Research On Semi-supervised Few-shot Learning Method Based On Ensemble Learning Strategy

Because of the development of deep learning technology,various methods based on this technology have achieved impressive application results in the fields of pattern recognition and natural language processing,which greatly promoted the implementation of intelligent algorithms.However,an important prerequisite for various methods based on deep learning theory to achieve good generalization performance is to use a large number of labeled samples to train the model,so building a deep learning model with excellent performance requires high labor and time costs.In fact,human beings have the ability to learn new things quickly.For an entirely new thing that has never been seen before,human beings can often abstract the important features of this thing in a limited number of observations and accurately identify the new entity of this thing.Inspired by this reality,researchers began to try to build a model with good generalization performance from a small number of labeled samples,and the problem of few-shot learning was born.The few-shot learning task requires the use of a small number of samples with labeled information to build a classifier model with good performance.In this paper,we analyzes the few-shot learning task in detail,and deeply studies the key factors that affect the performance of the few-sho learning models.Finally,we propose innovative methods from the perspective of sample feature distribution,sample size,and ensemble learning.The main research contents of this paper are as follows:1.Explain the source and research significance of few-shot learning problems;introduce the basic methods and theories involved in the few-shot learning;analyze the design ideas of various classic few-shot learning methods in detail;summarize the basic process of constructing a few-shot learning method.2.In order to solve the problem of weak nonlinear mapping ability of the feature extractor in classic few-shot learning methods,a directional feature shifting network is proposed.This network can shift the original feature vectors to the corresponding prototype estimates,so as to achieve the effect of clustering similar samples around the corresponding prototype estimates.Experiment shows that the directional feature shifting network can effectively reduce the intra-class distance,and reducing the difficulty of classification and improving the classification accuracy of the model.3.In order to solve the problem of low stability and low accuracy of original prototype estimation in the directional feature shifting network,two methods of data augmentation performed in feature space are proposed.Experiment shows that both feature augmentation methods will effectively improve the final classification effect of the directional feature shifting network.In addition,there are many random processes in the semi-supervised directional feature shifting network proposed,which is consistent with the requirements of ensemble learning methods.Therefore,we also attempts to use the ensemble strategy to build few-shot learning models.Experiment shows that the ensemble learning method is effective in the few-shot learning problem.