Federated Few-Shot Learning Method Based On Metric-Based Meta-Learning

The vigorous development of artificial intelligence depends on a large amount of labeled data.However,with the introduction of data security laws and regulations,the collection and acquisition of data has been greatly restricted,which will affect the development of AI-related applications.We can solve this problem from two different ways-data and model.From the perspective of data,federated learning method can be used to train the shared machine learning model with the fact that data does not leave the local.From the perspective of model,few-shot learning can improve the generalization level of the model by a small amount of training data.Federated few-shot method combines the advantages of federated learning and fewshot learning.It makes use of the privacy protection of federated framework to train the independent few-shot tasks distributed on different clients,aiming to get a few-shot model that can quickly adapt to new tasks.However,in the process of specific design and implementation,there are the following challenges.First of all,in order to make the low-power clients participate in federated training,the current federated few-shot method generally uses the metric-based meta-learning algorithm with low complexity.However,due to the simple structure,it also leads to the lack of adaptability of metric meta-learning to new tasks,which affects the prediction accuracy.Secondly,the problem of data heterogeneity will lead to the deviation between the local optimization objective of each participant and the global optimization objective,which makes the convergence of the federated few-shot algorithm difficult and affects the generalization effect of the model.Finally,in the federated learning scenario,the local few-shot tasks are more difficult because the data categories on a single participant are more similar.The difficult few-shot task will increase the difficulty of model learning and affect the generalization effect of the model.To cope with the above challenges,three federated few-shot algorithms are proposed in this paper.The main contributions are as follows:1.In order to solve the problem that traditional metric-based meta-learning methods have insufficient ability to fit new tasks,this paper proposes a new federated metric-based meta-learning few-shot algorithm based on deep mapping.The metric-based meta-learning methods commonly used in the current federated few-shot method have insufficient ability to fit new tasks.In order to solve the problem,new algorithm use siamese network to construct complex contrast relationship for label data.And we use a new deep mapping layer to make the model adaptively transform the sample metric space of the current task,so that the model can identify the differences between different categories in the case of a small amount of data.In this paper,experiments show that the algorithm improves the prediction accuracy and verifies the effectiveness of the algorithm.2.In order to solve the problem that the federated model is difficult to converge caused by data heterogeneity,this paper proposes a federated personalized few-shot algorithm.The new algorithm adjusts the local model structure of the participants,and divides the presentation layer into the basic presentation layer for learning global knowledge and the deep presentation layer for learning similar tasks.Through the ”global and personalized” federated updating method,each local model can learn global knowledge while paying attention to the characteristics of its own task.Compared with the traditional global federated model,the personalized federated model has stronger ability to resist the interference from heterogeneous data.In this paper,experiments verifies that the new algorithm can alleviate the convergence difficulty caused by heterogeneous data.3.In order to solve the problem that few-shot tasks are often more difficult in federated scenarios,this paper proposes a new federated personalized data augmentation algorithm.The personalized data augmentation algorithm proposed in this paper,taking into account the differences between categories,applies the data augmentation network to the personalized federated model to derive more data with category characteristics for each participant.so that the few-shot model can obtain more samples,reduce the difficulty of few-shot tasks,and improve the learning and recognition ability of the model for different types of samples.Finally,the effectiveness of the personalized data augmentation algorithm is verified by experiments.Generally speaking,this paper studies and improves the shortcomings of the current federated few-shot learning method.We propose a new federated metric-based meta-learning algorithm based on deep mapping,a new federated personalized fewshot algorithm and a new federated personalized data augmentation algorithm.Finally,a large number of experiments are carried out to verify the feasibility and effectiveness of the proposed methods.The research results have certain academic value and application value in the field of federated few-Shot learning.