Model-Aware Contrastive Learning: Towards Escaping Uniformity-Tolerance Dilemma in Training

Contrastive learning (CL) has achieved remarkable success in learning transferable representations. It has been identified that the temperature $ \tau $ of CL loss plays an essential role in automatically concentrating on hard negative samples. However, recent work also indicates a uniformity-tolerance dilemma (UTD) connected to $ \tau $, which will lead to unexpected performance degradation. We argue that it is the fixity of temperature that is inextricably linked to UTD and suboptimal embedding space. To tackle the challenge of UTD, we enrich the CL loss family by presenting a Model-Aware Contrastive Learning (MACL) strategy. In MACL, the temperature parameter is adaptive to the magnitude of alignment that reflects the basic confidence of the instance discrimination task. Lower alignment implies poor discrimination for the undertrained phase, then there is less possibility that the high similarity region contains latent positive samples (LPs). Thus, a small $ \tau $ can impose larger penalties on hard negative samples to learn uniformly informative embeddings. Instead, a larger $ \tau $ in the well-trained phase facilitates the exploration of semantic structures due to its increased tolerance for LPs. Besides, theoretically, we uncover why contrastive learning requires a large number of negative samples from a unified gradient reduction perspective. Based on MACL and these analyses, a new CL loss is proposed. Experimental results validate the effectiveness of our approach to escape UTD, which can achieve state-of-the-art performance and training with fewer negative samples.