Research On Weakly-supervised Learning Based On Sample Selection Strategy And Contrastive Learning

Great generalization performance of models in current deep learning heavily relies on large-scale and accurately annotated training data.However,in practical applications,it is often difficult to obtain high-quality training data,and learning algorithms face weak supervision environments with inaccurate or incomplete supervision information.Therefore,designing effective learning algorithms under weak supervision has always been a subject of great interest.Sample selection and contrastive learning are commonly used strategies in weakly-supervised learning.Models under weakly-supervised conditions are prone to be affected by label noise or wrong pseudo-labels,and designing reliable sample selection strategies is a core issue in weakly-supervised learning.At the same time,deep learning is essentially about representation learning,and designing effective contrastive learning strategies is crucial for solving weakly supervised problems.This thesis has conducted research on sample selection strategies and contrastive learning strategies under weak supervision conditions,focusing on the following aspects:To address the issue of missing boundary samples in sample selection strategies,from the perspective of prediction fluctuations,we design a novel selection strategy that can filter noisy labels while holding the boundary examples.By gradually filtering out the samples whose model prediction results are unstable,the boundary samples are retained.Meanwhile,this thesis designs a confidence penalty term based on the model output to prevent the model from overfitting to label noise.To address the issue of sensitivity to label noise in supervised contrastive learning,we design a noise-tolerant supervised contrastive learning framework consisting of a noise-robust contrastive loss and a weight-aware mechanism for label correction and selectively updating memory queues.In addition,we constructs a stochastic module for feature transformation to improve the representation ability of the contrastive learning framework.To address the issue of poor adaptability in pseudo-labeling methods with fixed thresholds,we propose a sample-aware threshold algorithm,which resorts to metalearning and automatically generates a sample-level confidence threshold through an additional meta-network.For updating the parameters of both the classification network and meta-network in this framework,we adopt the bi-level training strategy.Finally,we also provide a theoretical analysis of the learning framework,and experimentally give a solution to reduce the training complexity of the learning algorithm.The three algorithms proposed in this thesis achieved state-of-the-art performance on multiple public datasets for learning with noisy labels and semi-supervised tasks.The excellent performance on large-scale real-world datasets such as ImageNet and Clothing 1M demonstrates the effectiveness and practicality of our proposed algorithms.