Research On Few-shot Learning Method Based On Metric Learning

Few-shot learning holds significant importance in the field of machine learning,as practical applications often encounter the issue of limited training samples.In order to tackle this challenge,metric learning has emerged as an effective method,achieving remarkable success in few-shot learning.However,existing metric learning methods usually assume that intra-class samples contribute equally to the class prototype,an assumption that may lead to prototype computation being easily influenced by outliers or noisy data.Therefore,this thesis aims to investigate a method that can better capture the sample structure and thus enhance few-shot learning performance.For this purpose,the thesis proposes two methods: the Weighted Prototype Network and the Enhanced Weighted Prototype Network,in order to obtain better class prototypes.The Weighted Prototype Network is a method that improves the prototype network by introducing a weight factor.In the computation of class prototypes in this method,the contribution of each sample is no longer equal but is weighted according to the contribution of intra-class samples to the class prototype.This weighting strategy helps better capture the intra-class structure,thereby improving the performance of few-shot learning.To further improve performance,this thesis introduces an Enhanced Weighted Prototype Network.In this method,in addition to considering the contributions of intra-class samples to the class prototype,the contributions of inter-class samples to the class prototype are also considered.This adjustment strategy helps better distinguish different categories in the feature space,thus further improving classification performance.To validate the effectiveness of the proposed methods,experiments were conducted on two challenging benchmark datasets,mini Image Net and tiered Image Net.The experimental results show that,compared to other existing methods,the proposed Weighted Prototype Network and Enhanced Weighted Prototype Network methods achieved performance improvements in various few-shot learning tasks.Furthermore,a series of experiments were conducted to deeply explore the roles of each module in the proposed methods,and the final experimental results demonstrate the effectiveness of the proposed methods.In summary,this thesis delves into few-shot learning methods based on metric learning and proposes the Weighted Prototype Network and Enhanced Weighted Prototype Network methods.Experiments show that these two methods achieved excellent performance on the benchmark datasets mini Image Net and tiered Image Net,providing a new perspective for addressing few-shot learning problems.