Metalearning method for performing algorithm selection from learning curve data, electronic equipment and storage medium
By defining the meta-learning process as a reinforcement learning environment, using ALC and approximate greedy algorithms to optimize algorithm selection, the adaptability and resource consumption problems of the meta-learning system in diverse and complex scenarios are solved, and efficient algorithm selection and budget allocation are achieved. , improves the adaptability and resource utilization efficiency of the meta-learning system.
Patent Information
- Application Number
- CN202411061651.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-05
- Publication Date
- 2024-12-17
AI Technical Summary
The existing technology is still in its early stage in meta-learning, and cannot fully cope with diversified and complex practical application scenarios, high computing resource consumption, failure to effectively combine algorithm selection and budget allocation, resulting in poor resource utilization and performance performance and lack of Universal meta-learning strategy.
By defining the meta-learning process as a reinforcement learning environment and using ALC as the evaluation standard, the learning curve problem is reduced to the 0-1 backpack problem, and an approximate greedy algorithm is used to select the most cost-effective algorithm to optimize the meta-learning process and reduce the consumption of computing resources. Improve adaptability and resource utilization efficiency.
It realizes rapid achievement of excellent performance on the new data set, reduces computing resource consumption, optimizes algorithm selection and budget allocation, simplifies the meta-learning process, provides efficient solutions, and improves the adaptability and resource utilization efficiency of the meta-learning system.
Smart Images
Figure CN119150946A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of meta-learning technology, and in particular to a meta-learning method, electronic device and storage medium for algorithm selection from learning curve data. Background Art
[0002] With the rapid development of artificial intelligence (AI) and machine learning (ML) technologies, automated machine learning (AutoML) has become a hot research field in recent years. AutoML aims to optimize the construction and deployment process of machine learning models in an automated way, so that non-professionals can also use machine learning technology effectively. Among them, meta-learning plays an increasingly important role in automated machine learning. Meta-learning is defined as the process of "learning to learn" and is a technology that enables machine learning algorithms to use past experience to optimize future learning tasks.
[0003] In nature, organisms quickly adapt to changing environments by transferring acquired knowledge across tasks. This ability is common in organisms. However, compared with intelligent organisms in nature, artificial learning systems are still in a relatively "primary stage" in terms of meta-learning, and are currently mainly limited to knowledge transfer between very similar tasks. Current researchers have extensively explored multiple research areas such as few-shot learning, transfer learning, representation learning, continuous learning, lifelong learning, and meta-learning, and have proposed some methods to improve the efficiency of learning systems.
[0004] In the current technological context, society is increasingly concerned about the high consumption of computing resources by artificial intelligence. Therefore, it is particularly important to develop learning systems that can effectively reuse their knowledge. The development of such systems can not only save resources, but also improve learning efficiency, thereby accelerating the learning process and improving the performance of new tasks. To achieve this goal, researchers have proposed a variety of methods and algorithms, such as the aforementioned few-sample learning, transfer learning, representation learning, continuous learning, and lifelong learning.
[0005] One particularly interesting and promising area is meta-learning from learning curves, a relatively underexplored but fundamental area. Learning curves are an important tool for measuring the incremental performance improvements of algorithms, and can be evaluated as a function of training time, number of iterations, or number of samples. However, past challenges and research have often failed to fully exploit the potential of learning curves, viewing meta-learning only as a recommendation problem, or focusing only on specific aspects such as few-shot learning.
[0006] In response to this problem, organizers have proposed a new challenge to encourage academia and industry to conduct more in-depth research on the possibility of meta-learning from learning curves, focusing on the two simultaneous problems of algorithm selection and budget allocation. In the setting of this problem, the meta-learner needs to use the information of some training algorithms to reduce the cost of training them to convergence. In addition, this problem focuses on the all-time performance of the training curve on a new dataset, so ALC (Areaunder Learning Curve) is used to evaluate the proxy training strategy.
[0007] In general, although existing technologies have made some progress in the field of automated machine learning and meta-learning, they still have many shortcomings. For example: (1) Existing artificial learning systems are still in a relatively "primary stage" in terms of meta-learning, and are mainly limited to knowledge transfer between very similar tasks, and cannot fully cope with diverse and complex practical application scenarios.
[0008] (2) Existing learning systems consume a lot of computing resources, especially when multiple algorithms need to be frequently trained and evaluated. This problem is particularly prominent, hindering the practical application and promotion of the system.
[0009] (3) Although learning curves are an important tool for evaluating algorithm performance, past studies have often failed to fully utilize their potential. They have mainly viewed meta-learning as a recommendation problem, or have only focused on specific aspects such as learning with a small number of samples, lacking comprehensive and in-depth exploration and application.
[0010] (4) In existing research, algorithm selection and budget allocation issues are often treated separately without effectively combining the two, resulting in the inability to optimize resource utilization and performance in practical applications.
[0011] (5) Although some methods perform well in specific situations, there is generally a lack of a simple, efficient and universal meta-learning strategy to cope with diverse and complex practical application requirements.
[0012] Therefore, further innovation and optimization are needed to achieve a more efficient and universal learning system. Summary of the invention
[0013] The purpose of the present invention is to provide a meta-learning method, electronic device and storage medium for algorithm selection from learning curve data, to solve the problems existing in the prior art, to significantly improve the practical application value of automated machine learning and meta-learning systems, and to promote the application and development of artificial intelligence technology in a wider range of fields.
[0014] In order to achieve the above object, the solution of the present invention is: A meta-learning method for algorithm selection from learning curve data, comprising the following steps: Step 1. Data preparation and preprocessing 1-1. Collect a large amount of learning curve data of different algorithms trained on multiple data sets; 1-2. Arranging the learning curve data into structured data; Step 2. Definition of Meta-Learning Environment 2-1. Define the meta-learning process as a reinforcement learning environment consisting of an agent and an environment, where the agent takes actions in response to information and observations from the environment, and the algorithm plays the role of the agent. 2-2. Define the agent's actions and observations; an action is a tuple containing the algorithm index and the training data ratio, and an observation is a tuple containing the algorithm index, the training data ratio, the training time, the performance score on the training dataset, and the performance score on the validation dataset; Step 3. Determination of evaluation criteria Introducing ALC as an evaluation criterion; Step 4. Problem reduction The problem of meta-learning from learning curves is reduced to the 0-1 knapsack problem, and the following assumptions are made based on the physical meaning of the problem: Hypothesis 1: For the same algorithm, the larger the proportion of training data, the higher the time required and performance score; Assumption 2: For different algorithms, the time required for the same training data ratio is similar; Hypothesis 3: Under a limited time budget, the improvement of algorithm performance is related to the ratio of time and data; Step 5. Algorithm design and implementation 5-1. Sort the algorithms according to their cost-effectiveness and select the algorithm with the best cost-effectiveness; 5-2. Use pseudocode flowcharts to describe the algorithm's operation steps; update the training strategy based on observations in the actual environment and optimize the meta-learning process; Step 6. Algorithm evaluation and optimization 6-1. Test and evaluate the method flow of steps 1 to 5 during the development phase and final evaluation phase; 6-2. Compare with existing baseline methods, select cases for detailed analysis, and verify the effectiveness and advancement of the method.
[0015] In step 1, the structured data includes metadata of the model performance at a preset training time ratio in the training curve and the hyperparameter settings of the training algorithm.
[0016] In step 5, the cost performance of the algorithm refers to the ratio of the performance score to the training time.
[0017] An electronic device includes a processor, a memory and an application; the application is stored in the memory and is configured so that the processor executes the meta-learning method for algorithm selection from learning curve data.
[0018] A computer-readable storage medium stores a computer program; when the computer program is executed in a computer, the computer is caused to execute the meta-learning method for selecting an algorithm from learning curve data.
[0019] After adopting the above technical solution, the present invention has the following technical effects: The present invention aims to improve the adaptability and resource utilization efficiency of automated machine learning and meta-learning systems, and to achieve more efficient algorithm selection and budget allocation strategies, so as to quickly achieve excellent performance on new data sets; moreover, the present invention can improve the adaptability of meta-learning systems, reduce computing resource consumption, optimize algorithm selection and budget allocation, simplify the meta-learning process, and provide efficient solutions. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] Figure 1 It is a schematic diagram of the main steps of a specific embodiment of the present invention; Figure 2 4 is a comparison chart of the results of a specific embodiment of the present invention and other meta-learning methods. DETAILED DESCRIPTION
[0021] In order to further explain the technical solution of the present invention, the present invention is described in detail below through specific embodiments.
[0022] refer to Figure 1 As shown, the present invention discloses a meta-learning method for algorithm selection from learning curve data, comprising the following steps: Step 1. Data preparation and preprocessing 1-1. Collect a large number of learning curve data of different algorithms (including different hyperparameter settings of the same algorithm) trained on multiple data sets; 1-2. Organize the above learning curve data into structured data, including metadata such as model performance under the preset training time ratio in the training curve, hyperparameter settings of the training algorithm, etc., to ensure that it can be used in the subsequent meta-learning process; Step 2. Definition of Meta-Learning Environment 2-1. The meta-learning process is defined as a reinforcement learning environment consisting of an agent and an environment. The agent takes actions in response to information and observations from the environment, and the algorithm plays the role of the agent. 2-2. Define the agent's actions and observations; an action is a tuple containing the algorithm index and the training data ratio, and an observation is a tuple containing the algorithm index, the training data ratio, the training time, the performance score on the training dataset, and the performance score on the validation dataset; Step 3. Determination of evaluation criteria ALC (Area under Learning Curve) is introduced as an evaluation criterion to focus on the performance of the model at all time points to measure its ALL-TIME Performance; Step 4. Problem reduction The problem of meta-learning from learning curves is reduced to the 0-1 knapsack problem, and the following assumptions are made based on the physical meaning of the problem: Hypothesis 1: For the same algorithm, the larger the proportion of training data, the higher the time required and performance score; Assumption 2: For different algorithms, the time required for the same training data ratio is similar; Hypothesis 3: Under a limited time budget, the improvement of algorithm performance is related to the ratio of time and data; Step 5. Algorithm design and implementation 5-1. Since dynamic programming is not feasible in practical applications, a solution based on an approximate greedy algorithm is proposed: sort the algorithms according to their cost-effectiveness (performance score / training time) and select the algorithm with the best cost-effectiveness; 5-2. Implement the agent strategy of the method and provide specific action suggestions: Use pseudocode flowcharts to describe the operation steps of the algorithm; update the training strategy based on observations in the actual environment and optimize the meta-learning process; Step 6. Algorithm evaluation and optimization 6-1. During the development phase and final evaluation phase, the method flow of steps 1 to 5 is tested and evaluated.
[0023] 6-2. Compare with various existing baseline methods such as DDQN (Double Deep Q Network), FT (Freeze-Thaw Bayesian Optimization), random search, etc., select cases for detailed analysis, and verify the effectiveness and advancement of the method.
[0024] Through the above technical steps, the present invention aims to improve the adaptability and resource utilization efficiency of automated machine learning and meta-learning systems, realize more efficient algorithm selection and budget allocation strategies, and thus quickly achieve excellent performance on new data sets; and the present invention can achieve the following purposes: a. Improve the adaptability of meta-learning systems By meta-learning from the algorithm learning curves on past data sets, the algorithm selection on new data sets is optimized, enabling machine learning systems to better adapt to diverse and complex practical application scenarios.
[0025] b. Reduce computing resource consumption The learning system can effectively reuse knowledge to save computing resources, improve learning efficiency, accelerate the learning process, and improve the performance of new tasks.
[0026] c. Optimization algorithm selection and budget allocation ALC (Area under Learning Curve) is introduced as an evaluation criterion to focus on the ALL-TIME Performance of the model. Under limited budget conditions, the best training strategy can be implemented to ensure that the model performs well regardless of the training budget.
[0027] d. Provide efficient solutions By reducing the problem of meta-learning from learning curves to a 0-1 knapsack problem, a GreedyAgent method based on an approximate greedy algorithm is proposed. It can replace the dynamic programming algorithm in an environment with insufficient conditions, obtain an approximate optimal solution, and improve the efficiency and effectiveness of meta-learning.
[0028] e. Simplify the meta-learning process Through clear and complete mathematical arguments and concrete and available code, an efficient and concise meta-learning and transfer learning method is demonstrated, which enables a reliable strategy for meta-learning and transfer learning from the learning curve without consuming a lot of computing power.
[0029] The present invention also discloses an electronic device, comprising a processor, a memory and an application; the application is stored in the memory and is configured to execute the above-mentioned meta-learning method by the processor.
[0030] The present invention also discloses a computer-readable storage medium, which stores a computer program; when the computer program is executed in a computer, the computer executes the above-mentioned meta-learning method. The storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.
[0031] The above embodiments and drawings do not limit the product form and style of the present invention. Any appropriate changes or modifications made by ordinary technicians in the relevant technical field should be deemed to be within the patent scope of the present invention.
Claims
1. A meta-learning method for algorithm selection from learning curve data, characterized in that The following steps are involved: Step 1. Data preparation and preprocessing 1-1. Collect a large amount of learning curve data of different algorithms trained on multiple data sets; 1-2. Arranging the learning curve data into structured data; Step 2. Definition of Meta-Learning Environment 2-1. Define the meta-learning process as a reinforcement learning environment consisting of an agent and an environment, where the agent takes actions in response to information and observations from the environment, and the algorithm plays the role of the agent. 2-2. Define the agent's actions and observations; an action is a tuple containing the algorithm index and the training data ratio, and an observation is a tuple containing the algorithm index, the training data ratio, the training time, the performance score on the training dataset, and the performance score on the validation dataset; Step 3. Determination of evaluation criteria Introducing ALC as an evaluation criterion; Step 4. Problem reduction The problem of meta-learning from learning curves is reduced to the 0-1 knapsack problem, and the following assumptions are made based on the physical meaning of the problem: Hypothesis 1: For the same algorithm, the larger the proportion of training data, the higher the time required and performance score; Assumption 2: For different algorithms, the time required for the same training data ratio is similar; Hypothesis 3: Under a limited time budget, the improvement of algorithm performance is related to the ratio of time and data; Step 5. Algorithm design and implementation 5-1. Sort the algorithms according to their cost-effectiveness and select the algorithm with the best cost-effectiveness; 5-2. Use pseudocode flowcharts to describe the algorithm's operation steps; update the training strategy based on observations in the actual environment and optimize the meta-learning process; Step 6. Algorithm evaluation and optimization 6-1. Test and evaluate the method flow of steps 1 to 5 during the development phase and final evaluation phase; 6-2. Compare with existing baseline methods, select cases for detailed analysis, and verify the effectiveness and advancement of the method.
2. The meta-learning method for algorithm selection from learning curve data according to claim 1, characterized in that: In step 1, the structured data includes metadata of the model performance at a preset training time ratio in the training curve and the hyperparameter settings of the training algorithm.
3. The meta-learning method for algorithm selection from learning curve data according to claim 1, characterized in that: In step 5, the cost performance of the algorithm refers to the ratio of the performance score to the training time.
4. An electronic device, characterized in that: The invention comprises a processor, a memory and an application program; the application program is stored in the memory and is configured so that the processor executes the meta-learning method for algorithm selection from learning curve data as described in any one of claims 1 to 3.
5. A computer-readable storage medium, characterized in that: The storage medium stores a computer program; when the computer program is executed in a computer, the computer is caused to execute the meta-learning method for algorithm selection from learning curve data as described in any one of claims 1 to 3.