Research On Few-shot Learning Method Based On Deep Feature Metric

Recently,deep learning has made significant progress in tasks such as image and speech based on big data sets.However,these achievements depend on the fact that based on the depth model,a large number of labeled samples are used for training,and the model parameters are iteratively updated.This type of optimization is less effective in the face of smaller data sets,because training a deep network based on a small data set can cause a serious over-fitting of the network.In this case,few-shot learning in deep learning has become the focus of research.Few-shot learning tasks refer to new classes that are not in training.In the case where only a small number of labeled samples in these classes are given,the classifier can be correctly identified.This task can be difficult for deep learning,but it is quite easy for humans.In addition,there are many applications for this task,such as the identification and classification of rare cases in medical images for auxiliary diagnosis.The most obvious point is that it can greatly reduce the workload of sample labeling while ensuring a certain classification accuracy.The solutions to few-shot learning tasks are mainly in the following categories:data enhancement,meta-learning,and metric learning.Considering the characteristics of few-shot learning,data enhancement can be used to reduce a certain over-fitting,but due to the small data space,the transformation mode is very limited and can not solve the over-fitting problem.The meta-learning method is widely used for few-shot learning because it is based on the high-level strategy of learning similar tasks.It is widely used,but the network structure is more complicated and less efficient.The metric learning method is simpler and more efficient.It firstly learns the embedded vector of the sample by embedding the network,and then directly solves the nearest neighbor in the embedded space to achieve the purpose of predictive classification.Based on the metric learning method,this paper proposes an improved scheme for the feature network and the loss function.The best known effects are currently achieved while efficiently solving this type of problem.The main research contents of this paper are as follows:1.The basic composition and basic methods of few-shot learning are expounded.The basic composition of few-shot learning system is introduced,including feature network,feature space analysis,evaluation criteria and datasets.Then the basic method of few-shot learning is introduced in detail,including data augmentation methods,meta-learning methods and metric learning methods.Then we systematically introduce the characteristics of various methods and analyze their advantages and disadvantages.2.This paper is inspired by the prototype idea to propose a representative feature network.For it simply uses the mean of embeddings to express class prototype,it can not evaluate the different contributions of each support set sample eigenvector in the class to the class prototype,and propose the representative feature concept.The representative feature network consists of two serial modules;an embedded module and a representative feature module.First,use the embedded module to extract the individual support set sample embedding vectors,and then stack multiple embedding vector inputs to represent the feature modules,and get the representative feature vector at last.Here we use two methods to compute the representative feature module.We first use the fixed solution method to fully consider the support target concentration of different targets.The influence of the embedded vector of the sample on the prototype is assigned different weights,and then weighted and summed to obtain the representative feature vector,and good results are obtained.In-depth,we use the learnable method to solve the representative features,and use the multi-layer perceptron structure to verify the excellent results of this method on multiple public datasets.3.The study found that some of classes in the few-shot classification tasks were more similar,such as the dogs in the minilmageNet dataset;and the rodents in the Cifar100 dataset.And because of the characteristics of few-shot learning:the data domain is very small and it is easy to cause the similar category to be misclassified.Therefore,this paper believes that by extending the distance between classes,the probability of similar categories can be effectively reduced.Specifically,after the optimization,the distance between the embedded vector of the test sample and the similar representative feature becomes closer,and the distance from the representative feature of the heterogeneous variable becomes longer,thereby avoiding misclassification for similar categories.According to this paper,the mixture loss function is proposed to add the relative error loss term to the cross entropy loss.In addition,the weight balance is considered in the design of the mixture loss function,and the importance degree of different loss terms is distinguished.The experimental results show that the mixture loss function can converge well and can reduce the probability of similar categories,so that the effect can be further improved.