Research On Few-Shot Learning Based On Deep Learning

In recent years,with the development of deep learning in machine learning and pattern recognition,more and more artificial intelligence applications have entered the public's sight.However,deep learning's models have two major shortcomings.One is that those models must be trained multiple times by a large number of labeled samples.Besides,it is difficult to collect and mark the unlabeled samples in applications.The other is that the model must be retrained whenever the train data changes.The Few-Shot Learning(FSL)algorithm provides theoretical and methodological support for solving these problems,and it has become a hot issue in the field of machine learning.FSL aims to build a classification model with good generalization ability by using a small number of labeled samples.Since the theory was born,there are many research results in image recognition be proposed.However,as a hot issue in the field of machine learning,the existing few-shot learning theory and methods still have some problems that have not been solved so far.For example,due to the small number of labeled samples,the model performs overfitting and low generalization.Therefore,how to improve the generalization ability of FSL model is an important research direction of machine learning,which is receiving widespread attention in the academic community.This dissertation first analyzes the present situation and existing problems of FSL,then proposes two solutions to the existing problems to improve the recognition rate of the model.In one word,the main research work of this dissertation is shown in the following aspects:1)Aiming at the problem of data sensitivity,a Few-Shot Learning based on Bagging Model(FSLBM)is proposed.FSLBM use the bagging algorithm to reduce the impact of noisy data,making the model become better.Firstly,a random sampling method is used to generate multiple different training sets from the original training set.Secondly,each training set is randomly divided to generate a query set and a sample set respectively.Thirdly,create multiple asynchronous threads,and build a relation network model on each thread.Finally,multiple heterogeneous base models are trained in parallel,then the base models are fused using probability voting method.2)Aiming at the problem of the weak generalization ability of the existing FSL algorithms when the feature difference between the source domain data and the target domain data is large,a Few-Shot Learning based on Semi-Supervised with Pseudo-labels(FSLSP)is proposed.FSLSP use the pseudo-label semi-supervised algorithm to approximate the feature expression of the source domain data and the target domain data,so FSLSP can improve the generalization ability of the model.Firstly,build a relation network and pre-train it with source domain data.Secondly,the pre-trained network is used to predict the labels of the target domain data,then the predicted labels are recorded as pseudo labels.Finally,the source domain data with real labels and the target domain data with pseudo labels are used to hybrid train the pre-trained network,then the pseudo-labeled and hybrid training processes are repeated until the training is completed.The comparative experimental results show that in the 5 (6 5 ?,5 (6 1 ?,20 (6 1 ?,and 20 (6 5 ? classification scenarios,compared with other FSL algorithms,both the FSLBM and the FSLSP have higher classification recognition rates.