Research On Few-Shot Learning Algorithm Based On Prototypical Network

With the development of machine learning and artificial intelligence,deep learning algorithms have made some breakthroughs in computer vision and other directions.However,plenty of training data is required for some practical tasks,some of which are very difficult to obtain.Meanwhile,different tasks usually require different data for training model.So we hope to find a way to learn quickly with very little data.That way is called few-shot learning.Prototypical network is a few-shot learning algorithm based on metric learning with the characteristics of simple network structure,low parameter dependency and the features of fast model optimization.Prototypical network can incorporate prior knowledge into the nature of the embedded space.This article is based on the prototypical network,the research will be done on sample prototype calculation,embedding space selection and loss function.The specific content is as follows.(1)In prototypical network,the prototype is calculated by the mean of every sample for per classes.Noise and outliers will influence on the process of prototype Calculation.We propose a prototypical network which uses the inverse distance weighting of the IMQ function.We determine the weight using the sum of the distance between each sample and other sample points.The weighted sum of each sample is used to obtain the category prototype.The experiments on public data sets verity the effectiveness of this method.(2)In view of the lack of spatial classification ability in embedded space and the high dimension of feature space,suitable subspaces will be constructed by all of the samples from every classes.The distance is calculated in subspace to calculate the projection of the test sample on each subspace for classification.Experiments verify the effectiveness of the algorithm(3)In view of the problem of similar categories in few-shot learning,a hybrid loss function based on Multi-path contrastive loss function is proposed.Multi-path contrastive loss function is used to shorten the distance between samples within a class and lengthen the distance between samples without classes.The mixed loss function is combine with cross entropy loss and multipath contrastive loss function.The experiments were conducted on the public data set to verify the effect of the method.