Method for training a neural network model, method for performing information processing based on a neural network model, apparatus for training a neural network model, computer program, and electronic device

Abstract

[0006] Embodiments of the present disclosure provide a method, apparatus, computer program product, and storage medium for training a neural network model. The method includes: initializing an information-processing neural network model and a reference neural network model corresponding to the information-processing neural network model, where the information-processing neural network model includes a plurality of processing modules and the reference neural network model includes a plurality of reference modules; and updating parameters of the information-processing neural network model and the reference neural network model through multiple iterations. The method of the present disclosure can reduce the difficulty of training the neural network model, improve the training efficiency of the neural network model, and improve the confidentiality of the model training parameters, making it better suited to large-scale neural network model training scenarios or real-time neural network model training scenarios, and avoiding the collapse problem during the neural network model training process.

Description

[Technical Field]

[0001] The present disclosure relates to the field of artificial intelligence, and more particularly to methods, apparatus, computer program products, and storage media for training neural network models, and for performing information processing tasks based on neural network models. [Background technology]

[0002] Artificial intelligence (AI) refers to the theories, methods, technologies, and application systems that use digital computers, or machines controlled by digital computers, to simulate, extend, or enhance human intelligence, sense the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a branch of computer science that seeks to understand the nature of intelligence and create new intelligent machines that can respond in a manner similar to human intelligence. AI primarily studies the design principles and implementation methods of various intelligent machines, endowing them with sensing, reasoning, and decision-making capabilities. Research in the field of AI covers robotics control, natural language processing, computer vision, decision-making and reasoning, human-computer interaction, and information recommendation and retrieval.

[0003] Neural networks (NNs), an important branch of artificial intelligence, are network structures that process information by mimicking the behavioral characteristics of animal neural networks. A neural network is structured by interconnecting a large number of nodes (also called neurons), and achieves information processing goals by learning and training input information based on a specific computational model. A neural network includes an input layer, a hidden layer, and an output layer. The input layer is responsible for receiving input signals, the output layer is responsible for outputting the neural network's computation results, and the hidden layer is responsible for computational processes such as learning and training and is the network's memory unit. The memory function of the hidden layer is represented by a weight matrix, and typically, each neuron corresponds to one weight coefficient.

[0004] As the number of network layers in a neural network model increases, the scale of the network parameters also increases. As the scale of the network parameters increases, the difficulty of model training increases and the time required for training also increases. In particular, self-supervised learning (SSL) models are difficult to train because they may suffer from pattern and dimensional collapse. Therefore, how to effectively improve the training efficiency of neural network models and reduce the time required for model training has become an urgent issue that needs to be resolved. Summary of the Invention [Problem to be solved by the invention]

[0005] Conventional methods for training neural network models typically involve sequentially calculating the outputs of all processing modules in the neural network model through forward propagation, then sequentially calculating the gradient of the loss function for the parameters of each processing module through backpropagation, and then updating the parameters of the entire neural network based on the gradient. In this method, both the forward propagation and backpropagation processes involve all processing modules in the neural network. Furthermore, in the backpropagation process, network parameters can only be updated for each processing module once the network parameters of all subsequent processing modules have been determined. This reduces the training efficiency of the entire model. [Means for solving the problem]

[0006] To improve the training efficiency of a neural network model, the present disclosure provides a method for training a neural network model, the method including the steps of: initializing an information-processing neural network model and a reference neural network model corresponding to the information-processing neural network model, wherein the information-processing neural network model includes a plurality of processing modules and the reference neural network model includes a plurality of reference modules; and updating parameters of the information-processing neural network model and the reference neural network model through a plurality of iterations, wherein in each iteration, determining an output result corresponding to each of the plurality of reference modules based on a training sample and the parameters of the plurality of reference modules, determining an output result corresponding to each of the plurality of processing modules based on a mask sample obtained by masking the training sample and the parameters of the plurality of processing modules, updating the parameters of the plurality of processing modules based on the output result corresponding to each of the plurality of processing modules and the output result corresponding to each of the plurality of reference modules, and updating the parameters of the plurality of reference modules based on the updated parameters of the plurality of processing modules.

[0007] An embodiment of the present disclosure further provides an apparatus for training a neural network model, the apparatus including: an initialization module configured to initialize an information-processing neural network model and a reference neural network model corresponding to the information-processing neural network model, the information-processing neural network model including a plurality of processing modules, and the reference neural network model including a plurality of reference modules; and a parameter updating module configured to update parameters of the information-processing neural network model and the reference neural network model through a plurality of iterations, the parameter updating module being configured to: determine, in each iteration, an output result corresponding to each of the plurality of reference modules based on a training sample and parameters of the plurality of reference modules; determine an output result corresponding to each of the plurality of processing modules based on mask samples obtained by masking the training sample and parameters of the plurality of processing modules; update the parameters of the plurality of processing modules based on the output result corresponding to each of the plurality of processing modules and the output result corresponding to each of the plurality of reference modules; and update the parameters of the plurality of reference modules based on the updated parameters of the plurality of processing modules.

[0008] An embodiment of the present disclosure further provides a method for performing information processing based on a neural network model, the method including: acquiring target information, the target information including at least one of image information, text information, audio information, and video information; and processing the target information based on the information-processing neural network model to acquire a result of processing the target information, the information-processing neural network model being trained by the method for training a neural network model described above.

[0009] An embodiment of the present disclosure further provides a computer program product, which includes computer software code that, when executed by a processor, provides the above method.

[0010] An embodiment of the present disclosure further provides a computer-readable storage medium having stored thereon computer-executable instructions that, when executed by a processor, provide the above method. [Effects of the Invention]

[0011] The disclosed method for training a neural network model can train each processing module in an information processing neural network model independently, and can train each processing module in parallel, thereby effectively improving the training efficiency of the neural network model.

[0012] The disclosed method for training a neural network model can train a processing module based only on the output result of the processing module and the output result of the reference module corresponding to the processing module, which does not require specifying the gradient of the loss function with respect to the parameters of other processing modules, thereby reducing the difficulty of training the neural network model, effectively saving computer resources, and reducing the requirements for computer performance.

[0013] In addition, since each processing module in the information processing neural network model can be trained independently and there is no need to share the network parameters between the processing modules, the disclosed method for training a neural network model has good confidentiality, can better protect the parameters of each processing module, and can prevent the parameters from being stolen. Therefore, the disclosed method for training a neural network model has obvious advantages in application scenarios with high confidentiality requirements (e.g., those related to business secrets, privacy, etc.).

[0014] The disclosed method for training a neural network model can effectively reduce the amount of computation required by a computer and utilize computer resources to achieve parallel training of multiple processing modules, thereby reducing the training time of the neural network model. The disclosed method for training a neural network model can train each processing module separately, resulting in high training efficiency, making it more suitable for training scenarios of large-scale neural network models or real-time training scenarios of neural network models (i.e., training is performed while task processing is being performed, rather than the task processing process and the training process being completely independent).

[0015] Furthermore, compared to the conventional neural network model training method based on backpropagation, the neural network model training method disclosed herein can stably train the neural network model and effectively avoid the collapse problem during the neural network training process. [Brief explanation of the drawings]

[0016] In order to more clearly explain the configuration of the embodiments of the present disclosure, the following briefly introduces the drawings necessary for the description of the embodiments. Obviously, the drawings in the following description only show some exemplary embodiments of the present disclosure, and those skilled in the art can also obtain other drawings from these drawings without creative work. [Figure 1] 1 shows a schematic diagram of an application scenario according to an embodiment of the present disclosure; [Figure 2] FIG. 1 is an exemplary schematic diagram illustrating a scenario of information processing and training based on a neural network model according to an embodiment of the present disclosure. [Figure 3A] 1 is a schematic flow chart illustrating a method for training a neural network model according to an embodiment of the present disclosure. [Figure 3B] 1 is a schematic flow chart illustrating a method for training a neural network model according to an embodiment of the present disclosure. [Figure 4A] FIG. 1 is a schematic diagram illustrating a process for processing training samples according to an embodiment of the present disclosure. [Figure 4B] FIG. 1 is a schematic diagram illustrating a process for processing training samples according to an embodiment of the present disclosure. [Figure 4C] FIG. 1 is a schematic diagram illustrating a process for processing training samples according to an embodiment of the present disclosure. [Figure 4D] FIG. 1 is a schematic diagram illustrating a process for processing training samples according to an embodiment of the present disclosure. [Figure 4E] FIG. 1 is a schematic diagram illustrating a process for processing training samples according to an embodiment of the present disclosure. [Figure 5] FIG. 1 is a schematic diagram illustrating a process of training a neural network model according to an embodiment of the present disclosure. [Figure 6] 1 is an exemplary flowchart illustrating a method for performing information processing based on a neural network model according to an embodiment of the present disclosure. [Figure 7] FIG. 1 is a schematic diagram illustrating an information processing process based on a neural network model according to an embodiment of the present disclosure. [Figure 8] FIG. 1 is a structural schematic diagram illustrating an apparatus for training a neural network model according to an embodiment of the present disclosure. [Figure 9] FIG. 1 is a schematic diagram illustrating a configuration of an apparatus that performs information processing based on a neural network model according to an embodiment of the present disclosure. [Figure 10] 1 is an architecture illustrating a computing device according to an embodiment of the present disclosure. DETAILED DESCRIPTION OF THE INVENTION

[0017] In order to clarify the objectives, configurations, and advantages of the present disclosure, exemplary embodiments according to the present disclosure will be described in detail below with reference to the drawings. It is apparent that the described embodiments are only some of the embodiments of the present disclosure, and do not cover all of the embodiments of the present disclosure. It should be understood that the present disclosure is not limited by the exemplary embodiments described herein.

[0018] Furthermore, in this specification and drawings, steps and elements that are essentially the same or similar are denoted by the same or similar reference numerals, and redundant descriptions of these steps and elements are omitted.

[0019] Also, in the specification and drawings, according to the examples, elements are described in singular or plural. However, the singular and plural forms have been selected as appropriate for the presented circumstances merely for ease of interpretation and are not intended to limit the present disclosure. Thus, unless the context clearly dictates otherwise, the singular can include the plural and vice versa.

[0020] In this specification and drawings, steps and elements that are essentially the same or similar are denoted by the same or similar reference numerals, and redundant descriptions of these steps and elements are omitted. In addition, in the description of this disclosure, terms such as "first," "second," etc. are merely used for distinguishing and describing purposes, and cannot be understood as indicating or suggesting relative importance or rank.

[0021] Each of the following various neural networks (or neural network models) usable in the embodiments of the present disclosure may be an artificial intelligence model, particularly an artificial intelligence-based neural network model. Typically, an artificial intelligence-based neural network model is implemented as an acyclic graph, where neurons are arranged in different layers. Typically, a neural network model includes an input layer and an output layer, which are separated by at least one hidden layer. The hidden layer converts inputs received at the input layer into a representation useful for generating outputs at the output layer. Network nodes (i.e., neurons) are fully connected to nodes in adjacent layers via edges, and there are no edges between nodes within each layer. Data received at nodes in the input layer of a neural network is propagated to nodes in the output layer via any one of a hidden layer, an activation layer, a pooling layer, a convolutional layer, etc. Inputs and outputs of a neural network may take various forms, but the present disclosure is not limited thereto.

[0022] In this disclosure, we consider that the lack of rationality and local plasticity in backpropagation of neural network models (e.g., SSL models) remains a fundamental unsolved problem. Therefore, we propose a new biologically rational algorithm. This algorithm is implemented in a manner closer to satisfying neural circuit constraints and utilizes only locally available information. Under the framework of an energy-based model, all free variables in the proposed model are autonomously optimized, minimizing a predictive masking coding function to define local and global energy. This eliminates collapse and helps stabilize pre-training. The configuration in the embodiments of this disclosure can be used for various neural network models, including SSL models and supervised models.

[0023] Hereinafter, embodiments of the present disclosure will be further described with reference to the drawings.

[0024] First, an application scenario of a method and a corresponding device according to an embodiment of the present disclosure will be described with reference to Fig. 1. Fig. 1 shows a schematic diagram of an application scenario 100 according to an embodiment of the present disclosure. In the figure, a server 110 and multiple terminals 120 are schematically shown.

[0025] The information processing neural network model of the embodiment of the present disclosure may be specifically incorporated into various electronic devices, for example, any of the electronic devices among the server 110 and the multiple terminals 120 in FIG. 1 . For example, the information processing neural network model may be incorporated into the terminal 120. The terminal 120 may be, but is not limited to, a smartphone, a tablet computer, a laptop computer, a desktop computer, a personal computer (PC), a smart speaker, or a smart watch. Furthermore, for example, the information processing neural network model may be incorporated into the server 110. The server 110 may be an independent physical server, a server cluster or a distributed system consisting of multiple physical servers, or a cloud server that provides cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, a content delivery network (CDN), and base cloud computing services such as big data and artificial intelligence platforms. The terminals and the server may be directly or indirectly connected via wired or wireless communication. The present disclosure is not limited thereto.

[0026] As can be understood, the device for performing information processing based on a neural network model to which the embodiments of the present disclosure are applied may be a terminal, a server, or a system configured with a terminal and a server, and the information processing method to which the embodiments of the present disclosure are applied may be executed by the terminal, the server, or jointly executed by the terminal and the server.

[0027] The information processing neural network model provided in the embodiments of the present disclosure can be used to perform various information processing tasks, including, but not limited to, information extraction (e.g., key information extraction, information retrieval, feature extraction, etc.), information classification (e.g., image classification, disease diagnosis, junk information recognition), information restoration (e.g., image restoration, missing information prediction, etc.), information style transition (e.g., image style transition, tone conversion, etc.), information enhancement (e.g., image clarity enhancement, audio noise removal, etc.), information mining (e.g., big data mining, network information mining, etc.), information recognition (e.g., disease diagnosis, junk information recognition, etc.), machine translation, etc. In the present disclosure, the information processed based on the neural network model may include images, text, audio, video, number sequences, etc.

[0028] The information processing neural network model provided in the embodiments of the present disclosure may also be related to artificial intelligence cloud services in the field of cloud technology. Here, cloud technology refers to a hosting technology that integrates a series of resources, such as hardware, software, and networks, within a wide area network or a local area network to realize data calculation, storage, processing, and sharing. Cloud technology is a collective term for network technology, information technology, integration technology, management platform technology, and application technology applied based on the cloud computing business model, forming a resource pool and enabling on-demand utilization, which is flexible and convenient. Cloud computing technology is an important support. The back-end services of technology network systems require large amounts of computing and storage resources, such as video sites, picture sites, and more portal sites. With the rapid development and application of the Internet industry, in the future, individual items may have their own identification marks, which will all need to be transmitted to a back-end system for logic processing. Different levels of data will be processed separately, and various types of industry data will require powerful system support, which can only be achieved through cloud computing.

[0029] AI cloud services are commonly referred to as AIaaS (AI as a Service). This is the service model of the current mainstream AI platforms. Specifically, AIaaS platforms separate several standard AI services and provide them as independent or packaged services on the cloud. This service model is similar to opening an AI-themed shopping mall. All developers can access and use one or more AI services provided by the platform through an application programming interface (API). Some experienced developers can even use the AI ​​framework and AI infrastructure provided by the platform to deploy and manage their own cloud AI services.

[0030] FIG. 2 is an exemplary schematic diagram illustrating a scenario 200 for information processing and training based on a neural network model according to an embodiment of the present disclosure.

[0031] During the training phase, the server 110 may train a neural network model based on the training samples. Once training is complete, the server may deploy the trained neural network model to one or more servers (or cloud servers), thereby providing artificial intelligence services related to information processing based on the neural network model. It should be noted that all training samples used in this disclosure comply with legality, morality, and privacy regulations. Specifically, all training samples are from legitimate sources and are explicitly authorized by users during collection. Furthermore, all training samples used in this disclosure comply with privacy protection principles. These training samples are rigorously screened and cleaned, and will not be disclosed to unauthorized third parties.

[0032] At the stage of performing information processing based on the neural network model, it is assumed that a client or application (e.g., an image processing application, a text processing application, etc.) that interacts with the information processing server 110 is installed on the user terminal 120. The user terminal 120 may transmit an information processing request to the server 110 corresponding to the application via a network, requesting the neural network disposed on the server 110 to process the information. For example, after receiving the information processing request, the server 110 uses a trained neural network model to process the information in response to the information processing request, and feeds back the predicted information processing result to the user terminal 120. The user terminal 120 may receive the information processing result. The user terminal 120 may then perform further analysis or processing based on the information processing result.

[0033] It should be noted that the training sample data shown in FIG. 2 may be updated in real time. For example, a user may score the information processing result. For example, if the user recognizes that the information processing result is both highly relevant and accurate, the user may give the information processing result a high score, and the server 110 may use the information processing result as a positive sample for training the neural network model in real time. If the user gives the information processing result a low score, the server 110 may use the information processing result as a negative sample.

[0034] The training sample set shown in Figure 2 may be preset. For example, referring to Figure 2, the server may obtain training data (e.g., image training samples, text training samples, audio training samples, video training samples, etc.) from a database and then generate a training sample set for the neural network model. Of course, the present disclosure is not limited thereto.

[0035] FIG. 3A is a schematic flow chart illustrating a method 300 for training a neural network model according to an embodiment of the present disclosure.

[0036] In step S310, an information-processing neural network model and a reference neural network model corresponding to the information-processing neural network model are initialized. In one example, the relationship between the reference neural network model and the information-processing neural network model can be understood to be similar to the relationship between teacher-student models in the learning process of knowledge distillation. The purpose of establishing the reference neural network is to obtain the learning goal of the information-processing neural network, thereby training the information-processing neural network and stabilizing the training to prevent collapse. In one embodiment, the structure and parameters of the corresponding reference neural network model can be determined based on the structure and parameters of the information-processing neural network model.

[0037] Note that the information processing neural network model of the present disclosure may include multiple processing modules, and the reference neural network model may include multiple reference modules.

[0038] According to an embodiment of the present disclosure, the plurality of processing modules may be L processing modules, and the plurality of reference modules may be L reference modules, and the l-th processing module among the L processing modules and the l-th reference module among the L reference modules have the same structure, where L and l are positive integers, and l≦L.

[0039] It should be understood that each processing module of the plurality of processing modules may have the same structure or functionality as the other processing modules, or may have different structures or functionality, and similarly, each reference module of the plurality of reference modules may have the same structure or functionality as the other reference modules, or may have different structures or functionality.

[0040] According to an embodiment of the present disclosure, the initial parameters of the information-processing neural network model and the reference neural network model may be determined based on random numbers. Alternatively, the initial parameters of the information-processing neural network model and the reference neural network model may be determined based on the experience of an engineer. The initial parameters of the information-processing neural network model and the reference neural network model may be the same or different. Optionally, the information-processing neural network model may be pre-trained a predetermined number of times, and then the parameters of the information-processing neural network model after the pre-training may be used as the initial parameters of the reference neural network model.

[0041] Each of the processing modules or each of the reference modules may be a sub-neural network model or a neural network layer, such as a convolutional neural network, an attention-based neural network, a recurrent neural network, a recurrent neural network, a feedforward neural network, a generative adversarial neural network, or a deep neural network.

[0042] In step S320, the parameters of the information processing neural network model and the reference neural network model are updated through multiple iterations.

[0043] To more clearly explain step S320, Fig. 3B further illustrates the processing steps in each iteration. As shown in Fig. 3B, step S320 may specifically include step S321, step S322, step S323, and step S324.

[0044] In step S321, output results corresponding to each of the plurality of reference modules are determined based on the training samples and the parameters of the plurality of reference modules. In one example, the specific process of determining the output results corresponding to each of the plurality of reference modules may be to input the training samples to the plurality of reference modules, perform forward calculation, and output the corresponding calculation results. This process may not include gradient backward propagation. Therefore, the parameters of the reference modules in this process may be fixed so as not to change.

[0045] It should be understood that the training samples may include one or more of image training samples, text training samples, audio training samples, video (which may optionally be treated as both image and audio) training samples.

[0046] In step S322, output results corresponding to each of the multiple processing modules are determined based on mask samples obtained by masking the training samples and parameters of the multiple processing modules. In one example, the specific process of determining output results corresponding to each of the multiple processing modules may be to input the mask samples to the multiple processing modules, perform forward calculations, and output corresponding calculation results. This process may also not include gradient backpropagation. Therefore, the parameters of the processing modules in this process may be fixed so as not to change.

[0047] For example, in the case of the image training samples in FIG. 4A, masking can be performed on image training sample P0 to obtain mask samples P0-1, P0-2, and so on, and mask samples P0-i, all of which have different mask positions. For example, in the case of the text training samples in FIG. 4B, masking can be performed on text training sample T0 to obtain mask samples T0-1, T0-2, and so on, and mask samples T0-i, all of which have different mask positions. For example, in the case of the audio training samples in FIG. 4C, masking can be performed on audio training sample V0 to obtain mask samples V0-1, V0-2, and so on, and mask samples V0-i, all of which have different mask positions. In the examples of FIGS. 4A to 4C, i is a positive integer.

[0048] According to an embodiment of the present disclosure, masking the training samples may include masking the training samples with different sizes (e.g., lengths, dimensions, etc.), different positions, or different shapes (e.g., square, circular, irregular, etc.). For example, taking the image training sample in FIG. 4D as an example, masking the image training sample P1 can obtain mask samples P1-1, P1-2, and P1-j with different mask sizes and positions, where j is a positive integer.

[0049] By performing different mask processing on the training samples, the diversity of the mask samples processed by the processing module can be increased, improving the performance of the trained processing module, and thus more accurate processing results can be obtained based on the trained processing module.

[0050] In step S323, the parameters of the plurality of processing modules are updated based on the output results corresponding to each of the plurality of processing modules and the output results corresponding to each of the plurality of reference modules.

[0051] According to an embodiment of the present disclosure, for an l-th processing module among the L processing modules, the parameters of the l-th processing module may be updated based on the output result of the l-th processing module and the output result of a reference module corresponding to the l-th processing module.

[0052] Note that the reference module corresponding to the lth processing module may be, but is not limited to, the lth reference module among the L reference modules. Optionally, the reference module corresponding to the lth processing module may include one or more reference modules surrounding the lth reference module. For example, the reference module corresponding to the lth processing module may include the lmth reference module to the l+nth reference module. Here, m and n are integers and satisfy m≧0, n≧0, lm≧1, and l+n≦L, and the values ​​of m and n may be the same or different.

[0053] According to an embodiment of the present disclosure, an energy function value may be calculated based on output results corresponding to each of the plurality of processing modules and output results corresponding to each of the plurality of reference modules, and parameters of the plurality of processing modules may be updated based on the energy function value, where the energy function is for evaluating accuracy of the output results of the plurality of processing modules.

[0054] It should be understood that the energy function may be a global energy function (i.e., a common global optimization goal of all processing modules) for evaluating the difference between the output result of the Lth processing module and the output result of the Lth reference module, or may be a local energy function (i.e., an optimization goal of an individual processing module) for evaluating the difference between the output result of the lth processing module and the output result of a reference module corresponding to the lth processing module. Optionally, the energy function may be jointly determined based on the global energy function and the local energy function (e.g., by adding them together, calculating an average value, or calculating a weighted average value).

[0055] According to embodiments of the present disclosure, the form of the energy function (whether a global energy function or a local energy function) may vary.

[0056] For example, for an l-th processing module among the L processing modules, the value of the first local energy function may be determined based on the difference between the output result of the l-th processing module and the output result of a reference module corresponding to the l-th processing module.

[0057] Further, a predicted output result of the lth processing module may be determined based on the global energy function, and a value of a second local energy function may be determined based on the difference between the output result of the lth processing module and the predicted output result of the lth processing module.

[0058] Specifically, in the process of determining the expected output result of the l-th processing module based on the global energy function, a backward calculation difference of the output result of the l-th processing module may be determined based on the value of the global energy function, and then the expected output result of the l-th processing module may be determined based on the backward calculation difference of the output result of the l-th processing module and the output result of the l-th processing module. The backward calculation difference of the output result of the l-th processing module is obtained by sequentially propagating the difference between the output result of the L-th processing module and the output result of the L-th reference module to the l-th processing module.

[0059] For example, for the l+k-th processing module among the L-th to l-th processing modules, a partial derivative of the output result of the l+k-1-th processing module may be calculated (the output result of the l+k-1-th processing module is the input of the l+k-1-th processing module), and then a backward calculation difference of the output result of the l+k-th processing module may be calculated based on the partial derivative and the backward calculation difference of the output result of the l+k+1-th processing module, where k is an integer and satisfies k≧0 and l+k+1≦L.

[0060] According to an embodiment of the present disclosure, one overall local energy function may be determined based on different local energy functions, and the parameters of the l-th processing module may be updated based on the overall local energy function.

[0061] For example, a third local energy function may be determined based on the first and second local energy functions, and a parameter of the lth processing module may be updated based on the third local energy function. In one embodiment, the third local energy function may be determined by performing a weighting calculation on the first and second local energy functions.

[0062] According to an embodiment of the present disclosure, if different, the parameters of the l-th processing module may be updated based on different local energy functions.

[0063] For example, if the first local energy function corresponding to the lth processing module does not converge, the parameters of the lth processing module may be updated based on the value of the first local energy function corresponding to the lth processing module (i.e., taking into account only the optimization goal of each individual processing module), and if the first local energy function corresponding to the lth processing module converges, the parameters of the lth processing module may be updated based on the second local energy function (determined based on the global energy function) corresponding to the lth processing module (i.e., taking into account the common global optimization goal of all processing modules).

[0064] In step S324, the parameters of the plurality of reference modules are updated based on the updated parameters of the plurality of processing modules.

[0065] According to an embodiment of the present disclosure, the parameters of the plurality of reference modules may be updated using the values ​​of the parameters of the plurality of processing modules in at least one previous iteration. For example, the parameters of the plurality of reference modules may be updated using an exponential moving average value of the parameters of the plurality of processing modules in at least one previous iteration. That is, the updated parameters of the processing modules may be TIFF2026502558000002.tif4170 and the current reference module parameters are Assuming the file is TIFF2026502558000003.tif4170, the updated reference module parameters are: TIFF2026502558000004.tif5170, where TIFF2026502558000005.tif4170. The parameters of the reference modules may be updated using an arithmetic average or weighted average (optionally, the weight of the latest parameter of a processing module may be set greater than the weight of the previous parameter) of the parameters of the processing modules in at least one previous iteration. The present disclosure is not limited thereto.

[0066] It should be understood that the above method 300 for training a neural network model can be used to perform supervised training of the information-processing neural network model, or to perform self-supervised training of the information-processing neural network model.

[0067] Taking the image training sample shown in FIG. 4A as an example, when performing self-supervised training on the information processing neural network model, the process of processing the training sample may be as shown in FIG. 4E.

[0068] In step S1, the training image sample P0 may be masked at different positions to obtain mask samples P0-1...P0-i. Optionally, the masking of the training image samples may be performed based on information contained in the image training samples. For example, in the case of an information processing neural network for recognizing people in images, finer-grained masking (e.g., smaller mask sizes) may be performed on the portions of the image where people are located, while coarser-grained masking (e.g., larger mask sizes) may be performed on the portions of the image where the surrounding environment is located, so that the trained neural network model will focus on learning the features of the people.

[0069] In step S2, the structure and parameters of the corresponding reference neural network model may be determined based on the structure and parameters of the information-processing neural network model. For example, the structure of each processing module in the information-processing neural network model may correspond one-to-one to the structure of each reference module in the reference neural network model. Both can extract image features, but the difference is that the reference neural network model extracts image features based on complete training image samples P0, while the information-processing neural network model extracts image features based on mask samples. The parameters of each reference module in the reference neural network model may be determined based on the historical results of the parameters of the corresponding processing module. For example, before training the information-processing neural network model based on the reference neural network model, the information-processing neural network model may be subjected to a predetermined number of supervised trainings (e.g., but not limited to, 50 times) using annotated image training samples (note that each of the predetermined number of supervised trainings here refers to one epoch, not one batch, of training). Next, the parameters of the information-processing neural network model that have been updated a predetermined number of times are used as initial parameters of the reference neural network model, and then the information-processing neural network model is trained based on the reference neural network model, and the parameters of the information-processing neural network model and the reference neural network model are updated. During the training process of the information-processing neural network model, each time the parameters of the information-processing neural network model are updated, the parameters of the corresponding reference module are also updated using the parameters of the processing module after the update.

[0070] In step S3, a reference neural network model may be used to perform forward processing on the training image sample P0 to obtain a feature representation F0 of the training image sample P0.

[0071] Additionally, in step S4, a neural network model may be used to perform forward processing on the mask samples P0-1...mask samples P0-i, and predict a predicted feature representation F0-1...F0-i of the training image sample P0 based on the mask samples P0-1...mask samples P0-i. The predicted feature representation is composed of two parts: a feature representation for the unmasked portion (i.e., the portion of the mask sample shown in FIG. 4E that is filled with a diagonal texture) and a feature representation for the masked portion (i.e., the portion of the mask sample shown in FIG. 4E that is filled with a dotted texture).

[0072] The reference neural network model can obtain a complete feature representation F0 of the training image sample P0 by performing feature encoding on the complete training image sample P0, whereas the information processing neural network model performs feature encoding on the masked sample, and therefore can only obtain a predicted feature representation of the training image sample P0 by first obtaining a feature representation of the unmasked part and then decoding and predicting the feature representation of the masked part based on the feature representation of the unmasked part.

[0073] Next, in step S5, an energy function may be calculated based on the difference between the output result corresponding to the information-processing neural network model (i.e., the feature representations F0-1...F0-i of mask sample P0-1...mask sample P0-i) and the output result corresponding to the reference neural network model (i.e., the feature representation F0 of training image sample P0), and the parameters of the information-processing neural network model may be updated based on the value of the energy function.

[0074] In the embodiment of FIG. 4E, the information processing neural network model can be used to extract image features or predict (or repair) missing parts of an image. In the process shown in FIG. 4E, the training of the information processing neural network model can be achieved simply by predetermining the image training samples and the structure of the information processing neural network model. This process does not require providing labels for the training samples, and does not require the effort and labor costs of annotating the training samples, significantly improving the processing efficiency of the image processing neural network model. For example, in the process of training a neural network model for medical image processing, annotating medical images requires the careful judgment of a specialist, resulting in low annotation efficiency. The method disclosed herein can effectively reduce annotation costs and improve model training efficiency.

[0075] 4A and 4E, the present disclosure further discloses a method for training a neural network model for image processing. Optionally, the neural network model for image processing is an example of the above information processing neural network model. The method includes the steps of: initializing a neural network model for image processing and a reference neural network model corresponding to the neural network model for image processing, wherein the neural network model for image processing includes a plurality of processing modules and the reference neural network model includes a plurality of reference modules; and updating parameters of the neural network model for image processing and the reference neural network model through a plurality of iterations, wherein in each iteration, determining image features corresponding to each of the plurality of reference modules based on image training samples and parameters of the plurality of reference modules, determining image features corresponding to each of the plurality of processing modules based on mask samples obtained by masking the image training samples and parameters of the plurality of processing modules, updating parameters of the plurality of processing modules based on the image features corresponding to each of the plurality of processing modules and the image features corresponding to each of the plurality of reference modules, and updating parameters of the plurality of reference modules based on the updated parameters of the plurality of processing modules. Optionally, the trained neural network model for image processing may be used to extract features from image data and further complete tasks such as face recognition or disease diagnosis.

[0076] Furthermore, when there are predetermined training sample labels, the information-processing neural network model may be subjected to supervised training in a process similar to that of Fig. 4E. That is, the method 300 is suitable for both supervised training and self-supervised training. When the information-processing neural network model is subjected to supervised training, a global energy function may be determined based on the difference between the output result of the Lth processing module and the predetermined training sample labels (i.e., the result that the Lth processing module should output).

[0077] If the updated information processing neural network model satisfies a predetermined condition (e.g., the energy function converges, a preset number of training times is reached, etc.), it is determined that the training of the information processing neural network model is complete, and a trained information processing neural network model can be obtained. In this case, the trained information processing neural network model can be used for feature extraction (of multimedia information such as images, text, audio, and video), missing information prediction (including at least one of multimedia information such as images, text, audio, and video), etc.

[0078] Optionally, the trained information processing neural network model may be further trained (e.g., fine-tune the structure and parameters of the information processing neural network model) using information processing training samples corresponding to the information processing tasks to complete downstream information processing tasks such as information extraction, information classification, information restoration, information style transition, information enhancement, information mining, and information recognition.

[0079] FIG. 5 is a schematic diagram illustrating a process of training a neural network model according to an embodiment of the present disclosure.

[0080] FIG. 5 illustrates a process for training an information-processing neural network model using an audio training sample V1 as an example. It should be understood that a plurality of different mask samples V1-1, V1-2, ..., V1-j may be obtained by performing mask processing on the audio training sample V1 in the manner shown in FIG. 4C, where j is a positive integer. FIG. 5 illustrates only the mask sample V1-1 as an example. The process for training an information-processing neural network model using other mask samples is similar to that shown in FIG. 5, and further description thereof will be omitted here.

[0081] As shown in Figure 5, the information-processing neural network model M1 includes L processing modules (as shown by blocks M1-1, M1-2, ... M1-L filled with a dotted texture in Figure 5). Based on the information-processing neural network model M1 including L processing modules, a reference neural network model M2 including L reference modules (as shown by blocks M2-1, M2-2, ... M2-L filled with a diagonal line texture in Figure 5) can be determined. L is a positive integer. Each processing module and each reference module in Figure 5 may have the same network structure.

[0082] For example, each reference module may include a neural network layer, such as an attention-based neural network layer, a fully connected layer, or a linear layer, to jointly realize feature extraction of the audio training sample V1. Each reference module may adopt the same network structure. Similarly, each processing module may also include a neural network layer, such as an attention-based neural network layer, a fully connected layer, or a linear layer, to jointly realize feature extraction of the masked audio training sample V1-1. Each reference module may adopt the same network structure, i.e., the same network structure as the processing module.

[0083] It should be understood that in the example of Figure 5, the information processing neural network model M1 includes L processing modules for feature extraction with the same structure in order to more accurately obtain the features of the audio training sample. In fact, each processing module is a complete sub-neural network, which can be used to obtain the features of the audio training sample and predict the information of the missing part when some content of the audio training sample is missing.

[0084] Optionally, each processing module of the L processing modules or each reference module of the L reference modules may have a different network structure as needed.

[0085] In the information processing neural network model M1, the input of the lth processing module is TIFF2026502558000006.tif4170 (as shown by the black dots in Figure 5), and the output is It can be expressed as TIFF2026502558000007.tif5170, where x represents the feature vector, TIFF2026502558000008.tif5170 is the network parameter of the l-th processing module, where l is a positive integer and l≦L. In the example of FIG. 5, the processing modules are in a serial relationship, that is, the output of the l-th processing module is the input of the l+1-th processing module, that is, The file is TIFF2026502558000009.tif5170.

[0086] Similarly, in the reference neural network model M2, the input of the lth reference module is It can be expressed as TIFF2026502558000010.tif5170 and the output is It can be expressed as TIFF2026502558000011.tif5170, where: TIFF2026502558000012.tif5170 is the forward propagation function, TIFF2026502558000013.tif5170 is the network parameter of the l-th reference module, where l is a positive integer and l≦L. In the example of FIG. 5, the reference modules are in a serial relationship, that is, the output of the l-th reference module is the input of the l+1-th reference module, that is, The file is TIFF2026502558000014.tif5170.

[0087] In the embodiment shown in FIG. 5, each processing module can be trained independently. After calculating the forward calculation results of each processing module, it is not necessary to sequentially obtain the gradient of the loss function for the parameters of each processing module through backward calculation, and then update the parameters of the entire neural network based on the gradient.

[0088] For example, for the lth processing module among the L processing modules, the lth processing module may be trained based on the output result of the lth processing module and the output result of a reference module corresponding to the lth processing module, and the parameters of the lth processing module may be updated.

[0089] 5, the reference module corresponding to the lth processing module may include the lth reference module and reference modules surrounding the lth reference module. For example, the reference module corresponding to the lth processing module may include the l-1th reference module, the lth reference module, and the l+1th reference module, where l-1≧1 and l+1≦L. The reference module corresponding to the first processing module may include the first reference module, and the reference module corresponding to the Lth processing module may include the Lth reference module. Optionally, the reference module corresponding to the first processing module may include the first reference module and one or more subsequent reference modules, and the reference module corresponding to the Lth processing module may include the Lth reference module and one or more previous reference modules.

[0090] an output reference result for training the l-th processing module based on the output result of a reference module corresponding to the l-th processing module; TIFF2026502558000015.tif5170 (as shown by the cross in FIG. 5 ) can be determined. For example, if the reference modules corresponding to the l-th processing module include the l-1th reference module, the l-th reference module, and the l+1th reference module, the reference result corresponding to the l-th processing module can be determined by calculating the average or weighted average (for example, the weight of the output result of the l-1th reference module may be set larger than the weights of the output results of the other two reference modules) of the output result of the l-1th reference module, the output result of the l-th reference module, and the output result of the l+1th reference module. TIFF2026502558000016.tif5170 may be determined.

[0091] Next, the output reference result corresponding to the lth processing module TIFF2026502558000017.tif5170 and the output of the lth processing module TIFF2026502558000018.tif5170, an energy function for training the lth processing module may be determined.

[0092] Energy function for all information processing neural network models TIFF2026502558000019.tif4170 can be calculated using Formula 1.

[0093] [Formula 1] TIFF2026502558000020.tif6170 where, TIFF2026502558000021.tif5170 is local energy, and TIFF2026502558000022.tif6170, where TIFF2026502558000023.tif5170 represents the local energy corresponding to the lth processing module, TIFF2026502558000024.tif5170 is the global energy function, TIFF2026502558000025.tif4170 shows the degree of impact of global energy on total energy, TIFF2026502558000026.tif4170. If we only consider the optimization goals of individual processing modules, The file is TIFF2026502558000027.tif4170.

[0094] According to an embodiment of the present disclosure, to independently train each processing module, the lth processing module may be trained using a local energy function corresponding to the lth processing module, where the local energy function corresponding to the lth processing module is for evaluating the difference between the output result of the lth processing module and the output result of the corresponding reference module.

[0095] For example, considering the l-th processing module's own local optimization goal and its local optimization goal influenced by the common global optimization goal of all processing modules, the local energy function corresponding to the l-th processing module is TIFF2026502558000028.tif5170 can be calculated using Equation 2.

[0096] [Formula 2] TIFF2026502558000029.tif14170 where B is the batch size, T represents the masked frames (for simplicity, we omit the subscripts here), D is the dimension of the feature vector x, and Dec(·) represents the predicted feature representation for the masked part. TIFF2026502558000030.tif5170 represents the expected output result of the lth processing module, TIFF2026502558000031.tif5170 and The value of TIFF2026502558000032.tif5170 is either 0 or 1, and vice versa. If TIFF2026502558000033.tif5170 is 1 (i.e., only the optimization goal of the lth processing module is considered), TIFF2026502558000034.tif5170 (hereinafter referred to as the first local energy function) represents the difference between the feature representation predicted for the mask portion by the lth processing module and the feature representation encoded for the mask portion by the lth reference module. If TIFF2026502558000035.tif5170 is 1 (i.e., if a common global optimization goal for all processing modules is considered), TIFF2026502558000036.tif5170 (hereinafter referred to as the second local energy function) represents the difference between the output result of the lth processing module and the expected output result of the lth reference module, where the expected output result of the lth reference module is determined based on the output result of the Lth processing module.

[0097] According to an embodiment of the present disclosure, first, a first local energy function corresponding to an l-th processing module is used to train the l-th processing module, and then, when the first local energy function corresponding to the l-th processing module converges (in this case, TIFF2026502558000037.tif8170), and may update parameters of the lth processing module based on a second local energy function.

[0098] As shown in Equation 3, the second local energy function is TIFF2026502558000038.tif5170 (i.e., the feature representation Dec( TIFF2026502558000039.tif4170) and the output result of the Lth reference module The second local energy function reflects a common global optimization goal for all processing modules.

[0099] [Formula 3] TIFF2026502558000041.tif15170 Expected output of the lth processing module above TIFF2026502558000042.tif5170 is the global energy function For example, the expected output of the lth processing module can be determined by the following steps: TIFF2026502558000044.tif5170 may be determined.

[0100] When the local energy function converges, the energy function has the properties shown in Equation 4.

[0101] [Formula 4] TIFF2026502558000045.tif10170Therefore, for the lth processing module, Equation 5 can be obtained.

[0102] [Formula 5] TIFF2026502558000046.tif11170 where, TIFF2026502558000047.tif4170 represents the backward calculation difference. The backward calculation represents sequential calculation from the Lth processing module forward (i.e., the Lth processing module, the L-1th processing module, ..., the lth processing module).

[0103] Furthermore, the difference of the Lth processing module may be propagated to the lth processing module by determining the difference of the backward calculation for each processing module from the Lth processing module to the lth processing module in sequence.

[0104] For example, for the l+kth processing module from the Lth processing module to the lth processing module, the output result of the l+kth processing module Output result of the l+k-1th processing module for TIFF2026502558000048.tif5170 Partial derivatives for TIFF2026502558000049.tif4170 TIFF2026502558000050.tif8170, and then calculate this partial derivative Difference between TIFF2026502558000051.tif8170 and the output of the l+k+1th processing module TIFF2026502558000052.tif4170 and the l+kth processing module output result backward calculation difference TIFF2026502558000053.tif4170 may be calculated, where k is an integer and satisfies k≧0 and l+k+1≦L.

[0105] That is, according to the derivation, the backward calculation difference of the output result of the l+kth processing module can be determined by Equation 6. Based on TIFF2026502558000054.tif8170 I get TIFF2026502558000055.tif4170.

[0106] [Formula 6] TIFF2026502558000056.tif10170 Backward calculation difference of the output result of the lth processing module based on Equation 6 After determining TIFF2026502558000057.tif4170, the backward calculation difference of the output result of the lth processing module TIFF2026502558000058.tif4170 and the output result of the lth processing module TIFF2026502558000059.tif4170 and the expected output result of the lth processing module TIFF2026502558000060.tif5170 may be determined. For example, in Equation 7, TIFF2026502558000061.tif4170 and Based on the difference with TIFF2026502558000062.tif4170 The result is TIFF2026502558000063.tif5170, where TIFF2026502558000064.tif4170 represents the output of the lth processing module when the first local energy function converges.

[0107] [Formula 7] The process of updating the parameters of the l-th processing module based on the second local energy function can be implemented according to Equation 8.

[0108] [Formula 8] TIFF2026502558000066.tif15170 where, TIFF2026502558000067.tif5170 is the parameter of the l-th processing module. As can be seen from this, in Equation 8, the parameter of the l-th processing module The gradient of the second local energy function Elocal for TIFF2026502558000068.tif5170 is calculated, and then the gradient is used to update the parameters of the l-th processing module.

[0109] Furthermore, when the parameters of the lth processing module are determined, the parameters of the lth reference module may be updated based on the parameters of the lth processing module.

[0110] For example, referring to Equation 9, the exponential moving average of the parameters of the processing module in at least one previous iteration is used to calculate the parameters of the reference module. You may update TIFF2026502558000069.tif4170.

[0111] [Formula 9] TIFF2026502558000070.tif5170 where, TIFF2026502558000071.tif4170. As you can see, the parameters of the reference module are TIFF2026502558000072.tif4170 relates to the historical results of the processing module parameters.

[0112] After completing the training of the information processing neural network model based on the above process, the information processing neural network model can be used to extract audio features or predict (or repair) missing (or unclear) parts of audio.

[0113] Optionally, according to the examples of Figures 3A, 4B-4C, and 5, the present disclosure further discloses a method for training a neural network model for audio processing. Optionally, the neural network model for audio processing is an example of the above information processing neural network model. The method includes the steps of: initializing a neural network model for audio processing and a reference neural network model corresponding to the neural network model for audio processing, wherein the neural network model for audio processing includes a plurality of processing modules, and the reference neural network model includes a plurality of reference modules; and updating parameters of the neural network model for audio processing and the reference neural network model through a plurality of iterations, wherein, in each iteration, determining audio features corresponding to each of the plurality of reference modules based on audio training samples and parameters of the plurality of reference modules, determining audio features corresponding to each of the plurality of processing modules based on mask samples obtained by masking the audio training samples and parameters of the plurality of processing modules, updating parameters of the plurality of processing modules based on the audio features corresponding to each of the plurality of processing modules and the audio features corresponding to each of the plurality of reference modules, and updating parameters of the plurality of reference modules based on the updated parameters of the plurality of processing modules. Then, optionally, the trained neural network model for audio processing may be used to extract features of the audio data and further complete speech recognition tasks or sentiment analysis tasks, etc. The neural network model for audio processing may also be used to repair the audio data.

[0114] 3A to 5, the method for training a neural network model disclosed herein can independently train each processing module in an information-processing neural network model. This process does not require clarifying the gradient of the loss function with respect to the parameters of other processing modules, as in the conventional backpropagation algorithm, thereby reducing the difficulty of training the neural network model and enabling parallel training of each processing module, thereby effectively improving the training efficiency of the neural network model.

[0115] Since each processing module in the information processing neural network model can be trained independently and there is no need to share the network parameters between the processing modules, the disclosed method for training a neural network model has good confidentiality, can better protect the parameters of each processing module, and can prevent the parameters from being stolen. Therefore, the disclosed method for training a neural network model has obvious advantages in application scenarios with high confidentiality requirements (e.g., those related to business secrets, privacy, etc.).

[0116] The disclosed method for training a neural network model can effectively reduce the amount of computer calculations and save computer resources, and can utilize computer resources to realize parallel training of multiple processing modules, thereby reducing the training time of the neural network model. The disclosed method for training a neural network model can train each processing module separately, and has high training efficiency, so it is more suitable for training scenarios of large-scale neural network models or real-time training scenarios of neural network models (i.e., the task processing process and the training process are not completely independent, but training is performed while the task is being processed).

[0117] In the disclosed method for training a neural network model, each processing module can be trained by a different server, and there is no need to transmit network parameters between each server during the training process, which effectively reduces the amount of data to be transmitted between each server, reduces the requirements for server performance, and reduces the requirements for communication bandwidth between servers.

[0118] In fact, in the present disclosure, the method of independently training each processing module in an information processing neural network model is closer to the information processing process of the human brain (i.e., each neuron processes information independently, and the information processing process does not involve multiple neurons sequentially backpropagating information, but only involves the input or output of adjacent neurons), and is therefore more capable of producing intelligent machines that can respond in a manner similar to human intelligence.

[0119] FIG. 6 is an exemplary flowchart illustrating a method 600 for performing information processing based on a neural network model according to an embodiment of the present disclosure.

[0120] In step S610, the processing target information is acquired.

[0121] The information to be processed may include one or more of image information, text information, audio information, and video information, i.e., method 600 can be used to process one type of data (e.g., machine translation processes only text information) or multiple types of data simultaneously (e.g., pathology analysis based on both medical images and medical reports).

[0122] In step S620, the processing target information is processed based on the information processing neural network model, thereby obtaining the processing result of the processing target information.

[0123] The information processing neural network model may be trained by the method 300 shown in Figures 3A-3B. The information processing neural network model may be trained only based on training samples and mask samples to perform information processing tasks such as feature extraction and missing information prediction, or may be further trained based on the training samples of downstream information processing tasks for downstream information processing tasks. The information processing neural network model further trained for downstream information processing tasks can be used to complete downstream information processing tasks such as information extraction, information classification, information restoration, information style transition, information enhancement, information mining, and information recognition.

[0124] After obtaining the processing result of the information to be processed based on step S620, the processing result may be displayed on the terminal 120 shown in Fig. 1. Optionally, if necessary, the server 110 or the terminal 120 shown in Fig. 1 may perform further analysis or display processing on the processing result.

[0125] FIG. 7 is a schematic diagram illustrating an information processing process based on a neural network model according to an embodiment of the present disclosure.

[0126] As shown in FIG. 7, the information processing process based on the neural network model according to the embodiment of the present disclosure includes a neural network model training stage and a neural network model testing stage (or a neural network model using stage).

[0127] In the training stage of the neural network model, an information-processing neural network model may be constructed first. Then, based on the structure and parameters of the information-processing neural network model, the structure and parameters of a corresponding reference neural network model may be determined. For example, assuming that the information-processing neural network model includes L processing modules, the reference neural network model corresponding to the information-processing neural network model may include L reference modules, and the l-th processing module among the L processing modules and the l-th reference module among the L reference modules may have the same structure. Here, L and l are positive integers, and l≦L. Optionally, the initial parameters of the information-processing neural network model and the initial parameters of the corresponding reference neural network model may be initialized based on random numbers.

[0128] In the training process, training samples may be obtained, and the information processing neural network model and the reference neural network model may be trained based on the training samples, where the training samples may include image training samples, text training samples, audio training samples, video training samples, etc., or a combination thereof. The training process includes multiple iterations on the parameters of the information processing neural network model and the reference neural network model.

[0129] Specifically, in each iteration, an output result corresponding to each of the plurality of reference modules is determined based on a training sample and the parameters of the plurality of reference modules, an output result corresponding to each of the plurality of processing modules is determined based on a mask sample obtained by performing a mask process on the training sample and the parameters of the plurality of processing modules, and then the parameters of the plurality of processing modules are updated based on the output result corresponding to each of the plurality of processing modules and the output result corresponding to each of the plurality of reference modules, and the parameters of the plurality of reference modules are updated based on the updated parameters of the plurality of processing modules (for example, in each iteration, the parameters of the plurality of reference modules may be updated using the parameter values ​​of the plurality of processing modules in at least one previous iteration).

[0130] When each iteration of training is completed, it is determined whether the current training process satisfies a predetermined condition, and if so, it is determined that the training of the information-processing neural network model is completed, and the information-processing neural network model may be saved. If the current training process does not satisfy the predetermined condition, the training samples are continued to be used to train the information-processing neural network model and its corresponding reference neural network model, and the parameters of the information-processing neural network model and its corresponding reference neural network model are updated.

[0131] It should be understood that the predetermined condition may take a variety of forms. For example, the predetermined condition may be that a predetermined number of training runs is reached (e.g., 1,000 runs, but is not limited thereto. Note that each of the predetermined number of training runs here refers to one epoch of training, not one batch of training). The predetermined condition may be that an energy function converges. Here, the energy function is used to evaluate the accuracy of the output results of the multiple processing modules. Alternatively, the predetermined condition may be that two conditions, that a predetermined number of training runs is reached and that an energy function converges, are simultaneously satisfied. Alternatively, the predetermined condition may be that either one of the two conditions, that a predetermined number of training runs is reached and that an energy function converges, is satisfied. Optionally, the energy function may be a global energy function for evaluating the difference between the output result of the Lth processing module and the output result of the Lth reference module, or a local energy function for evaluating the difference between the output result of the lth processing module and the output result of a reference module corresponding to the lth processing module, or may be jointly determined based on the global energy function and the local energy function.

[0132] By setting the number of training times in advance, the training of the information processing neural network model can be stopped when the preset number of training times is reached, thereby preventing the occurrence of overfitting of the network or the phenomenon where the optimization effect of the network is not significant due to the training time being too long.

[0133] In the testing stage of the information processing neural network model, information to be processed may be provided to a computer. The computer loads the trained information processing neural network model, processes the information to be processed by the trained information processing neural network model, and outputs the results for user reference. Optionally, the user can determine whether further training or optimization of the information processing neural network model is necessary based on the information processing results.

[0134] Similar to training samples, the information to be processed may include information such as images, text, audio, video, or a combination thereof. Information processing neural network models can be used for feature extraction, missing information prediction, information enrichment, information mining, etc.

[0135] Similarly, the actual information processing process (i.e., the process of using the information processing neural network model) is similar to the testing stage of the information processing neural network model, and therefore further explanation will be omitted.

[0136] Note that the training stage of the information-processing neural network model shown in FIG. 7 and the testing stage (or use stage) of the information-processing neural network model may be performed independently. That is, the information-processing neural network model is trained first, and then the information-processing neural network model is tested or used after the training of the information-processing neural network model is completed. Also, the training stage of the information-processing neural network model shown in FIG. 7 and the testing stage (or use stage) of the information-processing neural network model may be performed alternately. For example, the task processing process and the training process are not completely independent, and training may be performed while the task processing is being performed. Therefore, the method of training a neural network model disclosed herein can be better applied to real-time training scenarios of neural network models.

[0137] Tests and verifications have shown that the disclosed method for training an information processing neural network model can train each processing module in the information processing neural network model independently, or can train each processing module in parallel, effectively improving the training efficiency of the neural network model and reducing the training time of the neural network model.Compared to the conventional neural network model training method based on backpropagation, the disclosed method for training a neural network model can effectively avoid the collapse problem in the neural network training process.

[0138] For example, as pointed out in related experiments, the data2vec2 model, which performs self-supervised training based on backpropagation, often suffers from a collapse problem during the training process (e.g., the feature vectors obtained by processing all training samples through a neural network are similar and cannot reflect the characteristics of different training samples). Compared to the data2vec2 model, the method disclosed herein has no collapse problem after repeated testing and validation. That is, the neural network model training method disclosed herein has good stability.

[0139] In addition, for example, the LS-100from960h dataset may be used to initially train an information processing neural network model, and then the LS-10from100h dataset may be used to retrain the information processing neural network model, thereby testing the performance of the neural network model training method disclosed herein and the method of self-supervised training based on backpropagation. Testing and verification have shown that the neural network model training method disclosed herein has significantly improved model training speed and model information processing accuracy compared to the method of self-supervised training based on backpropagation.

[0140] The neural network training method disclosed herein can improve the accuracy and efficiency of sample feature extraction, and when downstream information processing tasks are performed based on the model trained by this method, the accuracy and speed of task processing are also obviously improved.

[0141] FIG. 8 is a block diagram 800 illustrating an apparatus 800 for training a neural network model according to an embodiment of the present disclosure.

[0142] According to an embodiment of the present disclosure, the apparatus 800 for training a neural network model may include an initialization module 810 and a parameter update module 820.

[0143] Here, the initialization module 810 may be configured to initialize an information-processing neural network model and a reference neural network model corresponding to the information-processing neural network model, where the information-processing neural network model includes a plurality of processing modules, and the reference neural network model includes a plurality of reference modules.

[0144] According to an embodiment of the present disclosure, the plurality of processing modules may be L processing modules, and the plurality of reference modules may be L reference modules, and the l-th processing module among the L processing modules and the l-th reference module among the L reference modules have the same structure, where L and l are positive integers, and l≦L.

[0145] The parameter updating module 820 may be configured to update the parameters of the information processing neural network model and the reference neural network model through multiple iterations.

[0146] According to an embodiment of the present disclosure, in each iteration, the parameter update module 820 may determine an output result corresponding to each of the plurality of reference modules based on a training sample and parameters of the plurality of reference modules, determine an output result corresponding to each of the plurality of processing modules based on a mask sample obtained by performing mask processing on the training sample and parameters of the plurality of processing modules, update the parameters of the plurality of processing modules based on the output result corresponding to each of the plurality of processing modules and the output result corresponding to each of the plurality of reference modules, and then update the parameters of the plurality of reference modules based on the updated parameters of the plurality of processing modules.

[0147] According to an embodiment of the present disclosure, for an l-th processing module among the L processing modules, parameters of the l-th processing module may be updated based on an output result of the l-th processing module and an output result of a reference module corresponding to the l-th processing module, where the reference modules corresponding to the l-th processing module include the l-th reference module to the l+n-th reference module, where m and n are integers and satisfy m≧0, n≧0, lm≧1, and l+n≦L.

[0148] According to an embodiment of the present disclosure, in the parameter update process of the plurality of processing modules, an energy function value may be calculated based on the output results corresponding to each of the plurality of processing modules and the output results corresponding to each of the plurality of reference modules, and the parameters of the plurality of processing modules may be updated based on the energy function value, where the energy function is for evaluating accuracy of the output results of the plurality of processing modules.

[0149] The apparatus 800 for training a neural network model can be used to perform supervised training on the information-processing neural network model, and can also be used to perform self-supervised training on the information-processing neural network model.

[0150] It should be understood that the apparatus 800 for training a neural network model shown in FIG. 8 can implement various methods for training a neural network model described with reference to FIGS. 3A-3B.

[0151] FIG. 9 is a schematic diagram illustrating a configuration of an apparatus 900 for performing information processing based on a neural network model according to an embodiment of the present disclosure.

[0152] According to an embodiment of the present disclosure, an apparatus 900 for performing information processing based on a neural network model may include an information acquiring module 910 and an information processing module 920 .

[0153] Here, the information acquiring module 910 may be configured to acquire information to be processed, where the information to be processed includes at least one of image information, text information, audio information, and video information.

[0154] The neural network model 920 for generating interaction information may be configured to process the processing target information based on the information processing neural network model, thereby obtaining a result of processing the processing target information.

[0155] It should be understood that the device 900 for performing information processing based on a neural network model shown in Fig. 9 can implement various methods for performing information processing based on a neural network model described with reference to Fig. 6. The information processing neural network model may be trained by the method 300 shown in Figs. 3A-3B.

[0156] The device 900 for performing information processing based on a neural network model may be located in the server 110 shown in FIG. 1 or in the terminal 120 shown in FIG.

[0157] In general, various exemplary embodiments of the present disclosure may be implemented in hardware or special purpose circuits, software, firmware, logic, or any combination thereof. Some aspects may be implemented in hardware, while other aspects may be implemented in firmware or software executable by a controller, microprocessor, or other computing device. When aspects of embodiments of the present disclosure are shown or described as block diagrams, flowcharts, or some other graphical representation, it is understood that the blocks, apparatus, systems, techniques, or methods described herein may be implemented in, by way of non-limiting example, hardware, software, firmware, special purpose circuits or logic, general purpose hardware or controller or other computing device, or some combination thereof.

[0158] For example, a method or apparatus according to an embodiment of the present disclosure may be realized by the architecture of a computing device 3000 shown in FIG. 10. As shown in FIG. 10, the computing device 3000 may include a bus 3010, one or more CPUs 3020, a read-only memory (ROM) 3030, a random access memory (RAM) 3040, a communication port 3050 connected to a network, an input / output component 3060, a hard disk 3070, etc. A storage device in the computing device 3000, such as the ROM 3030 or the hard disk 3070, may store various data or files used in processing and / or communication of the methods provided in the present disclosure, as well as program instructions executed by the CPU. The computing device 3000 may also include a user interface 3080. Of course, the architecture shown in FIG. 10 is merely exemplary. When realizing a different device, one or more components in the computing device shown in FIG. 10 may be omitted according to actual needs.

[0159] According to another aspect of the present disclosure, there is also provided a computer-readable storage medium. The computer storage medium has computer-readable instructions stored thereon. When executed by a processor, the computer-readable instructions cause the implementation of the method according to the embodiment of the present disclosure, as described above with reference to the figures. The computer-readable storage medium in the embodiment of the present disclosure may be volatile or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory may be random access memory (RAM) used as an external cache. By way of example and not limitation, RAM is available in many forms, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), SyncLink dynamic random access memory (SLDRAM), and direct Rambus random access memory (DR RAM). It should be noted that the memory of the methods described herein is intended to comprise, but is not limited to, these and any other suitable types of memory. It should be noted that the memory of the methods described herein is intended to comprise, but is not limited to, these and any other suitable types of memory.

[0160] In accordance with an embodiment of the present disclosure, there is further provided a computer program product or computer program including computer instructions stored in a computer-readable storage medium, the computer instructions being read by a processor of a computing device from the computer-readable storage medium, and the processor executing the computer instructions causing the computing device to perform a method according to an embodiment of the present disclosure.

[0161] In summary, embodiments of the present disclosure provide a method, apparatus, computer program product, and storage medium for training a neural network model, the method including: initializing an information-processing neural network model and a reference neural network model corresponding to the information-processing neural network model, wherein the information-processing neural network model includes a plurality of processing modules and the reference neural network model includes a plurality of reference modules; and updating parameters of the information-processing neural network model and the reference neural network model through a plurality of iterations, wherein in each iteration, determining an output result corresponding to each of the plurality of reference modules based on a training sample and parameters of the plurality of reference modules, determining an output result corresponding to each of the plurality of processing modules based on mask samples obtained by masking the training sample and parameters of the plurality of processing modules, updating the parameters of the plurality of processing modules based on the output result corresponding to each of the plurality of processing modules and the output result corresponding to each of the plurality of reference modules, and updating the parameters of the plurality of reference modules based on the updated parameters of the plurality of processing modules.

[0162] The disclosed method for training a neural network model can train each processing module in an information processing neural network model independently, and can train each processing module in parallel, thereby effectively improving the training efficiency of the neural network model. The disclosed method for training a neural network model can train a processing module based only on the output result of the processing module and the output result of the reference module corresponding to the processing module. This process does not require specifying the gradient of the loss function with respect to the parameters of other processing modules, thereby reducing the difficulty of training the neural network model, effectively saving computer resources, and reducing the requirements for computer performance.

[0163] In addition, since each processing module in the information processing neural network model can be trained independently and there is no need for each processing module to share its network parameters, the disclosed method for training a neural network model has good confidentiality, can better protect the parameters of each processing module, and can prevent the parameters from being stolen. Therefore, the disclosed method for training a neural network model has obvious advantages in application scenarios with high confidentiality requirements.

[0164] The disclosed method for training a neural network model can effectively reduce the amount of computation required by a computer and utilize computer resources to achieve parallel training of multiple processing modules, thereby reducing the training time of the neural network model. The disclosed method for training a neural network model can train each processing module separately, resulting in high training efficiency, making it more suitable for training scenarios of large-scale neural network models or real-time training scenarios of neural network models.

[0165] In the disclosed method for training a neural network model, each processing module can be trained by a different server, and there is no need to transmit network parameters between each server during the training process, which effectively reduces the amount of data to be transmitted between each server, reduces the requirements for server performance, and reduces the requirements for communication bandwidth between servers.

[0166] Furthermore, compared to the conventional neural network model training method based on backpropagation, the neural network model training method disclosed herein can stably train the neural network model and effectively avoid the problem of collapse during the neural network training process.

[0167] In fact, in the present disclosure, the method of independently training each processing module in an information processing neural network model is closer to the information processing process of the human brain (i.e., each neuron processes information independently, and the information processing process does not involve multiple neurons sequentially backpropagating information, but only involves the input or output of adjacent neurons), and is therefore more capable of producing intelligent machines that can respond in a manner similar to human intelligence.

[0168] Tests and verifications have shown that the neural network model training method disclosed herein significantly improves the model training speed and the model information processing accuracy compared to the self-supervised training method based on backpropagation.

[0169] The flowcharts and block diagrams in the drawings illustrate possible system architectures, functions, and operations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flowcharts or block diagrams may represent a module, program segment, or portion of code, which includes at least one executable instruction for implementing a given logical function. It should also be noted that in some alternative implementations, the functions noted in the blocks may occur in a different order than that depicted in the drawings. For example, two blocks shown in succession may actually be executed essentially in parallel, or may be executed in the reverse order, depending on the functionality involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented in a dedicated hardware-based system for performing a given function or operation, or in a combination of dedicated hardware and computer instructions.

[0170] This disclosure uses specific terms to describe embodiments of the disclosure. For example, "first / second embodiment," "one embodiment," and / or "some embodiments" refer to certain features, configurations, or characteristics associated with at least one embodiment of the disclosure. Therefore, it should be emphasized and noted that "one embodiment," "one embodiment," or "one alternative embodiment" referenced more than once in different places in this specification do not necessarily refer to the same embodiment. Furthermore, some features, configurations, or characteristics of one or more embodiments of the disclosure may be combined as appropriate.

[0171] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. It should be further understood that terms as defined in ordinary dictionaries, unless expressly defined herein, should not be construed in an idealized or overly formal sense, but should be construed to have a meaning consistent with their meaning in the context of the relevant art.

[0172] The foregoing has described the present invention but should not be construed as limiting it. While several exemplary embodiments of the present invention have been described, those skilled in the art will readily appreciate that many modifications can be made to the exemplary embodiments without departing from the novel teachings and advantages of the present invention. Accordingly, all such modifications are intended to be included within the scope of the present invention as defined by the appended claims. It should be understood that the foregoing has described the present invention but should not be construed as limited to the particular embodiments disclosed, and that modifications of the disclosed embodiments and other embodiments are intended to be included within the scope of the appended claims. The present invention is limited by the claims and their equivalents.

Claims

1. 1. A method for training a neural network model, comprising: a step of initializing an information-processing neural network model and a reference neural network model corresponding to the information-processing neural network model, the information-processing neural network model including a plurality of processing modules, and the reference neural network model including a plurality of reference modules; updating parameters of the information processing neural network model and the reference neural network model through multiple iterations, wherein in each iteration: Based on a training sample and parameters of the plurality of reference modules, the training sample is input to the plurality of reference modules to perform forward calculation, and the calculation results are output, thereby obtaining output results corresponding to each of the plurality of reference modules; Based on a mask sample obtained by performing a mask process on the training sample and parameters of the plurality of processing modules, the mask sample is input to the plurality of processing modules to perform forward calculation, and the calculation results are output, thereby obtaining output results corresponding to each of the plurality of processing modules; updating parameters of the plurality of processing modules based on output results corresponding to each of the plurality of processing modules and output results corresponding to each of the plurality of reference modules; updating parameters of the plurality of reference modules based on the updated parameters of the plurality of processing modules; How to train a neural network model.

2. The step of updating the parameters of the plurality of reference modules based on the updated parameters of the plurality of processing modules includes: updating parameters of the plurality of reference modules using values ​​of parameters of the plurality of processing modules in at least one previous iteration; 10. The method of training a neural network model of claim 1.

3. The step of updating the parameters of the plurality of reference modules using values ​​of the parameters of the plurality of processing modules in at least one previous iteration includes: updating parameters of the plurality of reference modules using exponential moving averages of parameters of the plurality of processing modules in at least one previous iteration; The method of training a neural network model according to claim 2.

4. The plurality of processing modules are L processing modules, the plurality of reference modules are L reference modules, an l-th processing module among the L processing modules and an l-th reference module among the L reference modules have the same structure, L and l are positive integers, and l≦L; A method for training a neural network model according to any one of claims 1 to 3.

5. The step of updating parameters of the plurality of processing modules based on output results corresponding to each of the plurality of processing modules and output results corresponding to each of the plurality of reference modules includes: For the l-th processing module among the L processing modules, updating a parameter of the lth processing module based on an output result of the lth processing module and an output result of a reference module corresponding to the lth processing module, wherein the reference module corresponding to the lth processing module includes a l-mth reference module to a l+nth reference module, m and n are integers, and satisfy m≧0, n≧0, l-m≧1, l+n≦L; 5. The method of training a neural network model according to claim 4.

6. The step of updating parameters of the plurality of processing modules based on output results corresponding to each of the plurality of processing modules and output results corresponding to each of the plurality of reference modules includes: a step of calculating a value of an energy function based on an output result corresponding to each of the plurality of processing modules and an output result corresponding to each of the plurality of reference modules, and updating parameters of the plurality of processing modules based on the value of the energy function, wherein the energy function is for evaluating accuracy of the output results of the plurality of processing modules; A method for training a neural network model according to claim 4 or 5.

7. the energy function includes a global energy function and a local energy function; the global energy function is for evaluating the difference between an output result of an L-th processing module and an output result of an L-th reference module, and the local energy function is for evaluating the difference between an output result of an l-th processing module and an output result of a reference module corresponding to the l-th processing module; 7. The method of training a neural network model according to claim 6.

8. the local energy functions include a first local energy function and a second local energy function; The step of calculating a value of an energy function based on output results corresponding to each of the plurality of processing modules and output results corresponding to each of the plurality of reference modules includes: For the l-th processing module among the L processing modules, determining a value of the first local energy function based on a difference between an output result of the lth processing module and an output result of a reference module corresponding to the lth processing module; determining a predicted output result of the lth processing module based on the global energy function; determining a value of the second local energy function based on a difference between an output result of the lth processing module and an expected output result of the lth processing module; 8. The method of training a neural network model according to claim 7.

9. determining a predicted output result of the lth processing module based on the global energy function, determining a backward calculation difference of the output result of the lth processing module based on the value of the global energy function; determining a predicted output result of the lth processing module based on a backward calculation difference of the output result of the lth processing module and the output result of the lth processing module; 9. The method of training a neural network model according to claim 8.

10. determining a backward calculation difference of an output result of the lth processing module based on the value of the global energy function, For the l+k-th processing module among the L-th processing module to the l-th processing module, calculating a partial derivative of an output result of the l+k-th processing module with respect to an output result of the l+k-1-th processing module, the output result of the l+k-1-th processing module being an input to the l+k-th processing module; Calculating a backward calculation difference of the output result of the l+k-th processing module based on the partial derivative and the backward calculation difference of the output result of the l+k+1-th processing module, where k is an integer, and satisfies k≧0 and l+k+1≦L.

10. The method of training a neural network model according to claim 9.

11. The step of updating parameters of the plurality of processing modules includes: performing a weighting calculation on the first local energy function and the second local energy function corresponding to the lth processing module to obtain a third local energy function; updating a parameter of the lth processing module based on the third local energy function; A method for training a neural network model according to any one of claims 8 to 10.

12. The step of updating parameters of the plurality of processing modules includes: if the first local energy function corresponding to the lth processing module does not converge, updating a parameter of the lth processing module based on a value of the first local energy function corresponding to the lth processing module; If the first local energy function corresponding to the lth processing module converges, updating a parameter of the lth processing module based on a second local energy function corresponding to the lth processing module. A method for training a neural network model according to any one of claims 8 to 11.

13. Each processing module among the plurality of processing modules or each reference module among the plurality of reference modules is a sub-neural network model or a neural network layer; The processing module or the reference module may include one or more of a convolutional neural network, an attention-based neural network, a recurrent neural network, a recurrent neural network, a feed-forward neural network, a generative adversarial neural network, and a deep neural network; A method for training a neural network model according to any one of claims 1 to 12.

14. Used to perform supervised training on the information processing neural network model or to perform self-supervised training on the information processing neural network model. A method for training a neural network model according to any one of claims 1 to 13.

15. The step of masking the training samples includes the step of masking the training samples with at least one of different sizes, different positions, and different shapes; the training samples include at least one of image training samples, text training samples, audio training samples, and video training samples; A method for training a neural network model according to any one of claims 1 to 14.

16. If the updated information processing neural network model satisfies a predetermined condition, completing the training of the information processing neural network model to obtain a trained information processing neural network model; Further training the trained information processing neural network model using information processing training samples corresponding to information processing tasks, wherein the information processing tasks include at least one of information extraction, information classification, information restoration, information style transition, information enhancement, information mining, and information recognition. A method for training a neural network model according to any one of claims 1 to 15.

17. 1. A method for performing information processing based on a neural network model, comprising: acquiring information to be processed, the information to be processed including at least one of image information, text information, audio information, and video information; and a step of acquiring a result of processing the processing target information based on an information processing neural network model, the information processing neural network model being trained by the method for training a neural network model according to any one of claims 1 to 16. A method for performing information processing based on a neural network model.

18. 1. An apparatus for training a neural network model, comprising: an initialization module configured to initialize an information-processing neural network model and a reference neural network model corresponding to the information-processing neural network model, the information-processing neural network model including a plurality of processing modules, and the reference neural network model including a plurality of reference modules; a parameter updating module configured to update parameters of the information processing neural network model and the reference neural network model through a plurality of iterations, wherein in each iteration: Based on a training sample and parameters of the plurality of reference modules, the training sample is input to the plurality of reference modules to perform forward calculation, and the calculation results are output, thereby obtaining output results corresponding to each of the plurality of reference modules; Based on a mask sample obtained by performing a mask process on the training sample and parameters of the plurality of processing modules, the mask sample is input to the plurality of processing modules to perform forward calculation, and the calculation results are output, thereby obtaining output results corresponding to each of the plurality of processing modules; updating parameters of the plurality of processing modules based on output results corresponding to each of the plurality of processing modules and output results corresponding to each of the plurality of reference modules; a parameter update module that updates parameters of the plurality of reference modules based on updated parameters of the plurality of processing modules; A device for training neural network models.

19. A computer program product including computer software code which, when executed by a processor, causes the computer program product to implement the method of any one of claims 1 to 17.

20. 18. A computer-readable storage medium having stored thereon computer-executable instructions, the instructions, when executed by a processor, causing the method of any one of claims 1 to 17 to be implemented.

21. 18. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured by the computer program to execute a method according to any one of claims 1 to 17.

Images