Research On The Metric Few Shot Classification Based On Meta Learning Optimization

Today's deep learning algorithms mostly need to train neural networks through massive data sets.However,in many application scenarios,it is very inconvenient to obtain samples,which leads to the lack of sample data.The purpose of small sample learning task is to learn quickly from a single or a small number of training samples.Meta learning method,by summarizing and abstracting the general meta knowledge among different tasks,and then applying meta knowledge to new and never seen tasks quickly,so as to solve the problem of learning with few samples.Therefore,this paper takes the problem of image classification with a small number of samples as the research object,and studies how to use the meta learning method to quickly learn and summarize from a small number of image samples.The main contents of this paper are as follows:(1)The meta relation network is proposed to solve the problem of few shot classification.The model agnostic meta learning algorithm is used to find the optimal initialization parameters of the deep metric learning model.Therefore,the deep metric learning model can not only compare the similarity between the learning samples and the test samples,but also learn the useful information between the training samples.Moreover,the optimal initialization parameters of the model can make the algorithm only need a few gradient update steps to approach the optimal solution of the problem.The experimental results show that the meta relational network has strong generalization performance,and further improves the classification accuracy on on the benchmark set.(2)The meta attention query network is proposed to solve the problem of few shot classification.By querying the similarity coefficient between the samples and the learning samples in each class,the prototype representation of the centroid of each class can be obtained,which is more discriminative.Using the structure of self attention,we can customize a separate embedding space for each classification task.This embedding space can obtain the most distinguishable visual features for a given classification task.Finally,the point product attention is used to optimize the model agnostic meta learning algorithm,so that the algorithm can pay different attention to different tasks,so that different tasks pay more attention to the optimization of the model.The experimental results show that the meta attention query network can make the learning more effective and further improve the classification accuracy on the benchmark set.