Multi-modal large model fusion training method and system based on biological evolutionary algorithm

Abstract

The invention discloses a multi-modal large model fusion training method and system based on a biological evolutionary algorithm, and relates to a method and system for performing fusion and model selection on a multi-modal large model by using a framework of the biological evolutionary algorithm, selecting characteristic parameters and capabilities to form different models, performing fusion training in a manner similar to biological evolution, and obtaining high-performance AI (artificial intelligence). According to the concept, natural selection, genetic variation and other mechanisms in biological evolution are mainly used for reference and applied to the optimization process of an artificial intelligence model, a self-adaptive training system is formed, and the AI performance is continuously improved in continuous iterative training.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence model technology, and specifically to a multi-modal large model fusion training method and system based on a biological evolution algorithm. Background Art

[0002] In recent years, with the rapid development of the field of artificial intelligence, large models have achieved remarkable results in various tasks. However, large model training requires a lot of computing resources and data volume, which has become a bottleneck problem for the implementation of large models in the industry. At the same time, a single model is often difficult to handle complex scenarios and lacks the generalization ability to migrate to multiple scenarios. Therefore, how to integrate multiple models to improve the performance and generalization ability of the model has become a problem that needs to be solved urgently and is also a research direction with great potential.

[0003] Early model fusion methods mainly focused on ensemble learning and voting strategies, such as bagging, boosting, and random forests. With the development of deep learning methods, researchers have begun to try to integrate multiple neural network models. Common methods include model fusion, model stacking, and model distillation. In recent years, the Transformer architecture has made great progress in natural language processing and computer vision, including multimodal learning tasks. At the same time, with the widespread application of multimodal data, multimodal model fusion has become a research hotspot, and some new model fusion strategies and methods have been proposed, including those based on knowledge distillation and attention mechanisms. These methods aim to fuse information from different modalities (such as images, text, speech, etc.) into a unified model to improve the performance and generalization ability of the model.

[0004] However, the current methods have the following problems:

[0005] 1. Computational resource consumption: Model fusion methods usually require training and integrating multiple models, which consumes a lot of computing resources and time, especially on large-scale datasets and complex tasks.

[0006] 2. Reliance on expert knowledge and human intuition: Model fusion involves parameter tuning of multiple models and the selection of fusion strategies, which requires certain experience and skills. Currently, trainers select fusion strategies based on experience and intuition.

[0007] 3. Ensemble overfitting: In the process of model fusion, if the model or ensemble strategy is not selected properly, the ensemble model may overfit the training data and reduce the generalization ability of the model.

[0008] 4. Integration consistency: Differences between different models or data sources may lead to reduced consistency of the integrated model after model fusion, affecting the stability and robustness of the model.

[0009] 5. Data imbalance and noise: The model fusion process may face problems such as data imbalance and noise, which will affect the training and generalization capabilities of the model.

[0010] A system for multimodal large-model fusion training based on a biological evolutionary algorithm includes a model selection system, a model generation system, a model evaluation system, a model processing system and a model determination system. The model selection system, the model generation system, the model evaluation system, the model processing system and the model determination system are connected in sequence through electrical signals, and the model determination system is connected to the model generation system through electrical signals. Summary of the invention

[0011] 1. Technical issues to be solved

[0012] In view of the deficiencies in the prior art, the present invention provides a multi-modal large model fusion training method and system based on a biological evolution algorithm.

[0013] (II) Technical solution

[0014] To achieve the above object, the present invention provides the following technical solution: a multimodal large model fusion training method based on a biological evolution algorithm, the multimodal large model fusion training method based on a biological evolution algorithm comprises:

[0015] Initialize the population: According to the specific problem, select the initial model, initialize the model parameters, and form the initial model;

[0016] Model crossover: Use the previous generation model for crossover to generate a child model;

[0017] Model selection: Evaluate the fitness of offspring models and select offspring models with better performance;

[0018] Model mutation: select some offspring models and perform mutation operations;

[0019] Iterative optimization: Determine whether the evolution process meets the termination conditions. If not, continue the above process.

[0020] Furthermore, the present invention has been improved in that the population initialization operation includes: first, an initial model population needs to be created, and these models can be neural networks with different parameters and structures. The diversity of the initialization population is crucial to the subsequent evolutionary process because it determines the breadth and depth of the search space.

[0021] Furthermore, the present invention is improved in that model crossover can generate a new model by combining the characteristics of two or more excellent models, for example, methods such as structural crossover or parameter crossover of neural networks can be used.

[0022] Furthermore, the present invention has the following improvements: model selection is performed based on fitness. Fitness is an important indicator for measuring the survival and reproductive capacity of an individual. During the evolution of an AI model, fitness is usually determined by evaluating the performance of the model on a specific task. For example, indicators such as accuracy, recall, and F1 score can be used to measure the performance of the model.

[0023] Furthermore, the present invention has the following improvements: the mutation operation increases model diversity and expands the model search space by introducing new features to prevent falling into local optimal solutions; the mutation methods include randomly changing certain parameters or structures of the model and introducing noise using Gaussian functions.

[0024] Furthermore, the present invention has the following improvements: by continuously repeating the above process (i.e., model crossover, selection, and mutation operations), the model performance can be gradually optimized, and a maximum number of iterations or a fitness threshold can be set to determine whether to stop the search until the possible model space is fully explored and the optimal solution is found.

[0025] This application also provides the following technical solutions:

[0026] 1. Initial model selection: select basic models based on the problems to be solved and specific scenarios. These basic models correspond to the initial population in the genetic algorithm. The selection of the initial population has an important impact on the final integrated model and evolution efficiency. For a single model, it is necessary to comprehensively consider dimensions such as model performance, model complexity, and adaptability. At the same time, the diversity of models in the population also needs to be considered.

[0027] 2. Model Crossover, this application uses three model crossover methods;

[0028] a. Based on the parameter space, by integrating the parameters of multiple parent models, a model with the same architecture as the parent model is generated, and by linearly or nonlinearly combining the model parameters, the parameters of the child model are generated;

[0029] b. Based on the data flow space, keep the parameters of the parent model and change the hierarchical structure after input, for example, after the i-th layer of model A, connect to the j-th layer of model B. Choosing different data flow structures will generate different child models;

[0030] c. Based on both parameter space and data stream space, combined with methods based on parameter space and data stream space, different offspring models are generated through the parameter combination of the parent model and the hierarchical structure of the data stream. This method considers both parameter space and data stream space, increases the diversity of candidate offspring, and requires more computational cost due to the increase in search space.

[0031] 3. Offspring model selection: among the generated offspring models, evaluate them according to the specific scenario and optimization goal, and select the model with better performance. Common selection strategies include roulette selection, tournament selection, random selection, competitive selection, and mixed strategy selection. Common selection criteria include accuracy, recall, AUC, F1-Score, etc. When selecting offspring models, it is necessary to consider offspring diversity and convergence speed.

[0032] 4. Sub-generation mutation operation: increase model diversity and expand model search space through mutation operation;

[0033] a. Parameter variation: by adjusting the combined parameters or hyperparameters of some offspring models, adding random perturbations and changing them according to specific rules;

[0034] b. Structural variation: Structural variation is performed by adding, deleting or modifying the model's hierarchy, node connections or node attributes. For example, you can randomly add or delete hidden layers or nodes in a neural network, or modify the filter size or number of channels in a convolutional neural network;

[0035] c. Combined mutation: In addition to the mutation operation of a single model, you can also perform mutation by combining different models. Randomly select a group of models and combine them into a new model to obtain more powerful model performance.

[0036] Determine when to terminate the evolution process, based on the evolution termination conditions set in advance and the accuracy of the current offspring model. If the conditions are met, terminate. If not, continue the above process and iterate the optimization.

[0037] (III) Beneficial effects

[0038] Compared with the prior art, the present invention provides a multi-modal large model fusion training method and system based on biological evolution algorithm, which has the following beneficial effects:

[0039] 1. The present invention can integrate models across fields and solve complex problems. The integrated model has universality and generalization, and has characteristics that each basic model does not have;

[0040] 2. No need to train sub-models separately, reducing computational costs;

[0041] 3. It does not rely on manual selection of models by domain experts, greatly improving iteration efficiency.

[0042] In general, this method has the characteristics of global optimization and strong adaptability, has good results in solving complex problems, and is conducive to the industrial application and innovation of large models. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 Schematic diagram of the model selection steps based on evolutionary algorithm in the present invention. DETAILED DESCRIPTION

[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0045] See also Figure 1 The multimodal large model fusion training method based on the biological evolution algorithm of the present invention comprises:

[0046] Initialize the population: According to the specific problem, select the initial model, initialize the model parameters, and form the initial model;

[0047] Model crossover: Use the previous generation model for crossover to generate a child model;

[0048] Model selection: Evaluate the fitness of offspring models and select offspring models with better performance;

[0049] Model mutation: select some offspring models and perform mutation operations;

[0050] Iterative optimization: Determine whether the evolution process meets the termination conditions. If not, continue the above process.

[0051] Furthermore, the present invention has been improved in that the population initialization operation includes: first, an initial model population needs to be created, and these models can be neural networks with different parameters and structures. The diversity of the initialization population is crucial to the subsequent evolutionary process because it determines the breadth and depth of the search space.

[0052] Model crossover can generate a new model by combining the characteristics of two or more excellent models. For example, methods such as structural crossover or parameter crossover of neural networks can be used.

[0053] Model selection is performed based on fitness, which is an important indicator for measuring individual survival and reproduction capabilities. During the evolution of AI models, fitness is usually determined by evaluating the performance of the model on a specific task. For example, accuracy, recall, F1 score, and other indicators can be used to measure model performance.

[0054] The mutation operation introduces new features, increases model diversity, expands the model search space, and prevents falling into local optimal solutions. Mutation methods include randomly changing certain parameters or structures of the model, and using Gaussian functions to introduce noise.

[0055] By repeating the above process (i.e., model crossover, selection, and mutation operations), the model performance can be gradually optimized, and the maximum iteration rounds or fitness threshold can be set to determine whether to stop the search until the possible model space is fully explored and the optimal solution is found.

[0056] 1. Initial model selection: select basic models based on the problems to be solved and specific scenarios. These basic models correspond to the initial population in the genetic algorithm. The selection of the initial population has an important impact on the final integrated model and evolution efficiency. For a single model, it is necessary to comprehensively consider dimensions such as model performance, model complexity, and adaptability. At the same time, the diversity of models in the population also needs to be considered.

[0057] 2. Model Crossover, this application uses three model crossover methods;

[0058] a. Based on the parameter space, by integrating the parameters of multiple parent models, a model with the same architecture as the parent model is generated, and by linearly or nonlinearly combining the model parameters, the parameters of the child model are generated;

[0059] b. Based on the data flow space, keep the parameters of the parent model and change the hierarchical structure after input, for example, after the i-th layer of model A, connect to the j-th layer of model B. Choosing different data flow structures will generate different child models;

[0060] c. Based on both parameter space and data stream space, combined with methods based on parameter space and data stream space, different offspring models are generated through the parameter combination of the parent model and the hierarchical structure of the data stream. This method considers both parameter space and data stream space, increases the diversity of candidate offspring, and requires more computational cost due to the increase in search space.

[0061] 3. Offspring model selection: among the generated offspring models, evaluate them according to the specific scenario and optimization goal, and select the model with better performance. Common selection strategies include roulette selection, tournament selection, random selection, competitive selection, and mixed strategy selection. Common selection criteria include accuracy, recall, AUC, F1-Score, etc. When selecting offspring models, it is necessary to consider offspring diversity and convergence speed.

[0062] 4. Sub-generation mutation operation: increase model diversity and expand model search space through mutation operation;

[0063] a. Parameter variation: by adjusting the combined parameters or hyperparameters of some offspring models, adding random perturbations and changing them according to specific rules;

[0064] b. Structural variation: Structural variation is performed by adding, deleting or modifying the model's hierarchy, node connections or node attributes. For example, you can randomly add or delete hidden layers or nodes in a neural network, or modify the filter size or number of channels in a convolutional neural network;

[0065] c. Combined mutation: In addition to the mutation operation of a single model, you can also perform mutation by combining different models. Randomly select a group of models and combine them into a new model to obtain more powerful model performance.

[0066] Determine when to terminate the evolution process, based on the evolution termination conditions set in advance and the accuracy of the current offspring model. If the conditions are met, terminate. If not, continue the above process and iterate the optimization.

[0067] A multimodal large model fusion training system based on a biological evolution algorithm, comprising a model selection system, a model generation system, a model evaluation system, a model processing system and a model determination system, wherein the model selection system, the model generation system, the model evaluation system, the model processing system and the model determination system are connected in sequence through electrical signals, and the model determination system is connected to the model generation system through electrical signals;

[0068] Among them, the model selection system is used to select the basic model, the model generation system is used to cross models and generate offspring models, the model evaluation system is used to evaluate offspring models and make selections, the model processing system is used to process offspring models, and the model judgment system is used to determine whether the termination conditions are met.

[0069] In order to obtain a high-performance and universal large model, this application provides a model fusion framework based on the genetic algorithm paradigm, which can systematically solve the model fusion problem. This application provides an example of the fusion of the Chinese language model and the math problem model. The specific contents are as follows:

[0070] 1. To solve Chinese math problems, you need a Chinese prediction model and a math problem model. Select: bert-base-chinese, WizardMath-7B-V1.1, Abel-7B-002 respectively;

[0071] 2. Use the Math23K dataset to evaluate the effects of different ensemble models. Math23K contains a large number of Chinese math problems, covering basic math concepts and problem types. The evaluation criteria are whether the answer is correct and the reasoning process needs to be generated in Chinese.

[0072] 3. Use CMA-ES evolutionary algorithm. CMSA-ES is a gradient-free random optimization algorithm suitable for high-dimensional optimization problems and situations without gradient information. Its core idea is to optimize the problem by maintaining a multivariate Gaussian distribution representing the search space and constantly adjusting the mean and covariance matrix of the distribution. The process is as follows:

[0073] a. Initialization: Randomly initialize the mean and covariance matrix of the population and set the initial parameters;

[0074] b. Generate offspring: Generate new candidate solutions based on the current mean and covariance matrix, usually using multivariate Gaussian distribution to generate mutation vectors;

[0075] c. Selection: Evaluate candidate offspring through the fitness function and select individuals with higher fitness as the next generation population. Fitness is defined as the accuracy on the data set.

[0076] d. Update: Update the mean and covariance matrix according to the selected individuals to better fit the search space;

[0077] e. Convergence: Repeat the above steps until the stopping condition is met. The stopping condition is set to 10 million iterations of optimization, and the model with the highest accuracy is selected as the final model.

[0078] 4. Use the Optuna framework for model hyperparameter optimization. Optuna is an open source framework for automated hyperparameter optimization. It is used to define the hyperparameter search space and use parallelization-based optimization algorithms (such as Tree-structuredParzen Estimator, Bayesian Optimization, etc.) to search for the best hyperparameter configuration.

[0079] The CMA-ES algorithm is combined with the model fusion method, and uses the model fusion method based on parameter space and data stream space to generate offspring. The offspring model is tested for accuracy on the test set, and the offspring with better accuracy is selected as the next generation parent. Some offspring are selected for mutation operations to increase population diversity, and the process is iterated until the stopping condition is met.

[0080] In the description of this article, it should be noted that relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0081] While the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that many changes, modifications, substitutions and variations can be made to the embodiments without departing from the principles and spirit of the invention.

Claims

1. A multi-modal large model fusion training method based on a biological evolution algorithm, characterized in that: The multi-modal large model fusion training method based on the biological evolution algorithm includes: Initialize the population: According to the specific problem, select the initial model, initialize the model parameters, and form the initial model; Model crossover: Use the previous generation model for crossover to generate a child model; Model selection: Evaluate the fitness of offspring models and select offspring models with better performance; Model mutation: select some offspring models and perform mutation operations; Iterative optimization: Determine whether the evolution process meets the termination conditions. If not, continue the above process.

2. The multi-modal large model fusion training method based on biological evolution algorithm according to claim 1 is characterized in that: The population initialization operation includes: first, an initial model population needs to be created. These models can be neural networks with different parameters and structures. The diversity of the initialization population is crucial to the subsequent evolutionary process because it determines the breadth and depth of the search space.

3. The multi-modal large model fusion training method based on biological evolution algorithm according to claim 2 is characterized in that: Model crossover can generate a new model by combining the characteristics of two or more excellent models. For example, methods such as structural crossover or parameter crossover of neural networks can be used.

4. The multi-modal large model fusion training method based on biological evolution algorithm according to claim 3 is characterized by: Model selection is performed based on fitness, which is an important indicator for measuring individual survival and reproduction capabilities. During the evolution of AI models, fitness is usually determined by evaluating the performance of the model on a specific task. For example, accuracy, recall, F1 score, and other indicators can be used to measure model performance.

5. The multi-modal large model fusion training method based on biological evolution algorithm according to claim 4 is characterized in that: The mutation operation introduces new features, increases model diversity, expands the model search space, and prevents falling into local optimal solutions. Mutation methods include randomly changing certain parameters or structures of the model, and using Gaussian functions to introduce noise.

6. The multi-modal large model fusion training method based on biological evolution algorithm according to claim 5 is characterized by: By repeating the above process (i.e., model crossover, selection, and mutation operations), the model performance can be gradually optimized, and the maximum iteration rounds or fitness threshold can be set to determine whether to stop the search until the possible model space is fully explored and the optimal solution is found.

7. A system for multi-modal large model fusion training based on biological evolution algorithm according to any one of claims 1 to 6, characterized in that: It includes a model selection system, a model generation system, a model evaluation system, a model processing system and a model determination system. The model selection system, the model generation system, the model evaluation system, the model processing system and the model determination system are connected in sequence through electrical signals, and the model determination system is connected to the model generation system through electrical signals.

Images