Self-Contrastive Learning: Single-Viewed Supervised Contrastive Framework Using Sub-network

Contrastive loss has significantly improved performance in supervised classification tasks by using a multi-viewed framework that leverages augmentation and label information. The enables contrast with another view of single image but enlarges training time memory usage. To exploit the strength multi-views while avoiding high computation cost, we introduce multi-exit architecture outputs multiple features single-viewed framework. this end, propose Self-Contrastive (SelfCon) learning, which self-contrasts within from different levels network. efficiently replaces multi-augmented images various information layers We demonstrate SelfCon learning improves encoder network, empirically analyze its advantages terms single-view sub-network. Furthermore, provide theoretical evidence increase based on mutual bound. For ImageNet ResNet-50, accuracy +0.6% 59% 48% Supervised simple ensemble boosts up to +1.5%. Our code is available at https://github.com/raymin0223/self-contrastive-learning.