A Contrastive Learning Algorithm Based on Bayesian Random Semantic Feature Augmentation

Abstract The Simple Contrastive Learning (SimCLR) algorithm is a self-supervised method that projects high-dimensional features into a low-dimensional space to enhance the similarity between positive samples and the difference between negative samples. However, this projection may cause the loss of feature information, significantly impacting the performance of downstream tasks. Therefore, we propose A Bayesian Random Semantic Data Augmentation Algorithm (SimKFA), an improved contrastive learning algorithm. A learnable linear transformation layer, KLP, replaces the multilayer perceptron (MLP) to preserve feature information. A feature enhancement module (BFAB), which combines Bayesian inference and frequency-domain feature analysis, is proposed. This module adaptively captures key information, maintains the continuity of the information flow, and reduces feature loss during deep learning. The experimental results show that the proposed algorithm prevents loss of feature information caused by the projection head and achieves excellent performance on multiple benchmark datasets.