Improved contrastive learning with MoCo framework

Abstract

Self-supervised learning typically suffers from lacking contrastive pairs and extracting unrepresentative vectors. To handle above mentioned challenges, this paper introduces a novel self-supervised learning framework that integrates the location-based sampling manner and a well-designed dimensionality reduction module. In the location-based sampling module, this paper embeds a multi-crop sampling paradigm into the memory bank-based framework. In the dimensionality reduction module, this paper introduces a principal component dimensionality reduction to capture the most comprehensive features. Experiments on popular datasets demonstrate the superior performance of our proposed method.