A system and method capable of autonomously acquiring and learning knowledge

By combining structured and unstructured knowledge expression and calculation, an artificial intelligence system with knowledge learning and creativity is built, which solves the problem that deep neural network models are difficult to quickly learn new knowledge and incremental learning, and realizes self-consistent knowledge generation and incremental learning improves the adaptability and efficiency of the model.

CN118396052BActive Publication Date: 2024-11-22ZHIZHISHENGONG (SHANGHAI) INFORMATION TECH CO LTD
View PDF 7 Cites 0 Cited by

Patent Information

Application Number
CN202410475682.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2023-04-21
Filing Date
2024-04-19
Publication Date
2024-11-22
Estimated Expiration
2044-04-19

AI Technical Summary

Technical Problem

The existing deep neural network models rely on big data and large models, making it difficult to quickly learn new knowledge, and cannot effectively conduct incremental learning and knowledge exchange. The models can only be applied based on existing knowledge and cannot create new knowledge.

Method used

By combining structured and unstructured knowledge expression and calculation, an artificial intelligence system with knowledge learning, creation and reasoning capabilities is built, using generation modules and judgment modules. The generation module includes concept generators and conjecture deducers, and the judgment module includes simulators and knowledge base to achieve the update and creation of prior knowledge.

Benefits of technology

It realizes self-consistent knowledge generation, incremental learning and knowledge exchange, can effectively create new knowledge, improve the learning efficiency and adaptability of the model, and avoid exponential growth of model scale and data volume.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118396052B_ABST
    Figure CN118396052B_ABST
Patent Text Reader

Abstract

The present invention provides a system and method for autonomous knowledge acquisition and learning. The system is mainly used for concept generation, knowledge extraction and learning of input data (such as video images, codes, texts, etc.), including a generation module and a determination module. The generation module includes: a concept generator, which extracts data features of input image data X and generates vector feature values ​​based on primitive concepts; a conjecture derivation device, which generates concept evolution results Y based on user instruction reasoning; the determination module includes: a simulator, which refutes and verifies the conjecture knowledge Y=F(X) based on the existing empirical data in the database; a knowledge base, which stores the prior knowledge of the application field based on a graph model and predefines the representation of the primitive concepts; a knowledge translator, which pre-trains the vector mapping of all primitive concepts stored in the knowledge base, extracts the vector feature values ​​of the primitive concepts corresponding to the input image data X, and sends the vector feature values ​​and knowledge prompts of the prior knowledge to the conjecture derivation device. After adopting the above technical solution, it is possible to perform training and learning according to data, and update and create the existing knowledge system.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to a system and method capable of autonomously acquiring and learning knowledge. Background Art

[0002] The past decade has been a period of rapid development of artificial intelligence (AI) technology represented by deep learning and reinforcement learning. In particular, the large language model (LLM) represented by ChatGPT has further pushed deep neural network technology to the top, making all walks of life in the whole society face the transformation of new AI technology. However, for the current deep neural network technology that strongly relies on big data and large models, its essence is inductive learning based on statistical theory, so it also faces insurmountable method defects, which are mainly reflected in three aspects:

[0003] 1) Due to the reliance on statistical learning of big data, the dominant features of the data set (high-probability information) will become the focus of model learning, while the low-probability important features highlighted by the long-tail problem will be covered up, resulting in the model being unable to quickly learn new knowledge, that is, the model follows the crowd and does not innovate. To learn new knowledge, it is necessary to expand the data of new knowledge in large quantities, and the expanded data must also match the existing data volume to a certain extent to achieve data balance, which will lead to an increasing amount of data in the future and higher costs;

[0004] 2) Since the model is learning from historical experience (data of facts that have already occurred), it can only learn the correlation between the existing data features. Although it can be expressed through correlation reorganization output, the result of the reorganization essentially already exists, but it does not exist in the data set as an independent individual. The model cannot generate new knowledge by breaking away from this existing knowledge.

[0005] 3) The training of existing neural network models is essentially to classify data by adjusting network parameters. Supervised learning is to adjust network parameters according to data labels for classification, while unsupervised learning or self-supervised learning is to classify according to the correlation measurement of data features, which leads to the fact that the trained network can only be applied to the classified categories. If new categories are to be added, the entire network parameters will inevitably be adjusted comprehensively. As a result, on a network structure of the same size, the more categories the network can distinguish, the worse the accuracy of the distinction will be. In order to ensure the classification accuracy, the network scale must be further expanded and the full amount of data must be used for new training. A typical example is the development of the GPT series. From GPT1's 117 million parameters and 5GB of training data to GPT3's 175 billion parameters and 45TB of data, the entire model scale and data volume have grown exponentially. As a result, the model cannot effectively conduct incremental learning and knowledge exchange, but must constantly renovate itself and learn from scratch. (Incremental learning means that the model can maintain the classification accuracy of existing categories and improve the classification accuracy of newly added categories only by learning new category data. Knowledge exchange means that the model can improve its own accuracy in distinguishing new categories by learning from other models that have learned to distinguish new categories).

[0006] Therefore, there is a need for a new type of system and method that can autonomously acquire and learn knowledge, simulate human cognitive patterns to conduct cognitive learning of information, and realize knowledge creation and domain application. Summary of the invention

[0007] In order to overcome the above-mentioned technical defects, the purpose of the present invention is to provide a system and method that can autonomously acquire and learn knowledge. By combining structured and unstructured knowledge expression and calculation, an artificial intelligence system with knowledge learning, creation and reasoning capabilities is constructed, in which structured knowledge is used as the skeleton of the knowledge system to support the understanding and reasoning capabilities of prior knowledge; in addition, based on unstructured knowledge expression and calculation methods, it is possible to train and learn according to data, update and create existing knowledge systems, and realize the expansion of system intelligence.

[0008] The present invention discloses a system capable of autonomously acquiring and learning knowledge, comprising a generation module and a determination module, wherein

[0009] The generated modules include:

[0010] A concept generator receives an input image data X, extracts data features of the input image data X, and generates vector feature values ​​of the data features under the structure of the primitive concept based on a predefined primitive concept;

[0011] The conjecture derivation device inputs the concept representation vector, user instructions and related knowledge prompts, and infers the evolution result Y;

[0012] The determination module includes:

[0013] The simulator receives the conjectured knowledge Y = F(X) and refutes and verifies it based on the empirical data in the database;

[0014] The knowledge base stores prior knowledge of the application domain based on the graph model and predefines primitive concepts;

[0015] The knowledge translator pre-trains the vector mapping of all primitive concepts stored in the knowledge base. The knowledge translator is connected to the concept generator and the conjecture deriver, receives the input data X and extracts the representation vector under its corresponding primitive concept structure, and sends the knowledge prompts related to the concept corresponding to the representation vector to the conjecture deriver.

[0016] Preferably, the concepts of the vector feature values ​​generated by the concept generator of the generation module include primitive concepts and object concepts. The object concept is a vector representation of a high-dimensional space composed of elements of the primitive concept. The primitive concept represents the basic characteristics of the application field, and all primitive concepts and object concepts correspond to a vector space under their characteristic dimension.

[0017] Preferably, the knowledge base represents the knowledge concepts of prior knowledge and the conversion relationships between them in the form of a graph;

[0018] The attribute form of the conversion relationship can be expressed as F, so that an initial concept is converted into a target concept, where the initial concept is a primitive concept or an object concept, and the target concept is an object concept;

[0019] When the initial concept is a primitive concept, F represents the definition of the target concept represented by the combination of primitive concepts; if the initial concept is an object concept, F represents the inference relationship between the target concept and the object concept;

[0020] The derivation chain of prior knowledge related to the input image data X is extracted from the knowledge base, and the derivation results are verified one by one. The conjecture concept vector Y obtained from the previous derivation is used as the input for the next knowledge derivation to realize the derivation and verification of knowledge in stages.

[0021] Preferably, the knowledge translator implements the mapping between the concept symbols of the prior knowledge in the determination module and the conjecture concept vector Y in the generation module.

[0022] Preferably, the conjecture deriver translates the prior knowledge Y=F(X) in the knowledge base into a corresponding concept representation vector through a knowledge translator, wherein the vector translated from the initial concept X in F serves as the input of the conjecture deriver, and the vector translated from the target concept Y in F serves as the labeled output of the conjecture deriver, and the network parameters of the conjecture deriver are pre-trained based on supervised learning.

[0023] Preferably, the knowledge translator decides whether to pass the prior knowledge as a prompt to the conjecture derivation device according to the completeness of the knowledge base and the task goal;

[0024] When the confidence of the knowledge base is lower than a confidence threshold, the prompt is not input to the hypothesis derivation device, or the confidence of the prompt is set to a smaller value.

[0025] Preferably, when the conjecture derivation device generates conjecture knowledge Y=F(X), the simulator will refute and verify the knowledge based on similar empirical data in the database. If there is no counterexample, the knowledge is considered valid and the knowledge expression will be stored in the knowledge base.

[0026] Preferably, it also includes:

[0027] Randomizer;

[0028] When the simulator's refutation verification determines that the conjecture concept vector Y is invalid, the randomizer is activated, which randomly applies random perturbations to the parameters in the conjecture derivation network and activates the conjecture derivation device to reconverge and generate a new conjecture concept vector Y and a new knowledge expression.

[0029] The present invention also discloses a knowledge acquisition method that can be independently created, comprising the following steps:

[0030] A concept generator receives an input image data X, extracts data features of the input image data X, and generates vector feature values ​​of the data features under the structure of the primitive concept based on a predefined primitive concept;

[0031] The conjecture derivation device inputs the concept representation vector, user instructions and related knowledge prompts, and infers the evolution result Y;

[0032] The simulator receives the conjectured knowledge Y = F(X) and refutes and verifies it based on the empirical data in the database;

[0033] The knowledge base stores prior knowledge of the application domain based on the graph model and predefines primitive concepts;

[0034] The knowledge translator pre-trains the vector mapping of all primitive concepts stored in the knowledge base. The knowledge translator is connected to the concept generator and the conjecture deriver, receives the input data X and extracts the representation vector under its corresponding primitive concept structure, and sends the knowledge prompts related to the concept corresponding to the representation vector to the conjecture deriver.

[0035] Compared with the prior art, the above technical solution has the following beneficial effects:

[0036] 1) Theoretical learning: The pre-training inner loop consisting of the knowledge base, knowledge translator, and conjecture derivation device can efficiently pre-train the conjecture derivation device based on people's existing knowledge and experience, so that the conjecture knowledge generated by the conjecture derivation device based on the premise (concept) of prior knowledge is consistent with the prior knowledge, and the knowledge generation is self-consistent. Just like when people learn textbook knowledge, they can have a preliminary understanding of the expression and application of knowledge based on a small number of textbook knowledge cases. The knowledge base is like textbook knowledge, and the generation module is like the human brain;

[0037] 2) Statistical verification: The simulator can summarize the knowledge rules of the data in the knowledge base that are consistent with the concepts in the conjectured knowledge through statistical models to refute whether the rules of the verified and conjectured knowledge are correct, just like when people do experimental verification, they need to make multiple measurements to confirm whether there are errors in the experimental results;

[0038] 3) Create new knowledge (practical learning): The entire system, from input data to actuator output, and then the large closed loop of statistical judgment to verify the conjecture knowledge, can effectively realize the generation of new knowledge. For example, there is characteristic knowledge of horses in the knowledge base. If a horse is observed, the characteristic concept contained in the input data can effectively match the concept of horses in the knowledge base, so the conjecture derivation can generate a conceptual vector output of horses; but if a unicorn is observed, although it has most of the characteristics of a horse, the horns on its head and the wings on its body will be the conjecture derivation. Generate a vector expression far away from the horse's own concept vector (because the horse has no horns and wings, these two dimensions are 0, while the unicorn has a larger value in these two dimensions, which will widen the class distance with the horse), thereby creating a biological concept different from the horse, a new concept category vector representation.

[0039] 4) Genetic evolution: When the randomizer determines that the conjecture knowledge is invalid, it can trigger the mutation of the conjecture derivation network parameters, so that the network can converge towards a new derivation direction. And because the corresponding parameters of the original conjecture knowledge have been cached in the knowledge translator (but not entered into the knowledge base), the evolution of the conjecture derivation in the correct direction can be accelerated through continuous statistical verification, which can essentially achieve the same evolutionary effect as the genetic algorithm.

[0040] 5) Incremental learning: Since the entire system separates concept representation and knowledge deduction into two independent networks, the two networks will not be trained and updated synchronously, thus ensuring that the concept representation will not be changed when training knowledge deduction. In this way, the concepts that have been learned before will be retained without repeated learning, just like there are concepts of cats in different languages, but after learning the English word cat, there is no need to relearn the Chinese word cat. The two are completely equivalent concepts and do not affect the expression of all knowledge related to cats.

[0041] In addition, since all concepts are represented on the domain primitive concept structure, for example, all animals can be expressed based on the features of the head, body, limbs, hands and feet (four claws), etc., when learning cats, feature vectors of cats under different primitive concepts will be formed. These features will form the representation vector of the concept "cat", corresponding to a certain interval in the primitive space. After that, whether learning dogs, pigs, and other animals, it is just to divide different intervals in the primitive space to correspond to these new concepts, without affecting the representation of cats, thus achieving the effect of incremental learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a schematic diagram of the structure of a knowledge acquisition and learning system in a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0043] The advantages of the present invention are further described below in conjunction with the accompanying drawings and specific embodiments.

[0044] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present disclosure. Instead, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0045] The terms used in this disclosure are for the purpose of describing specific embodiments only and are not intended to limit the disclosure. The singular forms of "a", "said" and "the" used in this disclosure and the appended claims are also intended to include plural forms unless the context clearly indicates otherwise. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0046] It should be understood that although the terms first, second, third, etc. may be used in the present disclosure to describe various information, such information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the present disclosure, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "at the time of" or "when" or "in response to determining".

[0047] In the description of the present invention, it is necessary to understand that the terms "longitudinal", "lateral", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0048] In the description of the present invention, unless otherwise specified and limited, it should be noted that the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a mechanical connection or an electrical connection, or it can be the internal connection between two components. It can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms can be understood according to the specific circumstances.

[0049] In the following description, the suffixes such as "module", "component" or "unit" used to represent elements are only used to facilitate the description of the present invention, and have no specific meanings. Therefore, "module" and "component" can be used interchangeably.

[0050] See also Figure 1 , is a schematic diagram of the structure of a knowledge acquisition and learning system in accordance with a preferred embodiment of the present invention. In this embodiment, the knowledge acquisition and learning system includes a generation module and a determination module, wherein the generation module includes: a concept generator, receiving an input data X, extracting data features of the input data X, and generating vector feature values ​​of the data features under the structure of the primitive concept based on a primitive concept; a conjecture derivation device, receiving vector feature values ​​and knowledge prompts, and generating a conjecture concept vector Y; the determination module includes: a simulator, refuting and verifying the conjectured knowledge based on the empirical data in the database, and storing the conjectured knowledge in the knowledge base if the verification passes; a knowledge base, storing the prior knowledge of the application field based on the graph model, and predefining the primitive concepts; a knowledge translator, pre-training the vector mapping of all primitive concepts stored in the knowledge base, the knowledge translator is connected to the concept generator and the conjecture derivation device, receiving the input image data X and extracting the vector feature values ​​of the primitive concept corresponding to the input image data X, and sending the vector feature values ​​and the knowledge prompts of the prior knowledge to the conjecture derivation device.

[0051] More specifically, the architecture modules of the knowledge acquisition and learning system cover methods such as neural networks and graphical models, among which graphical models are used to represent knowledge concept expressions and concept relationship conversions, and neural networks are used to compress, calculate and expand information on concepts in graphical models. By combining graphical models with neural networks, the system ultimately achieves knowledge relationship expression based on prior knowledge, observation information processing based on knowledge, knowledge expansion based on real feedback, and information translation based on concept representation and neural network parameter mapping.

[0052] When each module works, it will be executed in the following ways:

[0053] b1- In the generation module, concepts are defined in the form of hierarchical concepts, which are divided into primitive concepts and object concepts. Object concepts are represented by vectors in a high-dimensional space composed of the elements of primitive concepts, while primitive concepts represent the basic features of the application field, and all concepts correspond to a vector space under their feature dimensions. For example, in mammal recognition, the objects are various mammals, such as cats, dogs, monkeys, pigs, etc., and the primitive concepts are the basic features such as head, eyes, nose, ears, body, hair, legs, claws, and tail. If the feature values ​​for the recognition of each primitive concept in the input data are within the vector space of an object concept A, then the object to be recognized in the figure is considered to be A.

[0054] b2-The knowledge base represents each knowledge concept and the conversion relationship between them in the form of a graph. The attribute form of knowledge conversion can be expressed as F (initial concept set → target concept) or target concept = F (initial concept set), where the initial concept can be a primitive concept or an object concept, and the target concept can only be an object concept. If the initial concept is a primitive concept, F represents the definition of the target concept represented by the combination of primitive concepts; if the initial concept is an object concept, F represents the derivation relationship between the target concept and other object concepts. The knowledge base will extract the knowledge derivation chain related to the input data concept, verify the derivation results one by one, and use the conjecture result Y obtained from the previous derivation as the input for the next knowledge derivation, thereby realizing the staged derivation and verification of knowledge.

[0055] b3-Knowledge Translator realizes the mapping between the knowledge concept symbols in the judgment module and the concept vectors in the generation module, as well as the conjecture derivation network parameters corresponding to each knowledge F.

[0056] The b4-hypothesis deriver is a deep neural network that realizes the derivation of knowledge F, that is, it inputs the primitive concept vector generated by the concept generator and derives the target concept vector generated under the knowledge F.

[0057] b4.1- Pre-training of conjecture derivation: The knowledge F (concept symbol) in the knowledge base is translated into corresponding concept vectors through the knowledge translator. The vector of the initial concept translation in F is used as the input of the conjecture derivation, and the vector of the target concept translation in F is used as the labeled output of the conjecture derivation. The conjecture derivation network parameters are pre-trained based on supervised learning. Through pre-training, the conjecture derivation network learns all the concept-to-concept knowledge derivations in the knowledge base.

[0058] b4.2- The generation of new knowledge includes the generation of new concepts and the generation of new derivational relationships between object concepts. 1) The generation of new concepts is essentially that the target object vector obtained based on the new primitive concept vector set is not in the vector space of any object concept in the knowledge translator, and the spatial distance with each object concept is large, then the new primitive input corresponds to a new concept generation. Since the conjecture derivation has learned the concept derivation in the knowledge base, when a new set of concept vectors is input, the conjecture derivation will output a new target concept vector, and then the knowledge translator can be used to guess whether it is a suspected new object concept (this is conjecture knowledge, not post-determination knowledge). For example, there is knowledge about cats in the knowledge base, but the newly input primitive vector expresses strong limbs, a swollen head, few textures and no long hair on the body, and a pompom at the end of the tail. Then the object concept vector obtained based on these primitive vector inputs may exceed the vector distribution space of the cat after pre-training, so it can be guessed that it is a new object concept, and the symbol X is preset in the knowledge translator as the concept symbol corresponding to this vector. Then, the simulator searches the database and finds data that is close to the original vector. If its overall features are consistent, it means that it is a valid new concept and can be stored in the knowledge base. If the database has a corresponding language description for this data, the concept X can be matched as "lion", thus realizing the generation, recognition and learning of new concepts.

[0059] 2) The generation of new deductive relationships between object concepts is essentially the input of an initial concept vector set X that is different from any knowledge in the knowledge base (i.e., the input object vector exceeds the spatial range of the object vector in the knowledge, or is different from the input object dimension listed in the knowledge). If the target object generated by the conjecture derivation is consistent with the vector space corresponding to a certain object concept Y in the knowledge base, then new conjecture knowledge F(X→Y) is obtained. After the subsequent judgment is valid, the knowledge can be judged as true knowledge and stored in the knowledge base. For example, there is a knowledge base that cats bite the neck when they prey on mice. After training, the conjecture derivation may learn this knowledge, such as "predation (cat→bite neck)", "predation (cat→mouse)", etc. When the concept of lion predation is input to the derivation, since the representation vector of lion is close to that of cat, the derivation may output concepts such as "bite neck" and "mouse", and generate conjecture knowledge "lions bite necks when preying" and "lions prey on mice". The simulator will then retrieve data related to lions from the database, and find out through statistics that lions basically bite their prey by the neck, but they rarely hunt mice. In this way, it can determine that the conjectured knowledge "predation (lion → bite the neck)" is true and "predation (lion → mouse)" is false, thereby enabling the system to autonomously acquire new knowledge about lions' predation.

[0060] b4.3-The knowledge translator will decide whether to pass the knowledge as a prompt to the conjecture derivation device based on the completeness (confidence) of the knowledge base and the task objectives. For example, if the system task is biased towards domain knowledge application rather than domain knowledge research, and the knowledge base has not been newly expanded after a long period of practical learning, the conjecture derivation device parameter structure will be too complex at this time, and inputting prompt will help the conjecture derivation device converge and output in a specific direction. If the knowledge base is not yet complete and the system may generate a lot of new knowledge in the future, then prompt will not be input, or prompt validity will be set to a smaller value. The form of prompt is obtained by vector transformation of the conjecture derivation network parameters corresponding to the single knowledge training during pre-training. Because the network parameters implicitly express the information of different knowledge and have a one-to-one mapping relationship with the knowledge, the parameter vector can effectively express the knowledge direction. For example, if the derivation network is fitted into the inverse function of the vector transformation, the knowledge parameters during data training can be perfectly reproduced.

[0061] b5- The simulator will search the database according to the input concepts involved in the new knowledge, and analyze whether the retrieved data contains the concepts derived from the new knowledge, as well as the difference between the data and the input data, so as to determine the boundary of the concept in the primitive space. Taking the generation of new concepts in b4.2 as an example, the system input X is a picture of a lion, but there is no concept of lion in the knowledge base. The conjecture deriver outputs a concept vector corresponding to the lion based on the primitive features identified by the concept generator, and the next step is to determine the interval boundary representing the lion in the primitive space around this concept vector. The simulator will retrieve data with high similarity in the database based on this concept vector. Since it is retrieved based on the primitive concepts of animals (head, body, limbs, etc.), these data may contain similar pictures of lions, tigers, leopards, etc., and the similarity value must be continuously reduced according to lions, tigers, and leopards. Then these pictures are matched with the pictures of lions as a whole, and it will be found that the pictures of tigers and leopards will have great differences from the pictures of lions, such as differences in skin pattern features and living environment features. In this way, the interval boundary of the concept of lion in the primitive space can be determined. In this way, the knowledge base will store the lion's characterization vector and the boundary of the characterization interval (which can be simplified to the characterization vector radius).

[0062] b6-If the simulator passes statistical verification and determines that the conjecture knowledge is invalid, the randomizer will be activated. The randomizer will randomly apply random perturbations to a small number of parameters in the conjecture derivation network and activate the conjecture derivation to reconverge and generate new conjecture knowledge. This process is equivalent to the evolution process of the control parameters of the controller by the genetic algorithm, which can accelerate the effective convergence of the conjecture derivation.

[0063] In addition, at different application stages, the generation module and the determination module work as follows:

[0064] Pre-training phase:

[0065] a1-Knowledge Base:

[0066] The knowledge base adopts a graphical model approach to store prior knowledge of the desired application domain, where the knowledge is expressed in a reasonable form, such as “relationship function F (initial concept set → target concept set)”.

[0067] a2-Knowledge Translator:

[0068] The knowledge translator pre-trains and stores the vector mapping of all concepts in the knowledge base, that is, it applies a unified neural network (such as compressed closed-loop transcription) to directly extract the feature vector expression of all concepts in the knowledge base from the data, realizes the feature expression of the primitive (the dimensional composition of the primitive space), and the vector expression of the object concept based on the space constructed by the primitive.

[0069] a3-Conjecture Deriver:

[0070] The conjecture deriver uses the vector of the initial concept in the relationship function as the output and the feature vector of the target concept as the supervisory signal to train the network, thereby learning the definitions of all object concepts in the knowledge base and the derivation relationship between object concepts, and storing the pre-trained network parameters in the corresponding knowledge F in the knowledge translator.

[0071] Application learning stage (image recognition application):

[0072] b1-Concept Generator:

[0073] When an image is input, the concept generator will extract the features of the input data and generate vector feature values ​​of the primitive concepts based on the primitive concepts defined in the knowledge base. For example, if you want to identify the cat in the picture, the extracted features will include primitive features such as head, body, legs, and tail. Since the object concept is represented by a spatial vector in a high-dimensional space constructed using primitive concepts, the concept generator will output a set of feature values ​​of the primitive concepts to the conjecture deriver, and at the same time match these primitive concept vectors with the corresponding knowledge concept symbols in the knowledge translator.

[0074] b2-Knowledge Translator:

[0075] The knowledge translator searches the knowledge base for relevant deduction logic based on the matching primitive concepts, and then passes the network parameter vector of the corresponding knowledge as a prompt to the conjecture deriver, and determines the confidence of the prompt based on the completeness of the knowledge base. For example, through the input features such as head, body, legs, and tail, the knowledge translator corresponds these feature values ​​to the definition of cats in the knowledge base, but in cats, these features may correspond to cats, tigers, leopards, lions, etc., then the knowledge translator passes the prompts of these knowledge to the conjecture deriver respectively, and the conjecture deriver will deduce and output the concept vectors of the identified objects for cats, tigers, leopards, and lions respectively.

[0076] b3-Conjecture Deriver:

[0077] The conjecture generator generates a conjecture concept vector Y based on the received concept vector feature value and knowledge prompt (including confidence), and transmits the conjecture concept back to the knowledge translator to confirm its corresponding concept symbol. If Y is consistent with a concept already in the knowledge translator, the knowledge translator establishes a mapping between the input concept of the conjecture deriver and the conjecture concept (conjecture knowledge, which is represented by a combination of concept symbols of all concepts involved in the knowledge); if Y is inconsistent with all concepts in the knowledge translator, a new concept is generated, and the preset symbol is used as the symbolic representation of the new concept, and the conjecture knowledge represented by the concept symbol is established. For example, the input of the conjecture deriver includes concept vectors such as horse head, horse body, horse legs, horse tail, unicorn, and wings. The deduced output concept vector is inconsistent with the concept of horse in the knowledge base, and there is no other consistent concept. Then, the preset symbol (category symbol + number, such as M1) will be used to represent the concept of "unicorn" and establish conjecture knowledge.

[0078] b4-simulator:

[0079] The simulator will verify the conjectured knowledge. Taking the unicorn described in b3 as an example, if the conjectured knowledge is "unicorns can fly" (since unicorns have wings, and it is easy to associate wings with flying through text learning, the conjecture derivation device may derive the conjectured knowledge of "unicorns can fly"), the simulator will retrieve image data similar to the concept of unicorn in the database, and then determine the truth or falsity of the conjectured knowledge by analyzing the feature values ​​in the image under the concept of flying.

[0080] b5 - Randomizer:

[0081] If the simulator determines that the conjectured knowledge is invalid, the randomizer is activated. The randomizer will randomly change the parameters of the conjecture derivation network, trigger the network to converge to the actual state distribution law, and output a new conjecture concept for the next judgment of the verifier.

[0082] Application learning stage (driving behavior learning):

[0083] b1-Concept Generator:

[0084] When an image is input, the concept generator will extract the features of the input data and generate vector feature values ​​of primitive concepts based on the primitive concepts defined in the knowledge base. For example, object concepts such as pedestrians, vehicles, markings, and traffic signs in the scene. The extracted features will include primitive features such as the object's motion, contour, and graphics. The concept generator will then set the generated object vectors to the conjecture derivation device and simultaneously match these object concept vectors with the corresponding knowledge concept symbols in the knowledge translator.

[0085] b2-Knowledge Translator:

[0086] The knowledge translator retrieves the relevant deduction logic in the knowledge base according to the matching object concept, and then passes the network parameter vector of the corresponding knowledge as a prompt to the conjecture deriver, and determines the confidence of the prompt according to the completeness of the knowledge base. For example, by inputting concepts such as pedestrians, lane markings, and crosswalks, the knowledge translator maps these concepts to the corresponding traffic regulations or driving test definitions in the knowledge base, and then the knowledge translator will pass the prompts of these knowledge to the conjecture deriver respectively, and the conjecture deriver will deduce the self-vehicle behaviors required by the traffic regulations and driving tests under the object concept combination and output the behavior concept vector.

[0087] b3-Conjecture Deriver:

[0088] The conjecture generator generates the conjecture concept vector Y based on the received concept vector and knowledge prompt (including confidence), and transmits the conjecture concept back to the knowledge translator to confirm its corresponding concept symbol. Since the vehicle behavior in driving is a complete set, there are only longitudinal acceleration and deceleration, parking, etc., and lateral left turn, right turn, straight driving, etc., and it also includes the negative expression of these behaviors, such as should not accelerate, should not turn, etc., then Y will inevitably match a certain behavior concept, and generate the mapping knowledge between the scene information corresponding to the input concept vector and the self-vehicle behavior conjecture Y (conjecture knowledge F (driving scene → driving behavior)).

[0089] b4-simulator:

[0090] The simulator will perform statistical analysis based on the data in the database that matches the concepts contained in the conjectured knowledge. These data will include the actual driver's behavior and safety result judgment (whether the driver's behavior results in safe driving or causing an accident) in the input scenario from which the conjectured knowledge comes, as well as historical experience data similar to the input scenario already in the database. The validity of the conjectured knowledge is determined by analyzing the distribution of the control amount of the steering wheel and acceleration and deceleration pedals in the experience data for safe behavior.

[0091] b5 - Randomizer:

[0092] If the simulator determines that the conjectured knowledge is invalid, the randomizer is activated. The randomizer will randomly change the parameters of the conjecture derivation network, trigger the network to converge according to the actual state distribution law, and output new conjecture knowledge for the next judgment of the simulator.

[0093] The present invention also discloses a method for autonomously acquiring and learning knowledge, comprising the following steps:

[0094] A concept generator receives an input data X, extracts data features of the input data X, and generates a representation vector of the data features under the structure of the primitive concept based on a primitive concept;

[0095] A conjecture deriver receives the representation vector of the input concept and the knowledge hint, deduce the conjecture concept vector Y, and generates the conjecture knowledge expression Y = F (X);

[0096] A simulator receives the conjectured knowledge Y=F(X) and refutes and verifies the knowledge based on similar empirical data in a database;

[0097] The knowledge base stores the knowledge of the application domain based on the graph model and predefines the primitive concepts;

[0098] A knowledge translator is pre-trained to store vector mappings of all primitive concepts in the knowledge base. The knowledge translator is connected to the concept generator and the conjecture deriver, receives input data X and extracts the concept representation vector corresponding to the input data X under the primitive structure, and sends the representation vector and the corresponding knowledge prompt to the conjecture deriver.

[0099] It should be noted that the embodiments of the present invention have better practicability and do not impose any form of limitation on the present invention. Any technician familiar with the field may use the technical content disclosed above to change or modify it into an equivalent effective embodiment. However, any modification or equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention are still within the scope of the technical solution of the present invention.

Claims

1. A system capable of autonomously acquiring and learning knowledge, characterized in that: It includes a generation module and a determination module, wherein The generation module comprises: A concept generator receives an input image data X, extracts data features of the input image data X, and generates vector feature values ​​of the data features under the structure of a predefined primitive concept based on the primitive concept, wherein the concepts of the vector feature values ​​generated by the concept generator of the generation module include primitive concepts and object concepts, the object concept is a vector representation of a high-dimensional space composed of elements of the primitive concept, the primitive concept represents the basic features of the application field, and all primitive concepts and object concepts correspond to a vector space under their feature dimensions; The conjecture derivation device inputs the representation vector of the concept, the user instruction and the related knowledge prompt, and the reasoning evolution result Y; the judgment module includes: The simulator receives the conjectured knowledge Y = F(X) and refutes and verifies it based on the empirical data in the database; The knowledge base stores the prior knowledge of the application domain based on the graph model and predefines the primitive concepts; A knowledge translator is pre-trained to store vector mappings of all primitive concepts in the knowledge base. The knowledge translator is connected to the concept generator and the conjecture derivation device, receives the input data X and extracts the representation vector under the corresponding primitive concept structure, and sends knowledge prompts related to the concept corresponding to the representation vector to the conjecture derivation device.

2. The system capable of autonomously acquiring and learning knowledge as claimed in claim 1, characterized in that: The knowledge base represents the knowledge concepts of prior knowledge and the conversion relationships between them in the form of a graph; The attribute form of the conversion relationship can be expressed as F, so that an initial concept is converted into a target concept, wherein the initial concept is a primitive concept or an object concept, and the target concept is an object concept; When the initial concept is a primitive concept, F represents the definition of the target concept represented by the combination of primitive concepts; if the initial concept is an object concept, F represents the inference relationship between the target concept and the object concept; The knowledge base extracts the derivation chain of prior knowledge related to the input image data X, verifies the derivation results one by one, and uses the conjecture concept vector Y obtained in the previous derivation as the input for the next knowledge derivation, so as to realize the derivation and verification of knowledge in stages.

3. The system capable of autonomously acquiring and learning knowledge as claimed in claim 2, characterized in that: The knowledge translator realizes the mapping between the concept symbols of the prior knowledge in the determination module and the conjecture concept vector Y in the generation module.

4. The system capable of autonomously acquiring and learning knowledge as claimed in claim 3, characterized in that: The conjecture derivation device translates the prior knowledge Y=F(X) in the knowledge base into a corresponding concept representation vector through a knowledge translator, wherein the vector translated from the initial concept X in F serves as the input of the conjecture derivation device, and the vector translated from the target concept Y in F serves as the labeled output of the conjecture derivation device, and the network parameters of the conjecture derivation device are pre-trained based on supervised learning.

5. The system capable of autonomously acquiring and learning knowledge as claimed in claim 4, characterized in that: The knowledge translator decides whether to pass the prior knowledge as a prompt to the conjecture derivation device according to the completeness of the knowledge base and the task goal; When the confidence of the knowledge base is lower than a confidence threshold, the prompt is not input to the hypothesis derivation device, or the confidence of the prompt is set to a smaller value.

6. The system capable of autonomously acquiring and learning knowledge as claimed in claim 1, characterized in that: When the conjecture derivation device generates conjecture knowledge Y=F(X), the simulator will refute and verify the knowledge based on similar empirical data in the database. If there is no counterexample, the knowledge is considered valid and the knowledge expression will be stored in the knowledge base.

7. The system capable of autonomously acquiring and learning knowledge as claimed in claim 1, characterized in that: Also includes: Randomizer; When the refutation verification of the simulator determines that the conjecture concept vector Y is invalid, the randomizer is activated, the randomizer randomly applies random perturbations to the parameters in the conjecture derivation network, and activates the conjecture derivation device to reconverge and generate a new conjecture concept vector Y and a new knowledge expression.

8. A method for acquiring knowledge that can be created autonomously, characterized in that: The following steps are involved: A concept generator receives an input image data X, extracts data features of the input image data X, and generates vector feature values ​​of the data features under the structure of a predefined primitive concept based on the primitive concept, wherein the concepts of the vector feature values ​​generated by the concept generator include primitive concepts and object concepts, the object concept is a vector representation of a high-dimensional space composed of elements of the primitive concept, the primitive concept represents the basic features of the application field, and all primitive concepts and object concepts correspond to a vector space under their feature dimensions; The conjecture derivation device inputs the concept representation vector, user instructions and related knowledge prompts, and infers the evolution result Y; The simulator receives the conjectured knowledge Y = F(X) and refutes and verifies it based on the empirical data in the database; The knowledge base stores the prior knowledge of the application domain based on the graph model and predefines the primitive concepts; A knowledge translator is pre-trained to store vector mappings of all primitive concepts in the knowledge base. The knowledge translator is connected to the concept generator and the conjecture derivation device, receives the input data X and extracts the representation vector under the corresponding primitive concept structure, and sends knowledge prompts related to the concept corresponding to the representation vector to the conjecture derivation device.

Citation Information

Patent Citations

  • Image semantic understanding and text generation based on reinforcement learning and contrast learning

    CN114972907A

  • Reinforcement learning knowledge graph reasoning method based on generative adversarial imitation learning

    CN115269861A

  • Method, system and device for automatically obtaining concepts in domain knowledge base and medium

    CN115270776A

  • Sentence-level question generation method based on syntactic perception prompt learning

    CN115759042A

  • Solid waste detection model construction method based on semi-supervised learning and solid waste detection method

    CN115965838A