Research On Semi-supervised Learning Models For Limited Labeled Data

Deep learning has been successful in numerous fields,such as computer vision,natural language processing,audiovisual speech recognition,biomedicine,and autonomous driving.State-of-the-art deep learning models are often trained on a large-scale and highquality labeled training dataset.However,manually labeling a massive dataset is impractical under limited resources and cost constraints.On the other hand,unlabeled data is usually abundant and easily accessible.Therefore,in the case of limited labeled data,utilizing a large amount of unlabeled data for semi-supervised learning is considered a promising approach.Traditional semi-supervised learning models rely on an ideal and closed environment where all possible classes and domains are known in advance.In contrast,realworld application scenarios are open and may involve the emergence of new categories.Building upon the foundation of comprehensive research on closed-world semi-supervised models and considering the limitations of labeled data,this dissertation proposes three new semi-supervised learning models for visual recognition in an open-world scenario,addressing the issue in representation learning and label assignment processes.The main research content is summarized as follows,Closed-world semi-supervised learning models have made significant progress on visual recognition tasks.However,there is a lack of clear criteria to classify and systematically study these models.Based on this deficiency,this dissertation classifies and studies the existing closed-world semi-supervised learning models according to the characteristics of their model architectures and loss functions,dividing these models into deep semi-supervised generative models,semi-supervised consistency regularization models,graph-based semi-supervised models,pseudo-labeling models,and hybrid models.In this dissertation,representative works is selected for each class of models,and comparisons are made in terms of model architecture,method idea,loss function,and test performance,to provide reference for subsequent model research.In open-world scenario,existing closed-world semi-supervised learning models are limited by their closed-world assumptions and are mostly unable to learn new semantic categories of unlabeled instances.This dissertation investigates semi-supervised learning models in a pragmatic but under-explored setting of generalized category discovery.Generalized category discovery aims to recognize categories in unlabeled dataset that contain unknown new categories based on the limited labled dataset.To address the issues of insufficient discriminative representation learning and assuming that the number of categories is known in previous works,this dissertation proposes a co-training-based framework named CoGCD that encourages clustering assignment consistency.Specifically,this work introduce strong and weak augmentation transformations to generate two sufficiently different views for the same instance.Then,based on the co-training assumption,a consistency representation learning strategy is proposed,which encourages consistency between the clustering assignments of the two views.Finally,the discriminative embeddings learned from the semi-supervised representation learning process are used to construct an original sparse network and use a modify Louvain algorithm to obtain the clustering results and the number of categories simultaneously.During the experiments,this dissertation finds that the clustering assignment consistency of strong and weak augmentation transformations in the representation learning process helps the model to discover novel categories.Similar to existing models,the CoGCD model proposed in this dissertation still cannot avoid the problem of introducing false negative samples in representation learning.To handle this problem,this dissertation proposes a novel framework for generalized category discovery,called GL-GCD.This framework incorporates both global distribution similarity and local cross-instance neighbor representations simultaneously.To learn global representations,the weakly and strongly augmented embeddings of the same instance are encouraged to have similar similarity relationships respected to other embeddings in the support queue.For local representation learning,the feature embeddings that exhibit high similarity to the current instance embedding in the support queue are used as positive pairs,and the rest of the features are used as negative pairs,thus constructing the neighbor contrastive learning loss function and obtaining the cross-instance local representations.In the experimental section,it is found that the joint global and local representation learning not only learns rich discriminative feature representations,but also effectively alleviates the problem of class collision issue caused by false negative paris.The previous models as well as the first two models proposed in this dissertation are all two-stage models,where the representation learning and label assignment processes are relatively independent.It has been shown that when the data is projected into a feature space with a dimensionality of the target cluster number,the rows and columns of its feature matrix correspond to the instance and cluster representation,respectively.Based on the observation,this dissertation proposes Bi GCD,a generalized category discover framework based on Bi-contrastive learning.This framework respectively conducts instanceand cluster-level contrastive learning,and integrates the label assignment process into the representation learning process to form a one-stage model.In instance-level contrastive learning,this dissertation builds on the research of CoGCD and GL-GCD models to construct unsupervised contrastive loss for strong and weak views as well as nearest neighbor contrastive loss.In cluster-level contrastive learning,consistency between the clusterlevel representations of strong and weak views is encouraged.In addition,in order to cope with the problem of poor pseudo-labeling quality,this dissertation proposes a confidencebased criterion for boosting both the instance-and cluster-level contrastive learning,which further improves the model performance.