Meta-Learning Fast Adaptation Method, Apparatus, Device, Medium and Program Product

Abstract

The present application relates to a meta - learning fast adaptation method, apparatus, device, medium and program product, and relates to the field of artificial intelligence technology. The method includes: by analyzing task characteristics, identifying the similarity between tasks, and accordingly performing task grouping and specific initialization of model parameters to achieve fast adaptation to new tasks. While maintaining the generalization ability of the model, the present invention significantly improves the adaptation speed and learning efficiency. In addition, the present invention not only provides an in - depth understanding of the fast adaptation ability of meta - learning algorithms in theory, but also demonstrates its high efficiency and reliability in practical applications. Through the present invention, the training of multi - modal meta - learning models becomes faster and more accurate, greatly expanding the application scope and practical value of meta - learning technology.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and particularly to a meta-learning fast adaptation method, device, equipment, medium, and program product. Background Art

[0002] In the field of artificial intelligence, especially with the rapid development of machine learning and deep learning, the concept of meta-learning, that is, "learning to learn", has become a research hotspot. The core goal of meta-learning is to endow the model with the ability to quickly adapt to new tasks through a small number of samples, which is of extremely important significance in the small-sample learning scenario. However, although existing meta-learning methods have achieved remarkable achievements in a single task distribution, when facing multi-modal task distributions, they still face performance bottlenecks and it is difficult to achieve fast and accurate adaptation.

[0003] Existing meta-learning methods mainly focus on accelerating the adaptation process of new tasks through gradient descent. These methods learn an excellent model initialization parameter, so that when encountering a new task, the model can quickly adjust the parameter through a small number of gradient update steps to meet the needs of the new task. However, when dealing with multi-modal task distributions, these methods often have difficulty finding a unified initialization strategy applicable to all tasks. In addition, existing methods are insufficient in considering the similarity or difference between tasks, resulting in weak adaptability to new tasks. Existing research does not have a deep enough understanding of the generalization ability of meta-learning algorithms, which limits the effectiveness and reliability of the algorithms in practical applications. Although existing research has tried to explain the generalization ability of meta-learning algorithms through theoretical analysis, these analyses are usually limited to specific algorithm variants or in the context of non-convex optimization problems and are difficult to provide a comprehensive explanation. Summary of the Invention

[0004] Based on this, it is necessary to provide a meta-learning fast adaptation method, device, equipment, medium, and program product for the above technical problems, which can solve the problem of weak adaptability of traditional meta-learning methods in multi-modal task distributions, especially in the small-sample learning scenario.

[0005] In the first aspect, this application provides a meta-learning fast adaptation method. The method includes:

[0006] Conduct feature analysis on task-related data to identify and quantify the similarity between tasks; the task-related data includes at least one of text, pictures, and videos;

[0007] Perform clustering analysis on tasks based on the similarity to generate several task groups; for each task group, obtain the corresponding model parameters, and fuse the model parameters shared across groups by different task groups to obtain the first model parameters;

[0008] Find a task group with similarity matching for the target task, and extract the first model parameters corresponding to the successfully matched task group as the initial model parameters of the target task; perform parameter optimization on the initial model parameters to obtain the target model parameters corresponding to the target task.

[0009] In one embodiment, an online learning mechanism is introduced when performing clustering analysis on tasks.

[0010] In one embodiment, for each task group, obtain the corresponding model parameters, and fuse the model parameters shared across different task groups. Obtaining the first model parameters includes:

[0011] Obtain the model parameters corresponding to each task group, and obtain the model parameters that can be shared across task groups;

[0012] For each task group, perform weighted summation on the model parameters corresponding to the task group and the model parameters that can be shared across the task group to obtain the first model parameters.

[0013] In one embodiment, finding a task group with similarity matching for the target task includes:

[0014] Perform feature analysis on the target task and the task group, and determine the task group that matches the target task according to the similarity between the target task and each task group.

[0015] In one embodiment, performing parameter optimization on the initial model parameters to obtain the target model parameters corresponding to the target task includes:

[0016] Use the gradient descent method to perform parameter iterative optimization on the initial model parameters, evaluate the fitness of the model parameters after each gradient descent iteration, and stop the iteration when the fitness exceeds the threshold, and output the model parameters at the end of the iteration as the target model parameters.

[0017] In one embodiment, after obtaining the initial model parameters, the method further includes:

[0018] Introduce the first model parameters corresponding to the task groups in the task group that do not match the target task into the initial model parameters to obtain the second model parameters;

[0019] Perform parameter optimization on the second model parameters corresponding to the initial model parameters to obtain the target model parameters.

[0020] In a second aspect, the present application also provides a meta-learning fast adaptation device. The device includes:

[0021] A task feature analysis module, configured to perform feature analysis on task-related data, identify and quantify the similarity between tasks; the task-related data includes at least one of text, pictures, and videos;

[0022] The task grouping and parameter modulation module is used to perform clustering analysis on tasks based on similarity to generate several task groups; for each task group, obtain the corresponding model parameters, and fuse the model parameters shared across different task groups to obtain the first model parameters.

[0023] The fast adaptation strategy module is used to find a task group with a similarity match for the target task, extract the first model parameters corresponding to the successfully matched task group as the initial model parameters of the target task; perform parameter optimization on the initial model parameters to obtain the target model parameters corresponding to the target task.

[0024] In a third aspect, the present application also provides a computer device. The computer device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps in the above-mentioned meta-learning fast adaptation method are implemented.

[0025] In a fourth aspect, the present application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in the above-mentioned meta-learning fast adaptation method are implemented.

[0026] In a fifth aspect, the present application also provides a computer program product. The computer program product includes a computer program, and when the computer program is executed by a processor, the steps in the above-mentioned meta-learning fast adaptation method are implemented.

[0027] The above-mentioned meta-learning fast adaptation method, device, equipment, medium and program product analyze task characteristics, identify the similarity between tasks, and perform task grouping and specific initialization of model parameters accordingly to achieve fast adaptation to new tasks. Compared with traditional methods, the present invention significantly improves the adaptation speed and learning efficiency while maintaining the model generalization ability. The present invention not only provides an in-depth understanding of the fast adaptation ability of meta-learning algorithms theoretically, but also demonstrates its high efficiency and reliability in practical applications, such as in the fields of autonomous driving, medical diagnosis, intelligent robots, etc. Through the present invention, the training of multi-modal meta-learning models becomes faster and more accurate, greatly expanding the application scope and practical value of meta-learning technology. Description of the Drawings

[0028] Figure 1 It is a schematic flowchart of the meta-learning fast adaptation method in an embodiment;

[0029] Figure 2 It is a structural block diagram of the meta-learning fast adaptation device in an embodiment;

[0030] Figure 3 It is a schematic structural diagram of the task feature analysis module in an embodiment;

[0031] Figure 4Schematic diagram of the task grouping and parameter modulation module in an embodiment;

[0032] Figure 5 Schematic diagram of the fast adaptation strategy module in an embodiment;

[0033] Figure 6 Flowchart of accident detection and response in an intelligent transportation system in an embodiment;

[0034] Figure 7 Flowchart of medical image diagnosis based on multi-modal data in an embodiment. Detailed implementation manners

[0035] In order to make the objectives, technical solutions and advantages of the present application more clear and understandable, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0036] The embodiment of the present application provides a meta-learning fast adaptation method, which is multi-modal meta-learning based on task similarity perception, as Figure 1 shown, the method includes the following steps:

[0037] Step 102: Perform feature analysis on task-related data to identify and quantify the similarity between tasks. The task-related data includes at least one of text, pictures, and videos;

[0038] This step is responsible for deeply analyzing task data to identify and quantify the similarity between tasks. By using existing data mining techniques, the internal features of task data are captured. Based on the task features extracted from the task-related data, the similarity between tasks is quantitatively analyzed. This process is crucial for subsequent work because it provides the necessary similarity measurement information to quickly adapt to new tasks.

[0039] Among them, performing feature analysis on task-related data requires first extracting features from the task-related data. Feature extraction is to convert the original data into a numerical vector that can characterize the task characteristics.

[0040] Then, a similarity measurement method is used to quantify the similarity based on the extracted task features, that is, to measure the proximity of two tasks in the feature space. Cosine similarity can be used as an index for similarity measurement, so as to realize the quantification of the similarity between tasks.

[0041] In an embodiment, before performing task feature analysis, the data can be preprocessed first, including steps such as normalization, denoising, and feature extraction, to improve the effectiveness of the data.

[0042] Step 104: Perform clustering analysis on the tasks based on similarity to generate several task groups; for each task group, obtain the corresponding model parameters, and fuse the model parameters shared across different task groups to obtain the first model parameters.

[0043] Based on the results of task feature analysis, the tasks are divided into different groups, and a specific set of model parameters is learned for each group. This process utilizes the similarity between tasks, enabling the model to quickly adapt to similar tasks and thus improving the learning efficiency.

[0044] Specifically, in this embodiment, clustering analysis is used to group the tasks according to the similarity between tasks. Tools such as the k-means clustering algorithm can be used for clustering analysis.

[0045] To enhance the generalization ability of the model, this embodiment introduces a cross-group parameter sharing and coordination mechanism. By allowing a certain degree of sharing of model parameters between different groups, the model's adaptability to new tasks can be improved. The task groups are fused with the model parameters that can be shared across different groups to obtain the first model parameters.

[0046] It should be noted that the model parameters in this step are the model parameters after parameter optimization modulation through the objective function to adapt to each task in the corresponding task group.

[0047] In other embodiments, each task group corresponds to a meta-learner, which can quickly adapt to new tasks within the group. When constructing the meta-learner, model parameter optimization modulation is performed through the objective function, and a regularization term is introduced during the optimization modulation process to control the complexity of the model and prevent overfitting.

[0048] Step 106: Search for a task group with a similarity match for the target task, extract the first model parameters corresponding to the successfully matched task group as the initial model parameters for the target task; perform parameter optimization on the initial model parameters to obtain the target model parameters corresponding to the target task.

[0049] When the target task arrives, this step is responsible for quickly matching the task group to which it belongs and initializing it with the first model parameters of that task group. Subsequently, the first model parameters are iteratively adjusted through an optimization algorithm to adapt to the target task. This strategy not only speeds up the model's adaptation speed but also enhances the model's generalization ability on new tasks.

[0050] Since the first model parameters have been well adapted to each task in the task group, and then the first model parameters of the task group are migrated to the model of the target task. Due to the similarity matching between the target task and the task group, the target task has a good adaptation basis for the first model parameters. At this time, based on the first parameter model, the optimization of the model parameters can be completed within a small number of optimization iterations, thereby accelerating the adaptation speed of the model to the target task as a new task, and at the same time enhancing the generalization ability of the model on the new task.

[0051] In one embodiment, an online learning mechanism is introduced when performing clustering analysis on tasks in step 104.

[0052] During the task feature analysis process, it is necessary to be able to update dynamically to adapt to the arrival of new tasks. For this reason, an online learning mechanism is introduced, so that the clustering center and feature extraction can be adaptively adjusted according to new data, thereby continuously improving the model accuracy.

[0053] In one embodiment, for each task group, the corresponding model parameters are obtained, and the model parameters shared across different task groups are fused. Obtaining the first model parameters includes: obtaining the model parameters corresponding to each task group, and obtaining the model parameters that can be shared across task groups; for each task group, performing a weighted sum of the model parameters corresponding to the task group and the model parameters that can be shared across task groups with the task group to obtain the first model parameters.

[0054] Learn a set of specific model parameters θ for each task group m , the first model parameter θ m ' can be expressed as:

[0055]

[0056] where θ shared is the model parameter that can be shared across groups, and β is the weight of the model parameter that can be shared across groups.

[0057] In one embodiment, finding a task group with similarity matching for the target task includes: performing feature analysis on the target task and the task group, and determining the task group that matches the target task according to the similarity between the target task and each task group.

[0058] First, determine the feature vector of the target task through task feature analysis technology, then use the similarity measurement method to determine the similarity between the target task and the known task group, and match the task group with the highest similarity.

[0059] In one embodiment, parameter optimization is performed on the initial model parameters to obtain the target model parameters corresponding to the target task, including: performing parameter iterative optimization on the initial model parameters using the gradient descent method, evaluating the fitness of the model parameters after each gradient descent iteration, stopping the iteration when the fitness exceeds the threshold, and outputting the model parameters at the end of the iteration as the target model parameters.

[0060] Gradient descent is an optimization algorithm used to find the minimum of a loss function in machine learning and deep learning. Its basic idea is to gradually adjust the model parameters to continuously reduce the value of the loss function until a local or global optimum is reached.

[0061] The fitness function is usually related to the objective function of the optimization problem. For example, if the goal is to maximize a certain metric, the fitness function can be the value of that metric; if the goal is to minimize a certain metric, the fitness function can be the negative value or reciprocal of the objective function. Set the fitness function according to the actual situation to evaluate and guide the optimization direction of the model parameters.

[0062] In one embodiment, after obtaining the initial model parameters, the method further includes: introducing the first model parameters corresponding to the task groups in the task group that do not match the target task into the initial model parameters to obtain the second model parameters; performing parameter optimization on the second model parameters corresponding to the initial model parameters to obtain the target model parameters.

[0063] To enhance the adaptability of the target task and achieve cross-group knowledge transfer, it can be achieved by introducing the first model parameters of other task groups into the initial model parameters. At the same time, to emphasize the dominant position of the initial model parameters and the influence degree of other task groups on the task group corresponding to the target task, a weight system is introduced to optimize the composition of the second model parameters. Assume the second model parameter is θ init , and its formula is expressed as:

[0064]

[0065] where, θ m is the initial model parameter, θ j is the first model parameter of the j-th task group, γ j is the weight of the first model parameter of the j-th task group, and .

[0066] In one embodiment, the multi-modal meta-learning fast adaptation method based on task similarity perception includes the following steps:

[0067] S1. Task feature analysis

[0068] Task feature analysis is one of the core steps of the present invention, and its purpose is to identify the similarity between tasks by quantitatively analyzing task data. S1 includes:

[0069] S1.1. Data preprocessing and feature extraction

[0070] Before performing task feature analysis, it is first necessary to preprocess the data, including steps such as normalization, denoising, and feature extraction.

[0071] When performing feature extraction, let the original data set be , where N is the number of samples. The goal of feature extraction is to find the mapping to convert the data into a feature vector .

[0072] S1.2. Similarity measurement

[0073] In this embodiment, cosine similarity is used as the index for similarity measurement. For two feature vectors f i and f j , the cosine similarity is defined as:

[0074]

[0075] where · represents the dot product of vectors, represents the norm of the vector.

[0076] S1.3. Cluster analysis

[0077] In this embodiment, the k-means clustering algorithm is used to complete the cluster analysis. Let be the k cluster centers. Each cluster center c m is determined by the average value of the feature vectors of its members:

[0078]

[0079] where G m is the set of elements belonging to the m-th cluster, and m takes integer values from 1 to k.

[0080] S1.4. Dynamic update of task features

[0081] The task feature analysis module needs to be able to update dynamically to adapt to the arrival of new tasks. For this purpose, an online learning mechanism is introduced so that the cluster centers and the feature extraction model can be adjusted according to the new task data. Let be the learning rate. The update formula for the cluster center is:

[0082]

[0083] where t represents the time step, is the set of tasks belonging to cluster m at time step t.

[0084] S1.5, Optimization and Regularization of Feature Analysis

[0085] To prevent overfitting during feature extraction and improve the generalization ability of the model, an L2 regularization term is introduced to optimize the feature extraction model. The optimization objective function can be expressed as:

[0086]

[0087] where is the task loss function, θ is the model parameter, and λ opt is the regularization coefficient, and y i is the model output.

[0088] The task feature analysis step takes into account the dynamics and diversity in the environment, can accurately identify the similarities between tasks, and provides the necessary information for task grouping and parameter modulation.

[0089] S2, Task Grouping and Parameter Modulation

[0090] Task grouping and parameter modulation aims to effectively group tasks based on the results of task feature analysis and learn specific model parameters for each group of tasks. S2 includes:

[0091] S2.1, Task Grouping Strategy

[0092] The task grouping strategy divides tasks into different groups according to the similarity measurement results between tasks. This strategy can be based on the output of cluster analysis, such as each cluster center c m representing a task group. Task T i is assigned to the group represented by the cluster center c m to which it is most similar:

[0093]

[0094] where G m is the task group formed around the cluster center c m .

[0095] S2.2, Parameter Modulation Method

[0096] For each task group G m , a set of specific model parameters θ m is learned, and these parameters will subsequently be used as the initial model parameters for new tasks within the group. Parameter modulation can be achieved by optimizing the following objective function:

[0097]

[0098] where are the model parameters θ mOn task T i The loss function on it.

[0099] S2.3, Construction of the meta-learner

[0100] Each task group G m has a corresponding meta-learner that can quickly adapt to new tasks within the group. The construction of the meta-learner is based on the following optimization problem:

[0101]

[0102] where, is the regularization term used to control the complexity of the model and prevent overfitting, and α is the regularization coefficient.

[0103] S2.4, Cross-group parameter sharing and coordination

[0104] To enhance the generalization ability of the model, this embodiment also implements a cross-group parameter sharing and coordination mechanism. By allowing a certain degree of parameter sharing between different groups, the adaptability of the model to new tasks can be improved. Specifically, the first model parameters obtained after being processed by the cross-group parameter sharing and coordination mechanism are:

[0105]

[0106] where, θ m is the model parameter, θ shared is the model parameter that can be shared across groups, and β is the weight of the model parameter that can be shared across groups.

[0107] S3, Fast adaptation strategy

[0108] This step aims to achieve fast adaptation to new tasks. Using the results of the task grouping and parameter modulation module, the model is quickly adjusted to adapt to new tasks through a small number of gradient descent steps. S3 includes:

[0109] S3.1, Task identification and grouping

[0110] When the target task T new arrives, first determine its feature vector f new through task feature analysis technology. Then, use the similarity measurement method to determine the similarity between T new and the known task groups, and assign it to the most suitable group G m : [[ID=�6]]

[0111]

[0112] where, m best is the index of the task group that is most similar to the target task.

[0113] S3.2. Parameter Initialization

[0114] Once the target task is assigned to a specific task group, use the first model parameter θ m ' of this group as the initial model parameter θ init to adapt to the target task:

[0115] θ init = θ m ' (10)

[0116] S3.3. Gradient Descent Adaptation

[0117] Next, apply the gradient descent algorithm to the dataset D related to the target task new to adjust the initial model parameter θ init and find the optimized target model parameter θ adapted :

[0118]

[0119] where, is the learning rate, is the loss function on the target task dataset, t is the number of iterations of gradient descent, θ t+1 is the model parameter output at the (t + 1)-th iteration, θ t is the model parameter output at the t-th iteration.

[0120] S3.4. Fitness Evaluation

[0121] After each gradient descent iteration, evaluate the fitness A(θ t ) of the model parameter to determine whether further iterations are needed or a satisfactory adaptation level has been reached. A(θ t ) is expressed as:

[0122]

[0123] If the fitness exceeds the preset threshold, stop the iteration.

[0124] S3.5. Cross-Group Knowledge Transfer

[0125] To enhance the adaptability of the target task and achieve cross-group knowledge transfer. This can be achieved by introducing the model parameters of other task groups into the initial model parameter θ init , and the formula is expressed as:

[0126]

[0127] where, γ j is the weight of the j-th group of parameters, and , The second model parameters after being fused with the first model parameters of other task groups. Based on the second model parameters, the model parameters are further optimized to obtain the final target model parameters.

[0128] In summary, the multi-modal meta-learning fast adaptation method based on task similarity perception proposed by the present invention has achieved remarkable beneficial effects in solving the fast adaptation problem in the multi-modal task distribution. Compared with the prior art, the present invention has realized the improvement of performance and the effective utilization of resources through the following innovative points:

[0129] 1. Enhanced model adaptability: The present invention deeply explores the internal features of task data through task feature analysis, and realizes the accurate identification of the similarity between tasks. This similarity-based task grouping method enables the model to quickly find a similar task group when facing a new task, and initialize with the model parameters of this group, significantly improving the adaptation speed of the model to new tasks. On the CUB Birds dataset, compared with the existing method MAML (Model-Agnostic Meta-Learning), the method of the present invention has achieved a performance improvement of more than 3% in classification accuracy.

[0130] 2. Improved generalization ability: Through task grouping and parameter modulation, the present invention learns specific model parameters related to the task group. These parameters not only accelerate the adaptation speed of the model, but also enhance the generalization ability of the model on new tasks. In the experiment of cross-domain task adaptation, compared with the baseline method ANIL (Almost No Inner Loop) on the miniImageNet dataset, the method of the present invention shows a higher generalization accuracy, achieving an approximately 2% gain in 10-shot accuracy.

[0131] 3. Optimized resource allocation: The fast adaptation strategy of the present invention quickly adjusts the model through a small number of gradient descent steps, reducing the demand for computing resources. Compared with traditional meta-learning methods, the present invention reduces a certain amount of computation while maintaining the same model performance, which is particularly beneficial for resource-constrained environments.

[0132] 4. Improved data utilization efficiency: The present invention adopts advanced data mining techniques in the task feature analysis stage, making the features extracted from the same amount of data more abundant and representative, thereby improving the data utilization efficiency.

[0133] 5. Advantages of Cross-modal Learning: The present invention is specifically designed for multi-modal tasks and can effectively process information from different data sources and modalities. Through modal feature fusion and cross-modal correlation learning, the method of the present invention can make full use of the complementary information between different modalities and improve the model's understanding and adaptation ability for multi-modal tasks.

[0134] 6. Dynamic Update and Adaptive Adjustment: The dynamic update of task features in the present invention allows the system to continuously adjust and optimize the clustering centers and feature extraction models according to the arrival of new tasks, maintaining the continuous update and adaptive ability of the model. This dynamic update mechanism enables the method of the present invention to adapt to the change of task distribution over time and maintain long-term learning effects.

[0135] 7. Facilitating Knowledge Transfer: The fast adaptation strategy of the present invention is not only limited to the fast learning of a single task but also enables cross-task knowledge transfer. This means that the knowledge learned in one task can be quickly transferred to another related task, improving the overall efficiency of the learning process.

[0136] It should be understood that although the steps in the flowcharts involved in the above-described embodiments are shown in sequence according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless there is a clear description in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above-described embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily executed at the same moment but can be executed at different moments, and the execution order of these steps or stages is not necessarily sequential but can be executed alternately or in turn with at least a part of other steps or steps or stages in other steps.

[0137] Based on the same inventive concept, the embodiments of the present application also provide a meta-learning fast adaptation device for implementing the above-mentioned meta-learning fast adaptation method. The solution provided by this device for solving problems is similar to the solution described in the above method. Therefore, the specific limitations in one or more embodiments of the meta-learning fast adaptation device provided below can refer to the limitations on the meta-learning fast adaptation method in the above text, and will not be repeated here.

[0138] In one embodiment, as Figure 2 shown, a meta-learning fast adaptation device is provided, which is multi-modal meta-learning based on task similarity perception. The device includes: a task feature analysis module, a task grouping and parameter modulation module, and a fast adaptation strategy module.

[0139] Among them, the task feature analysis module is used to perform feature analysis on task-related data, identify and quantify the similarity between tasks; the task-related data includes at least one of text, pictures, and videos. This module is responsible for deeply analyzing task data, identifying and quantifying the similarity between tasks. Through data preprocessing, feature extraction, similarity measurement, clustering analysis, and dynamic update of task features, this module can capture the intrinsic features of task data.

[0140] The task grouping and parameter modulation module is used to perform clustering analysis on tasks based on similarity to generate several task groups; for each task group, obtain the corresponding model parameters, and fuse the model parameters shared across different task groups to obtain the first model parameters; based on the results of task feature analysis, this module divides tasks into different groups and learns a specific set of model parameters for each group. This process utilizes the similarity between tasks and improves learning efficiency through the construction of a meta-learner, cross-group parameter sharing and coordination, and dynamic update of parameter modulation.

[0141] The fast adaptation strategy module is used to find a task group with a similarity match for the target task, extract the first model parameters corresponding to the successfully matched task group as the initial model parameters of the target task; perform parameter optimization on the initial model parameters to obtain the target model parameters corresponding to the target task; when the target task arrives, this module is responsible for quickly identifying the group to which it belongs and initializing it with the model parameters of that group. Subsequently, through gradient descent adaptation, fitness evaluation, and cross-group knowledge transfer, the model is quickly adjusted to adapt to the target task.

[0142] Figure 2 Among them, the data flow and information transfer are represented by arrows for each part. The output of the task feature analysis module is passed to the task grouping and parameter modulation module through an arrow, indicating the transfer of the task feature analysis results; the output of the task grouping and parameter modulation module is passed to the fast adaptation strategy module through an arrow, indicating the transfer of the task grouping and parameter modulation results; the fast adaptation strategy module feeds back the adapted model parameters to the task feature analysis module and the task grouping and parameter modulation module according to the fast adaptation strategy, indicating the transfer of the model adaptation results.

[0143] As Figure 3 shown, the task feature analysis module mainly includes:

[0144] The data preprocessing and feature extraction module is used to be responsible for preprocessing operations such as normalizing and denoising the original data, and extracting key information that can characterize the task characteristics from it;

[0145] The similarity measurement module is used to measure the similarity of the feature vectors of different tasks using methods such as cosine similarity to provide a basis for task grouping;

[0146] The clustering analysis module is used to divide tasks into different groups by using clustering algorithms such as k-means according to the similarity measurement results, so as to achieve the effective organization of tasks;

[0147] The task feature dynamic update module is used to adopt an online learning mechanism to dynamically update the clustering center and feature extraction model according to the features of the target task, so as to adapt to the changes in task distribution.

[0148] Figure 3 In it, the output of the data preprocessing and feature extraction module undergoes feature similarity measurement by the similarity measurement module, thereby triggering the clustering analysis of the clustering analysis module. The clustering results of the clustering analysis module guide the task feature dynamic update module to perform dynamic update of task features. The update results of the task feature dynamic update module are fed back to the clustering analysis module to optimize the accuracy of similarity measurement and clustering analysis. Finally, the outputs of all modules are aggregated to the end point, indicating the completion of the task feature analysis process.

[0149] As Figure 4 shown, the task grouping and parameter modulation module mainly includes:

[0150] The task grouping strategy module is used to be responsible for effectively dividing tasks into different groups according to the similarity measurement results between tasks;

[0151] The parameter modulation method module is used to learn a set of specific model parameters for each task group, and these parameters will be used as the initial model parameters of the target tasks within the group;

[0152] The meta-learner construction module is used to construct the meta-learner for each task group, which can quickly adapt to the target tasks within the group;

[0153] The cross-group parameter sharing and coordination module is used to implement the cross-group parameter sharing and coordination mechanism, allowing a certain degree of parameter sharing between different groups to improve the generalization ability of the model;

[0154] The parameter modulation dynamic update module is used to be responsible for dynamically updating the parameter modulation module to maintain its effectiveness and adapt to new task data;

[0155] The task combination and splitting decision module is used to make decisions on merging or splitting task groups according to the changes in task distribution to adapt to the dynamic similarity changes between tasks.

[0156] The task grouping result of the task grouping strategy module guides the parameter modulation method module to perform parameter modulation, thereby triggering the construction of the meta-learner by the meta-learner construction module. The cross-group parameter sharing and coordination mechanism of the cross-group parameter sharing and coordination module supports the execution of the parameter modulation method module and the meta-learner construction module. The dynamic update of the parameter modulation dynamic update module is fed back to the parameter modulation method module and the meta-learner construction module to optimize the performance of parameter modulation and the meta-learner. The merge and split decision of the task combination and split decision module adjusts the task grouping of the task grouping strategy module to adapt to the change of task distribution. Finally, the outputs of all modules are aggregated to the end point, indicating the completion of the task grouping and parameter modulation process.

[0157] As Figure 5 shown, the fast adaptation strategy module mainly includes:

[0158] The target task identification and grouping module is responsible for receiving the target task and identifying its features, and then allocating it to the most suitable group according to the similarity between the features and the known task groups.

[0159] The parameter initialization module is used to use the model parameters of the group where the target task is located as the initial model parameters to provide a starting point for the fast adaptation of the target task.

[0160] The gradient descent adaptation module is used to apply the gradient descent algorithm to the dataset of the target task and quickly adjust the model parameters through a small number of iterations to adapt to the target task.

[0161] The fitness evaluation module is used to evaluate the model fitness after each iteration to determine whether further iterations are needed or a satisfactory adaptation level has been reached.

[0162] The cross-group knowledge transfer module is used to achieve cross-group knowledge transfer and enhance the adaptability of the target task by introducing the parameter information of other groups into the initial model parameters.

[0163] The adaptation completion decision module is used to make a decision on whether to complete the adaptation according to the fitness evaluation result. If the fitness reaches the preset threshold, the adaptation process ends.

[0164] The arrival of the target task triggers the recognition and grouping of target tasks by the target task recognition and grouping module, and then the parameter initialization module performs parameter initialization based on the grouping result. The gradient descent adaptation module receives the initial model parameters from the parameter initialization module and performs gradient descent adaptation. Subsequently, the fitness evaluation module evaluates the fitness. If the evaluation result of the fitness evaluation module indicates that further iteration is required, it returns to the gradient descent adaptation module to continue adaptation; if the fitness is satisfactory, it triggers the adaptation completion decision of the adaptation completion decision module. The cross-group knowledge transfer module realizes cross-group knowledge transfer between the parameter initialization module and the gradient descent adaptation module to enhance adaptability. Finally, the outputs of all modules are aggregated to the end point, indicating the completion of the fast adaptation strategy process.

[0165] This embodiment proposes a multi-modal meta-learning fast adaptation device based on task similarity perception. The task feature analysis module uses advanced data mining techniques to deeply analyze task data and identify and quantify the similarity between tasks. The task grouping and parameter modulation module groups tasks according to the results of task feature analysis and learns a specific set of model parameters for each group to achieve fast adaptation. When a new task arrives, the fast adaptation strategy module quickly identifies the group to which it belongs, initializes with the model parameters of that group, and then quickly adjusts the model through a small number of gradient descent steps to adapt to the new task. This embodiment not only improves the adaptation speed of the meta-learning algorithm on new tasks but also enhances its generalization ability, especially its effectiveness in dealing with multi-modal task distributions, bringing important research value and broad application prospects to the field of artificial intelligence.

[0166] Each module in the above multi-modal meta-learning fast adaptation device based on task similarity perception can be implemented in whole or in part by software, hardware, and their combination. The above modules can be embedded in the processor of a computer device in hardware form or be independent of it, or can be stored in the memory of a computer device in software form for the processor to call and execute the operations corresponding to the above respective modules.

[0167] In other embodiments, an application scenario of the method disclosed in the present invention is also proposed. As Figure 6 shown, the method of the present invention is applied to an intelligent transportation system to achieve fast detection and effective response to traffic accidents. The specific implementation process is as follows:

[0168] S1. Multi-source data collection: Collect traffic data through multi-source devices such as traffic monitoring cameras, vehicle sensors, and weather stations.

[0169] S2. Data fusion and feature extraction: Fusion data from different sources and extract features helpful for accident detection, such as vehicle speed, traffic flow, weather conditions, etc.

[0170] S3. Accident type similarity analysis: Use the similarity measurement method of the present invention to perform similarity analysis and grouping on different types of traffic accidents.

[0171] S4. Accident response strategy initialization: Design response strategies for each accident type and initialize the strategy parameters using the parameter modulation method of the present invention.

[0172] S5. Real-time accident detection and rapid adaptation: Monitor the traffic flow in real time. Once signs of an accident are detected, quickly identify the accident type and adjust the corresponding response strategy.

[0173] S6. Cross-strategy knowledge transfer: During the process of adjusting the response strategy, achieve cross-accident type strategy knowledge transfer to improve the flexibility and effectiveness of the response.

[0174] S7. Accident handling and traffic guidance: According to the adjusted response strategy, conduct accident handling and traffic guidance to minimize the impact of the accident.

[0175] S8. System performance evaluation and optimization: Collect accident handling results and traffic data, evaluate the effectiveness of the response strategy, and perform iterative optimization of the system performance.

[0176] In other embodiments, another application scenario of the method disclosed in the present invention is also proposed. As Figure 7 shown, apply the method of the present invention to a medical image diagnosis system to improve the accuracy and efficiency of diagnosis. The specific implementation process is as follows:

[0177] S1. Multi-modal data collection: In a medical environment, collect multi-modal image data of patients through different imaging techniques (such as CT, MRI, X-ray, etc.).

[0178] S2. Data preprocessing and feature extraction: Preprocess the collected raw image data, including contrast enhancement, noise reduction, etc., and extract features helpful for diagnosis from it.

[0179] S3. Similarity measurement and task grouping: Use the task similarity measurement method of the present invention to group the image data of different patients according to feature similarity.

[0180] S4. Parameter modulation and model initialization: Learn specific model parameters for each group and initialize the diagnostic model with these parameters to adapt to the image data of different groups.

[0181] S5. Rapid adaptation and diagnostic model adjustment: When new patient image data is received, quickly identify the group it belongs to and perform rapid adaptation and adjustment using the corresponding model parameters.

[0182] S6. Cross-modal knowledge transfer: During the model adjustment process, cross-modal knowledge transfer is introduced to utilize the complementary information between different imaging techniques to enhance the performance of the diagnostic model.

[0183] S7. Diagnostic result generation and evaluation: Generate the final diagnostic result and evaluate it to ensure the accuracy of the diagnosis.

[0184] S8. Feedback and model iterative optimization: Use the diagnostic results and patient feedback for the iterative optimization of the model to continuously improve the performance of the diagnostic system.

[0185] The above two embodiments are only examples to illustrate the scenarios where the present invention can be applied, and do not limit the application scenarios of the present invention.

[0186] In one embodiment, a computer device is provided, including a memory and a processor. A computer program is stored in the memory, and when the processor executes the computer program, the steps in all the above method embodiments are implemented.

[0187] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in all the above method embodiments are implemented.

[0188] In one embodiment, a computer program product is provided, including a computer program. When the computer program is executed by a processor, the steps in all the above method embodiments are implemented.

[0189] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use, and processing of relevant data need to comply with the relevant laws, regulations, and standards of relevant countries and regions.

[0190] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, database, or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memories. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc. The databases involved in the embodiments provided in the present application can include at least one of relational databases and non-relational databases. Non-relational databases can include distributed databases based on blockchain, etc., without limitation. The processors involved in the embodiments provided in the present application can be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, data processing logics based on quantum computing, etc., without limitation.

[0191] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

[0192] The above-described embodiments only represent several implementation manners of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the patent scope of the present application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A multi-modal meta-learning fast adaptation method based on task similarity perception, characterized in that, The method includes: Performing feature analysis on task-related data to identify and quantify the similarity between the tasks; the task-related data includes text, pictures or videos; the tasks are target classification tasks; Performing clustering analysis on the tasks based on the similarity to generate several task groups; for each task group, obtaining corresponding model parameters, and fusing the model parameters shared across different task groups to obtain first model parameters; Searching for the task group with a similarity match for the target task, extracting the first model parameters corresponding to the successfully matched task group as the initial model parameters of the target task; performing parameter optimization on the initial model parameters to obtain the target model parameters corresponding to the target task; Wherein, the step of for each task group, obtaining corresponding model parameters, and fusing the model parameters shared across different task groups to obtain first model parameters includes: Obtaining the model parameters corresponding to each task group, and obtaining the model parameters that can be shared across task groups; For each task group, performing weighted summation on the model parameters corresponding to the task group and the model parameters that can be shared with the task group across groups to obtain the first model parameters.

2. The method according to claim 1, wherein: An online learning mechanism is introduced during the clustering analysis of the tasks.

3. The method according to claim 1, characterized in that The step of searching for the task group with a similarity match for the target task includes: Performing feature analysis based on the target task and the task groups, and determining the task group that matches the target task according to the similarity between the target task and each task group.

4. The method according to claim 1, wherein The step of performing parameter optimization on the initial model parameters to obtain the target model parameters corresponding to the target task includes: Using the gradient descent method to perform parameter iterative optimization on the initial model parameters, evaluating the fitness of the model parameters after each gradient descent iteration, and stopping the iteration when the fitness exceeds the threshold, and outputting the model parameters at the end of the iteration as the target model parameters.

5. The method according to claim 1, wherein After obtaining the initial model parameters, the method further includes: Introducing the first model parameters corresponding to the task groups that do not match the target task in the initial model parameters to obtain second model parameters; Performing parameter optimization on the second model parameters corresponding to the initial model parameters to obtain the target model parameters.

6. A multi-modal meta-learning fast adaptation device based on task similarity perception, characterized in that The device includes: A task feature analysis module, configured to perform feature analysis on task-related data to identify and quantify the similarity between the tasks; the task-related data includes text, pictures or videos; the tasks are target classification tasks; A task grouping and parameter modulation module, configured to perform clustering analysis on the tasks based on the similarity to generate several task groups; for each task group, obtaining corresponding model parameters, and fusing the model parameters shared across different task groups to obtain first model parameters; A fast adaptation strategy module is used to find the task groups with similarity matching for the target task, extract the first model parameters corresponding to the successfully matched task groups as the initial model parameters of the target task; perform parameter optimization on the initial model parameters to obtain the target model parameters corresponding to the target task; Among them, for each of the task groups, obtaining the corresponding model parameters and fusing the model parameters shared across different task groups to obtain the first model parameters includes: Obtaining the model parameters corresponding to each of the task groups and obtaining the model parameters that can be shared across different task groups; For each of the task groups, perform weighted summation on the model parameters corresponding to the task group and the model parameters that can be shared across different task groups with the task group to obtain the first model parameters.

7. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the method according to any one of claims 1 to 5.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the method according to any one of claims 1 to 5.

Images