Research On Few-shot Learning Methods Based On Data Augmentation And Model Fine-tuning

In recent years,few-shot learning is one of the research hotspots in the field of machine learning,which widely exists in many fields,such as disease diagnosis,dialogue system,automatic driving,face recognition and so on.Few-shot learning aims to obtain models with good performance in the case of limited sample.The main issue of few-shot learning is that the sample size of dataset is insufficient,leading to unreliable empirical risk.At present,the methods of few-shot learning are mainly studied from three perspectives: data,model and optimization algorithm.Data based methods use the information of samples and characteristics to augment the experience,model based methods carefully design the structures and parameters of models to reduce the size of hypothesis space,and optimization algorithm based methods change search strategies under a given hypothesis space to increase the probability of searching the optimal hypothesis.This thesis studies few-shot learning from the perspectives of data and model,and focuses on data augmentation methods and model fine-tuning methods.And the main research works are as follows.(1)It is difficult for model fine-tuning methods to obtain relative datasets,and data augmentation methods may introduce bias.To address these issues,we study few-shot learning by combining perspectives of data and model,and propose a few-shot learning framework based on data augmentation and model fine-tuning,called DAMFT＿FSL.In this framework,there are two modules: data augmentation module and model fine-tuning module,the former for increasing training size,the latter for correcting the bias introduced by data augmentation module.Furthermore,few-shot learning algorithms are proposed under this framework,according to the characteristics of few-shot learning.(2)Aiming at the issue of collecting relative datasets for few-shot data,under the framework DAMDT＿FSL,this thesis proposes a few-shot learning algorithm based on generative adversarial network and model head fine-tuning,called hGAN.To begin with,the generative adversarial network is trained on few-shot dataset to generate a certain number of relative samples.At the same time,the distribution learned by the generative adversarial network with few-shot dataset is always biased,and the classification model trained on the generated dataset cannot obtain unbiased estimations of the optimal hypothesis of few-shot task.Therefore,this model should be fine-tuned by original datasets.In order to constrain few-shot learning by the model pre-trained on relative samples better,the model head fine-tuning method is used to tine the pre-trained model.Experimental results on experimental datasets show that this method can effectively improve the performance of few-shot classification.(3)Aiming at the issue that the distributions of generator and discriminator learned on few-shot dataset are biased,this thesis proposes a few-shot learning algorithm based on reparameterized sampling for ensemble of generative adversarial networks and multihead loss fine-tuning,called MhERGAN.Generally,the learned distribution of discriminator is more accurate than generator.Based on this thought,Markov Chain Monte Carlo sampling,whose target distribution is the learned distribution of discriminator,is adopted to correct the learned distribution of generator.Meanwhile,when the sample size is insufficient,there is also a gap between the learned distribution of discriminator and real distribution.To address this issue,ensemble method is used on discriminator,so as to improve the stability of discriminator and correct the bias of discriminator.In addition,two model fine-tuning strategies,increasing the iteration number and multi-head loss,are used to improve the stability of model fine-tuning and the performance of classification.Experimental results on experimental datasets show that MhERGAN is more suitable for few-shot learning,which verifies the effectiveness of this algorithm.The few-shot learning framework based on data augmentation and model fine-tuning is suitable for few-shot learning,and the algorithms under this framework can alleviate the issue of model unreliability caused by insufficient sample size in few-shot learning,and can effectively improve the classification performance of few-shot learning.This research has important theoretical significance for few-shot learning.