short list.

[0665] In some embodiments, the computer system 1700 may add additional answers to the short list based on prior experience with errors encountered in easily confused output categories. For example, HNLMS may maintain a confusion matrix of errors caused by previous versions of the network under development or other systems trained for the same classification task.

[0666] In some embodiments, the computer system 1700 may use abductive reasoning to evaluate each candidate answer on a short list. For example, in some embodiments, the computer system 1700 may apply abductive reasoning to explain the potential reasons for low scores on candidate answers. As a specific example, if candidate words in a speech recognition task match well except for one phoneme based on formant tracking, the computer system 1700 may examine the hypothesis that formant identification in formant tracking may be incorrect because two formants with close frequencies may form a single peak in the spectrum.

[0667] In block 808, in some embodiments, computer system 1700 may determine whether to perform additional preliminary classification. If not, computer system 1700 proceeds to block 809. If yes, computer system 1700 returns to block 803 to perform additional preliminary classification. In some embodiments, computer system 1700 may perform more complex classification based on the previous preliminary classification. In some embodiments, computer system 1700 may perform a new preliminary classification that is designed to be different from the previous preliminary classification.

[0668] In block 809, in some embodiments, computer system 1700 may perform tests to detect whether data items received in block 802 are subject to adversarial attacks or other interference that could alter classification. In some embodiments, computer system 1700 may examine the network to verify whether nodes and activation functions satisfy [the relevant parameters / conditions]. Figure 2 The associated basic first-level sensitivity rules. In some embodiments, to detect potential adversarial attacks or other interference, the computer system 1700 may use a diverse set of canary networks, such as combining... Figure 4The discussion is in box 415.

[0669] In block 810, in some embodiments, computer system 1700 may acquire additional data. In some embodiments, the additional data may include additional training data. In some embodiments, the additional data may include data obtained during the operation of the current classifier system or from other deployed classifier systems. In some embodiments, the data may be generated or synthesized data. In some embodiments, computer system 1700 may generate additional data in a region selected by computer system 1700 by analyzing the results of the initial classification.

[0670] In block 811, in some embodiments, computer system 1700 may apply techniques of continuous learning and growth, such as combining Figure 1 The technology discussed.

[0671] In some embodiments, the computer system 1700 may add to and modify the network, and these additions and modifications are tailored to the data items received in box 802.

[0672] In block 814, in some embodiments, computer system 1700 may optionally perform controlled semi-supervised learning using unlabeled data. In some cases, it may be impossible to verify whether the classification is correct during deployment. In some embodiments, computer system 1700 may acquire additional unlabeled or unclassified data. In some embodiments, during deployment, an end user or another person may explicitly or implicitly confirm some classification results, while other classification results may remain unconfirmed.

[0673] In some embodiments, the computer system 1700 may perform additional training, including unlabeled data obtained during deployment by temporarily labeling each unlabeled result with the best score label from the classifier. This process of using unlabeled labels from the classifier is called semi-supervised learning, which is well-known to those skilled in the art of machine learning. Semi-supervised learning typically improves the performance of machine learning systems when the amount of labeled training data is limited. On the other hand, in some cases, semi-supervised learning can lead to a degraded performance of machine learning systems, sometimes to an extreme degree. In fact, there is a theorem that as the amount of unlabeled data in semi-supervised learning approaches infinity, the performance of semi-supervised learning converges to that of unsupervised learning.

[0674] In some embodiments, the computer system 1700 may limit the relative amount of unconfirmed data with respect to the training data and confirmed labeled data obtained during deployment. In some embodiments, the computer system 1700 may use labeled data other than that used for training to validate the network performance after semi-supervised learning. In some embodiments, the computer system 1700 may examine the network performance after semi-supervised learning by comparing it with classification results obtained by other systems not trained on unconfirmed semi-supervised labeled data.

[0675] In box 815, in some embodiments, computer system 1700 may save the trained network to a network repository and the data to a data repository.

[0676] In a preferred embodiment, computer system 1700 can return to block 802 to continue lifelong learning.

[0677] Figure 9 This is a schematic diagram of an automatic encoder with parameter control that can be used by computer system 1700 in several aspects of the present invention.

[0678] A conventional automatic encoder includes input data 901, which is provided as input to an encoder network 902 by a computer system 1700. The computer system 1700 also provides the input data 901 as an output target to a decoder network 905. In a conventional automatic encoder, the output node 904 of the encoder 902 is also an input value to the decoder 905. In a parameter-controlled automatic encoder, the computer ...

Claims

1. A method comprising: (a) A generative AI system is developed by a programmed computer system by adding one or more interpretable network elements to the generative AI system, wherein: The generative AI system includes one or more trained machine learning networks that generate text paragraphs in response to prompts and context; and Train each of the one or more interpretable network elements to identify two or more sets of interpretable training data items for the generative AI system; (b) After adding one or more interpretable network elements, the generative AI system with the added one or more interpretable network elements is subjected to update training by a programmed computer system; (c) Following (a) and (b), the programmed computer system determines whether the continued development of the generative AI system is necessary; When it is determined at (c) that the continued development of generative AI systems is necessary, repeat steps (a) through (c); and When it is determined in step (c) that the continued development of the generative AI system is no longer necessary, the generative AI system is deployed to generate text paragraphs.

2. The method according to claim 1, wherein, The generative AI system includes an autoregressive next-word predictor.

3. The method according to claim 1, wherein, The generative AI system includes large language models (LLMs).

4. The method according to claim 3, wherein, The LLM includes a converter network.

5. The method according to claim 4, wherein, The converter network includes one or more attention blocks, and each of the one or more attention blocks includes one or more attention heads.

6. The method according to claim 1, wherein: The generative AI system includes a hybrid network; and One or more interpretable network elements include a cell of the hybrid network.

7. The method according to claim 1, wherein: The generative AI system includes neural networks; and One or more interpretable network elements include nodes added to the neural network of the generative AI system.

8. The method according to claim 1, wherein, The one or more interpretable network elements comprise multiple nodes organized into a new layer, which is added to one of the machine learning networks of the generative AI system.

9. The method according to claim 7, wherein, The neural network includes a converter network, wherein the converter network includes one or more attention blocks, and wherein each of the one or more attention blocks includes one or more attention heads.

10. The method according to claim 7, wherein, At least one of the one or more interpretable network elements includes a first node, which is trained to classify each data item into a specific set belonging to a set of two or more interpretable sets.

11. The method according to claim 10, wherein, Adding the one or more interpretable network elements includes: The target node of the generative AI system is selected by a programmed computer system; The first node is trained by a programmed computer system based on regression calculations to identify a first set of one or more classification categories and a second set of one or more classification categories; and Add the first node to the generative AI system.

12. The method according to claim 11, wherein, Training the first node for identification includes: For two or more classification categories, a programmed computer system calculates a regression of the number of instances in each classification category based on the activation value of the target node; A programmed computer system selects a first set of one or more classification categories with positive regression coefficients; A second set of classification categories with negative regression coefficients is selected by a programmed computer system; and The first node is trained by a programmed computer system to distinguish between the first set and the second set.

13. The method according to claim 11, wherein, Training the first node for identification includes: For two or more classification categories, a programmed computer system calculates a regression of the number of instances in each category based on the activation value of the target node; A programmed computer system selects a first set of one or more classification categories whose regression coefficients are greater than a first threshold. A programmed computer system selects a second set of one or more classification categories whose regression coefficients are less than a second threshold, wherein the second threshold is less than or equal to the first threshold; and The first node is trained by a programmed computer system to distinguish between the first set and the second set.

14. The method according to any one of claims 11 to 13, wherein, The regression includes linear regression.

15. The method according to any one of claims 11 to 13, wherein, The regressions include monotonic regressions.

16. The method according to any one of claims 11 to 13, wherein, Each of the aforementioned categories includes words.

17. The method according to any one of claims 11 to 13, wherein: The neural network includes a converter network; and Each category includes the named state of the hidden model of the converter network.

18. The method according to claim 10, wherein, The update training includes backpropagation to the weights on the direct input connections of the first node, but the backpropagation depth will not exceed the weights on the direct input connections of the first node.

19. The method according to claim 7, wherein: The one or more interpretable network elements include a first node and a second node; and The update training includes soft-binding the first and second nodes.

20. The method according to claim 19, wherein, The update training also includes rebinding the connection to the first node to the corresponding connection to the second node.

21. The method according to claim 7, wherein: The one or more interpretable network elements include a first node and a second node; and The update training includes applying node-to-node regularized links between the first and second nodes.

22. The method according to claim 1, wherein, Deploying the generative AI system to generate text paragraphs includes: The generative AI system generates the first and second text paragraphs; The first and second text paragraphs are presented to the user via a user interface by a backend computer system that includes the generative AI system and the programming computer system. The backend computer system receives the user's selection of one of the first and second text paragraphs; and The generative AI system is further trained by a programmed computer system based on the selection.

23. The method according to claim 1, wherein, Deploying the generative AI system to generate text paragraphs includes: The generative AI system generates text paragraphs; An interpretation system of a back-end computer system generates an interpretation related to a text paragraph generated by the generative AI system, the back-end computer system including the generative AI system and a programmed computer system; The backend computer system presents the text paragraphs and explanations to the user via a user interface; The backend computer system receives feedback from the user regarding the explanation; and The programmed computer system further trains the generative AI system based on the feedback.

24. The method according to claim 23, wherein: Generating the interpretation includes generating a first interpretation and a second interpretation, wherein each of the first interpretation and the second interpretation is related to a text paragraph generated by the generative AI system; and The feedback includes the user's choice of the first or second interpretation.

25. The method according to claim 23, wherein, The feedback includes ratings from users who provided the explanation.

26. The method according to claim 1, wherein, Developing the generative AI system includes adding a probabilistic model to the generative AI system, wherein the probabilistic model is associated with one or more interpretable network elements.

27. The method according to claim 26, wherein, Adding the probability model involves adding the probability model from the repository of the programmed computer system.

28. The method according to claim 26, wherein, The probability model includes a nonparametric probability model.

29. The method according to claim 28, wherein, The nonparametric probability model includes the nonparametric conditional probability model.

30. The method of claim 26, wherein: The one or more interpretable network elements include first and second interpretable network elements; and The probabilistic model includes a nonparametric correlation correction model, which is conditioned on the activation values ​​of the first and second interpretable network elements.

31. The method according to claim 26, wherein, The probability model includes a parametric probability model.

32. The method according to claim 26, wherein, The probability model includes template-type models.

33. The method according to claim 32, wherein, The template-type model represents the probability corresponding to the activation values ​​of a specified set of one or more interpretable network elements when a specified event in a part of a sequence not yet observed by the generative AI system is conditioned during the training of the generative AI system.

34. The method according to claim 26, wherein, Deploying the generative AI system for each position in the sequence of text paragraphs to be generated includes: The generative AI system calculates a list of multiple candidate positions in the sequence based on the context of the text paragraph; The generative AI system estimates the probability of each candidate in a list of multiple candidate positions in the sequence; and The generative AI system adds one of the candidates in the list to the position in the sequence based on the probability.

35. The method according to claim 34, wherein: The generative AI system includes multiple generative subsystems that collectively generate the text paragraphs; and The method further includes: The generative AI system broadcasts the activation values ​​of selected embedded nodes from each of the multiple generative subsystems to the other generative subsystems; and The generative AI system modifies the selected embedded nodes in each generative subsystem using data-related regularization links between the selected embedded nodes.

36. The method according to claim 35, wherein, The data-related regularization links include "equals" relational regularization links.

37. The method according to claim 36, wherein, Deploying the generative AI system also includes: The generative AI system selects a set of key language units relevant to the context, wherein each language unit is a word or phrase; and The generative AI system loads multiple example text paragraphs, wherein one or more key language units from the set of key language units appear in each of the multiple example text paragraphs. The calculation of a list of multiple candidates includes: calculating the list based at least in part on (i) the word count in the example text paragraph and (ii) the autoregressive prediction score of the word at the said position.

38. The method according to claim 28, wherein, Deploying the generative AI system for each position in the sequence of text paragraphs to be generated includes: The generative AI system selects a set of key language units that are context-dependent, where each language unit is a word or phrase; The generative AI system loads multiple example text paragraphs, wherein one or more key language units from the set of key language units appear in each of the multiple example text paragraphs; The generative AI system calculates a list of multiple candidate positions in the sequence based on the context of the text paragraph, wherein calculating the list of multiple candidate positions includes: calculating the list based on (i) the number of words in the example text paragraph and (ii) the output of a nonparametric probability model; The generative AI system estimates the probability of each candidate in a list of multiple candidate positions in the sequence; and The generative AI system adds one of the candidates in the list to the position in the sequence based on the probability.

39. The method according to claim 38, wherein: Loading multiple sample text paragraphs includes: testing the semantic similarity between candidate text paragraphs and the context by a semantic analysis system of a backend computer system; and The backend computer system also includes the generative AI system and the programming computer system.

40. The method according to claim 39, wherein, Deploying the generative AI system to generate text paragraphs includes: The generative AI system generates the first and second text paragraphs; The back-end computer system presents the first and second text paragraphs to the user via a user interface; The backend computer system receives the user's selection of one of the first and second text paragraphs; and The computer system programmed thereon further trains the generative AI system based on the selection.

41. The method according to claim 39, wherein: The backend computer system also includes an interpretation system; and Deploying the generative AI system to generate text paragraphs includes: The generative AI system generates text paragraphs; The interpretation system generates interpretations related to the text paragraphs generated by the generative AI system. The backend computer system presents the text paragraphs and explanations to the user via a user interface; The backend computer system receives feedback from the user regarding the explanation; and The programmed computer system further trains the generative AI system based on the feedback.

42. A system comprising: A generative AI system comprising one or more trained machine learning networks, such that the generative AI system generates text paragraphs in response to prompts and context; and A programmed computer system that communicates with the generative AI system, wherein the programmed computer system is configured to: (a) The generative AI system is developed by adding one or more interpretable network elements to the generative AI system, wherein each of the one or more interpretable network elements is trained to identify two or more sets of interpretable training data items of the generative AI system. (b) After adding one or more interpretable network elements, perform update training on the generative AI system using the added one or more interpretable network elements; (c) Following (a) and (b), determine whether continued development of the generative AI system is necessary; and When it is determined at (c) that the continued development of generative AI systems is necessary, repeat steps (a) through (c); and Wherein, when it is determined at step (c) that the continued development of the generative AI system is not required, the generative AI system is configured to generate text paragraphs.

43. The system according to claim 42, wherein, The generative AI system includes an autoregressive next-word predictor.

44. The system according to claim 42, wherein, The generative AI system includes large language models (LLMs).

45. The system according to claim 44, wherein, The LLM includes a converter network.

46. ​​The system according to claim 45, wherein, The converter network includes one or more attention blocks, and each of the one or more attention blocks includes one or more attention heads.

47. The system according to claim 42, wherein: The generative AI system includes a hybrid network; and One or more interpretable network elements include a cell of the hybrid network.

48. The system according to claim 42, wherein: The generative AI system includes neural networks; and One or more interpretable network elements include nodes added to the neural network of the generative AI system.

49. The system according to claim 42, wherein, The one or more interpretable network elements comprise multiple nodes organized into a new layer, which is added to one of the machine learning networks of the generative AI system.

50. The system according to claim 48, wherein, The neural network includes a converter network, wherein the converter network includes one or more attention blocks, and wherein each of the one or more attention blocks includes one or more attention heads.

51. The system according to claim 48, wherein, At least one of the one or more interpretable network elements includes a first node, which is trained to classify each data item into a specific set belonging to a set of two or more interpretable sets.

52. The system according to claim 51, wherein, The programmed computer system is configured to add the one or more interpretable network elements in the following manner: Select the target node of the generative AI system; The first node is trained based on regression calculation to identify a first set of one or more classification categories and a second set of one or more classification categories; as well as Add the first node to the generative AI system.

53. The system according to claim 52, wherein, The programmed computer system is configured to train the first node to perform identification in the following manner: For two or more categories, a regression is calculated based on the activation value of the target node to determine the number of instances in each category; Select a first set of one or more classification categories that have positive regression coefficients; Select a second set of one or more classification categories with negative regression coefficients; as well as The first node is trained to distinguish between the first set and the second set.

54. The system according to claim 52, wherein, The programmed computer system is configured to train the first node to perform identification in the following manner: For two or more classification categories, a regression is calculated based on the activation value of the target node to determine the number of instances in each category; Select a first set of one or more classification categories whose regression coefficients are greater than a first threshold; Select a second set of one or more classification categories whose regression coefficients are less than a second threshold, wherein the second threshold is less than or equal to the first threshold; as well as The first node is trained to distinguish between the first set and the second set.

55. The system according to any one of claims 52 to 54, wherein, The regression includes linear regression.

56. The system according to any one of claims 52 to 54, wherein, The regressions include monotonic regressions.

57. The system according to any one of claims 52 to 54, wherein, Each of the aforementioned categories includes words.

58. The system according to any one of claims 52 to 54, wherein: The neural network includes a converter network; and Each category includes the named state of the hidden model of the converter network.

59. The system according to claim 51, wherein, The programmed computer system is configured to perform the update training by backpropagating partially to the weights on the direct input connections of the first node, but the backpropagation depth does not exceed the weights on the direct input connections of the first node.

60. The system according to claim 48, wherein: The one or more interpretable network elements include a first node and a second node; and The programmed computer system is configured to perform the update training by partially soft-binding the first and second nodes.

61. The system according to claim 60, wherein, The programmed computer system is configured to perform the update training by partially unbinding the connection to the first node to the corresponding connection to the second node.

62. The system according to claim 48, wherein: The one or more interpretable network elements include a first node and a second node; and The programmed computer system is configured to perform the update training by partially applying node-to-node regularization links between the first and second nodes.

63. The system according to claim 42, wherein: The system includes a back-end computer system, which includes a generative AI system and a programming computer system; The generative AI system is configured to generate first and second text paragraphs; The backend computer system is configured as follows: The first and second text paragraphs are presented to the user via a user interface; and Receive user selection of one of the first and second text paragraphs; and The programming computer system is also configured to train the generative AI system based on the selection.

64. The system according to claim 42, wherein: The system includes a back-end computer system, which includes a generative AI system, a programming computer system, and an interpretation system. The generative AI system is configured to generate text paragraphs; The interpretation system is configured to generate interpretations related to text paragraphs generated by the generative AI system; The backend computer system is configured as follows: The text paragraphs and explanations are presented to the user via a user interface; and Receive feedback from users regarding the explanation; and The programmed computer system is configured to further train the generative AI system based on the feedback.

65. The system according to claim 64, wherein: The interpretation system is configured to generate the interpretation by partially generating a first interpretation and a second interpretation, wherein each of the first interpretation and the second interpretation is related to a text paragraph generated by the generative AI system; and The feedback includes the user's choice of the first or second interpretation.

66. The system according to claim 64, wherein, The feedback includes ratings from users who provided the explanation.

67. The system according to claim 42, wherein, The programmed computer system is configured to develop the generative AI system by partially adding a probabilistic model to the generative AI system, wherein the probabilistic model is associated with one or more interpretable network elements.

68. The system according to claim 67, wherein, The programmed computer system is configured to add the probability model from the programmable computer system's repository.

69. The system according to claim 67, wherein, The probability model includes a nonparametric probability model.

70. The system according to claim 69, wherein, The nonparametric probability model includes the nonparametric conditional probability model.

71. The system according to claim 67, wherein: The one or more interpretable network elements include first and second interpretable network elements; and The probabilistic model includes a nonparametric correlation correction model, which is conditioned on the activation values ​​of the first and second interpretable network elements.

72. The system according to claim 67, wherein, The probability model includes a parametric probability model.

73. The system according to claim 67, wherein, The probability model includes template-type models.

74. The system according to claim 73, wherein, The template-type model represents the probability corresponding to the activation values ​​of a specified set of one or more interpretable network elements when a specified event in a part of a sequence not yet observed by the generative AI system is conditioned during the training of the generative AI system.

75. The system according to claim 67, wherein, In order to generate text paragraphs, the generative AI system is configured to, for each position in the sequence of text paragraphs to be generated: Based on the context of the text paragraph, calculate a list of multiple candidate positions in the sequence; Estimate the probability of each candidate in a list of multiple candidate positions in a sequence; as well as One of the candidates in the list is added to the position in the sequence based on the probability.

76. The system according to claim 75, wherein: The generative AI system includes multiple generative subsystems that collectively generate the text paragraphs; and The generative AI system is configured as follows: Broadcast the activation value of the selected embedded node from each of the multiple generative subsystems to the other generative subsystems; as well as Modify the selected embedded nodes in each generative subsystem using data-related regularization links between the selected embedded nodes.

77. The system according to claim 76, wherein, The data-related regularization links include "equals" relational regularization links.

78. The system according to claim 77, wherein, In order to generate text paragraphs, the generative AI system is configured to partially: Select a set of key language units relevant to the context, where each language unit is a word or phrase; and Load multiple sample text paragraphs, wherein one or more key language units from the set of key language units appear in each of the multiple sample text paragraphs; and The list is calculated at least in part based on (i) the word count in the example text paragraph and (ii) the autoregressive prediction score of the word at the stated position.

79. The system according to claim 69, wherein, In order to generate text paragraphs, the generative AI system is configured to work partially for each position in the sequence of text paragraphs to be generated: Select a set of key language units that are relevant to the context, where each language unit is a word or phrase; Load multiple example text paragraphs, wherein one or more key language units from the set of key language units appear in each of the multiple example text paragraphs; Calculating a list of multiple candidate positions in the sequence based on the context of the text paragraph, wherein calculating the list of multiple candidate positions includes: calculating the list based on (i) word counts in the example text paragraph and (ii) the output of a nonparametric probability model; Estimate the probability of each candidate in a list of multiple candidate positions in the sequence; and One of the candidates in the list is added to the position in the sequence based on the probability.

80. The system of claim 79, further comprising a back-end computer system, the back-end computer system including the generative AI system, the programmable computer system, and a semantic analysis system configured to test the semantic similarity between candidate text paragraphs and the context.

81. The system according to claim 80, wherein, The generative AI system is configured to generate text paragraphs in part by: The generative AI system generates first and second text paragraphs; The back-end computer system presents the first and second text paragraphs to the user via a user interface; The backend computer system receives the user's selection of one of the first and second text paragraphs; and The computer system programmed thereon further trains the generative AI system based on the selection.

82. The system according to claim 80, wherein: The backend computer system also includes an interpretation system; and The generative AI system is configured to generate text paragraphs in part by: The generative AI system generates text paragraphs. The interpretation system generates interpretations related to the text paragraphs generated by the generative AI system. The backend computer system presents the text paragraphs and explanations to the user via a user interface; The backend computer system receives feedback from the user regarding the explanation; and The programmed computer system further trains the generative AI system based on the feedback.

Images