Abstract classification and extraction method of artificial intelligence skill program

Abstract

The application relates to an abstract classification and extraction method of an artificial intelligence skill program, and comprises the following steps: 1. creating an artificial intelligence skill program library, and defining the data structure of a single skill program in a triple; 2. creating a vocabulary table of all elements contained in all skill programs in the skill program library, and classifying the skill programs by using cosine similarity; 3. calculating the cluster and the centroid of the triple set of each skill program in each class by using a clustering algorithm, and obtaining program abstraction; and 4. distributing and storing skill program files of all classes, and establishing an index in the unit of the skill program. The application can improve the intuition and generalization ability of a model, and further improve the efficiency of searching and combining skill programs for solving problems or completing specified tasks when the model faces problems and tasks that have not been encountered, so that the artificial intelligence technology paradigm is changed from deep learning with a model and value as cores to an interpretable system with program combination as a core.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence program synthesis, specifically involving a method for abstracting, classifying, and extracting skill programs based on a triplet data model and cluster analysis. Background Technology

[0002] In the field of artificial intelligence, to enable models to possess extreme generalization capabilities for abstraction and reasoning, deep learning techniques can be combined with program synthesis techniques. Program synthesis refers to the process of searching for suitable program components in a vast program space and generating new programs. Due to the enormous number of permutations and combinations during the process of searching for program components and synthesizing new programs, a combinatorial explosion often occurs. To solve the combinatorial explosion problem in program synthesis with fewer computational resources, it is necessary to abstract and classify skill programs, constructing a hierarchical structure of classes to improve the efficiency of program search and combination selection.

[0003] Existing technologies lack a systematic abstract classification mechanism for skill procedures, making it difficult to construct hierarchical structures. This limits the accuracy and generalization ability of model program intuition, and fails to meet the needs of efficient program synthesis for complex tasks. Therefore, there is an urgent need for a method that can accurately abstract and classify skill procedures and optimize the storage and retrieval process to solve the combinatorial explosion problem in program synthesis and improve the extreme generalization ability of models. Summary of the Invention

[0004] This invention provides an abstract classification and extraction method for artificial intelligence skill programs. This method can effectively summarize different skill programs into higher-dimensional and higher-level abstractions, improve the accuracy and reliability of model program intuition, and further enhance the extreme generalization ability of the model.

[0005] To address the aforementioned problems, this invention provides a method for abstracting, classifying, distributing, and searching and retrieving skill programs from terminal hardware, which can effectively improve the efficiency of program synthesis.

[0006] This invention provides a method for abstracting, classifying, and extracting artificial intelligence skill programs, comprising the following steps:

[0007] 1) Create an AI skill program library: SKILLS={S1, S2, ..., Sn}, where Si (1≤i≤n) is the i-th skill program, and n is the total number of skill programs. Define the data structure of a single skill program using a triple Si={Di, Ri, Mi}, where: Di is the set of object elements of Si {Di1, Di2, ... Dij} (Di has j elements); Ri is the set of relation elements defined on Di {Ri1, Ri2, ... Rik} (Ri has k elements); Mi is the set of operation elements defined on Di {Mi1, Mi2, ... Mil} (Mi has l elements).

[0008] 2) Create a vocabulary containing all object elements, relational elements, and operation elements contained in all skill programs Si in the above skill program library SKILLS, encode them with one-hot encoding, and classify the skill programs using cosine similarity.

[0009] 3) Learn word vectors (word embeddings) of words in the vocabulary list from the corpus to obtain the multi-features of the vocabulary. The feature values ​​are normalized using the Softmax function. Use the K-means clustering algorithm to calculate the classification and centroid of the set of all object elements, the set of relation elements, and the set of operation elements for each skill program, so as to summarize the abstraction of the skill program.

[0010] 4) Distribute and store different types of skill programs (including their abstractions) across different storage units. Create an index based on the categorized and distributed skill programs.

[0011] Another aspect of the present invention provides a terminal hardware device, which may be a personal computer, a server or a robot, including at least one processor and a memory.

[0012] Memory, used to store data files of program instructions and skill programs;

[0013] The processor is used to call and execute program instructions stored in memory; data reading is achieved through asynchronous I / O.

[0014] The improvement of this invention compared to the prior art lies in the fact that by abstracting and classifying the skill programs in the system, the intuition and generalization ability of the model are improved, and the model is further improved in terms of efficiently searching and combining skill programs that can solve problems or complete specified tasks when faced with problems and tasks that have not been encountered before. This promotes the transformation of the artificial intelligence technology paradigm from deep learning, which is centered on models and values, to interpretable systems, which are centered on program synthesis.

[0015] The standardized definition of skill procedures is achieved by using a triplet data structure, which solves the problem of chaotic and difficult-to-process program elements and provides a data foundation for classification and abstraction.

[0016] By combining one-hot encoding and cosine similarity classification, the accuracy of skill procedure classification is improved, providing a precise grouping basis for subsequent abstraction. By training word vectors and K-means clustering through the Transformer architecture, efficient abstraction of skill procedures is achieved, enhancing the model's abstract reasoning ability.

[0017] Distributed storage and index building, combined with parallel retrieval from multiple I / O ports, significantly improve the retrieval and reading efficiency of skill programs, effectively solve the combinatorial explosion problem in program synthesis, and reduce computing power consumption.

[0018] It supports multi-level abstraction, enabling dimensional upgrades of skill programs and driving the transformation of artificial intelligence technology from deep learning to interpretable systems centered on program synthesis. Applicable scenarios cover various terminal devices such as personal computers, servers, and robots. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the steps of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0021] Example 1

[0022] The following is combined with Figure 1 The embodiments of the present invention will be further described as follows:

[0023] See Figure 1 As shown, this invention provides a method for abstracting, classifying, and extracting artificial intelligence skill programs. The method includes:

[0024] 1) Create an AI skill program library: SKILLS={S1, S2, ..., Sn}, where Si (1≤i≤n) is the i-th skill program, and n is the total number of skill programs. Define the data structure of a single skill program using a triple Si={Di, Ri, Mi}, where: Di is the set of object elements of Si {Di1, Di2, ... Dij} (Di has j elements); Ri is the set of relation elements defined on Di {Ri1, Ri2, ... Rik} (Ri has k elements); Mi is the set of operation elements defined on Di {Mi1, Mi2, ... Mil} (Mi has l elements).

[0025] In this embodiment, the data source can be an existing open-source dataset or a self-constructed dataset. If a self-constructed dataset is used, the text description of the skill program needs to be cleaned. Data cleaning includes, but is not limited to, extracting, deduplicating, filtering, and refining the raw data to obtain concise and high-quality data. After data cleaning, the data needs to be tokenized and the tokens classified. The tokens are stored in the sets D, R, and M of the above triplet data structure according to "object element", "relation element", and "operation element", respectively.

[0026] 2) Create a vocabulary containing all object elements, relational elements, and operation elements contained in all skill programs Si in the above skill program library SKILLS, encode them with one-hot encoding, and classify the skill programs using cosine similarity.

[0027] In this embodiment, each element of Di, Ri, and Mi in Si corresponds to the index of the word in the vocabulary. The one-hot encoding of Si is a one-dimensional tensor of the form [1, 0, 1, 0, ..., 0, 0, 0], which is used... This indicates that the position of element i corresponds to the index of the word in the vocabulary of each element in the sets Di, Ri, and Mi in Si.

[0028] The formula for calculating cosine similarity is:

[0029] cosine_similarity( , )= ,

[0030] in, Representing vectors The dot product;

[0031] They are vectors The norm of .

[0032] 3) Word vectors (word embeddings) of words in the vocabulary are learned from the corpus to obtain the multi-features of the vocabulary. The feature values ​​are normalized using the Softmax function. The K-means clustering algorithm is used to calculate the classification and centroid of the set of all object elements, the set of relation elements, and the set of operation elements for each skill program, in order to summarize the abstraction of that skill program. The word vectors (word embeddings) can be obtained using a deep learning training method based on the Transformer architecture with an attention mechanism.

[0033] In this embodiment, it is assumed that S1, S2, ..., Sm are the same type of skill program.

[0034] Where S1 = {D1, R1, M1}, ..., {Dm, Rm, Mm},

[0035] Let S' = {D1 ⋃D2 ⋃... ⋃Dm, R1 ⋃R2 ⋃... ⋃Rm, M1 ⋃M2 ⋃... ⋃Mm},

[0036] Calculate Nd cluster centroids of D1 ⋃D2 ⋃... ⋃Dm: d1', ..., dnd', D' = {d1', ..., dnd'}, where:

[0037] Min[N(Di)]≤Nd≤ (1≤i≤m), N(Di) is the number of elements in Di;

[0038] Similarly, R1 ⋃R2 ⋃... ⋃Rm has Nr cluster centroids: r1', ..., rnr', R'={r1', ..., rnr'}, where:

[0039] Min[N(Ri)]≤Nr≤ (1≤i≤m), N(Ri) is the number of elements in Ri;

[0040] Similarly, M1 ⋃M2 ⋃... ⋃Mm has Nm clusters of centroids: m1', ..., mnr', M'={m1', ..., mnr'}, where:

[0041] Min[N(Mi)]≤Nm≤ (1≤i≤m), N(Mi) is the number of elements in Mi;

[0042] ={D', R', M'} is the abstraction of this class.

[0043] It should be understood that, based on the above steps, continuing the operations from 1) to 3) on the already abstracted skill program class can obtain an abstraction of abstraction, that is, an abstraction of a higher dimension and level.

[0044] 4) Distribute and store different types of skill programs (including their abstractions) across different storage units. Create an index based on the categorized and distributed skill programs.

[0045] Example 2

[0046] A terminal hardware device, which may be a personal computer, a server, or a robot, includes at least one processor and a memory. The memory is used to store program instructions and data files of skill programs; the processor is used to call and execute program instructions stored in the memory, and data reading is achieved through asynchronous I / O.

[0047] In this embodiment, when searching for skill programs, the skill programs most likely to complete the specified task can be matched according to the semantics and structure of the specified task. Different types of skill programs can be retrieved and read through multiple I / O read / write ports of hardware (including personal computers, servers or robots) to achieve parallel computing. The computing refers to applying the searched skill programs to several instances to verify whether the skill programs of this type can complete the specified task.

[0048] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely software embodiment, a completely hardware embodiment, or an embodiment combining software and hardware. This application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0049] This invention is not limited to the embodiments described above. Those skilled in the art can make various adjustments and improvements without departing from the principles of this invention, and these adjustments and improvements should also be considered within the scope of protection of this invention. Contents not described in detail in this specification are prior art known to those skilled in the art.

Claims

1. A method for abstract classification and extraction of artificial intelligence skill programs, characterized in that, Includes the following steps: 1) Create an AI skill program library, defining the data structure of a single skill program using triples; 2) Create a vocabulary of all elements contained in all skill programs in the above skill program library, and classify the skill programs using cosine similarity; 3) Learn word vectors from the corpus, and use clustering algorithms to calculate the clusters and centroids of each tuple in each skill program triplet to obtain program abstraction; 4) Distributed storage in various skill programs.

2. The method for abstracting, classifying, and extracting artificial intelligence skill programs according to claim 1, characterized in that, If you build your own dataset, you need to clean and segment the original data after step 1) and before step 2).

3. The method for abstract classification and extraction of artificial intelligence skill programs according to claim 2, characterized in that, The data cleaning process includes extracting, deduplicating, filtering, and refining the raw data.

4. The method for abstracting, classifying, and extracting artificial intelligence skill programs according to claim 2, characterized in that, After word segmentation, the tokens are categorized into "object elements", "relational elements", and "operation elements" and stored in sets D, R, and M of the triplet data structure, respectively.

5. The method for abstract classification and extraction of artificial intelligence skill programs according to claim 1, characterized in that, One-hot encoding is performed on the vocabulary, and skill programs are classified using cosine similarity. The formula for calculating the cosine similarity is as follows: cosine_similarity( , )= , in, Representing vectors The dot product; They are vectors The norm of .

6. The method for abstract classification and extraction of artificial intelligence skill programs according to claim 1, characterized in that, Word vectors (word embeddings) can be acquired using a deep learning training method based on the Transformer architecture with an attention mechanism. The clustering algorithm uses K-means to calculate the clusters and centroids of each set of tuples in various skill program triples.

7. The method for abstract classification and extraction of artificial intelligence skill programs according to claim 1, characterized in that, Parallel computing is achieved by retrieving and reading skill program files through multiple I / O read / write ports of hardware, including personal computers, servers, or robots.

8. A terminal hardware device, characterized in that, Includes at least one processor and one memory; Memory, used to store data files of program instructions and skill programs; The processor is used to call and execute program instructions stored in memory, so that the terminal hardware device executes the abstract classification and extraction method of artificial intelligence skill programs as described in any one of claims 1-7.

9. The terminal hardware device according to claim 8, characterized in that, The memory supports a distributed storage architecture, and data reading is achieved through asynchronous I / O. The processor controls multiple I / O read / write ports to complete the parallel retrieval and extraction of skill programs.

Images