239

ff-MICL: Understanding and Generalizing InfoNCE-based Contrastive Learning

Abstract

In self-supervised contrastive learning, a widely-adopted objective function is InfoNCE, which uses the heuristic cosine similarity for the representation comparison, and is closely related to maximizing the Kullback-Leibler (KL)-based mutual information. In this paper, we aim at answering two intriguing questions: (1) Can we go beyond the KL-based objective? (2) Besides the popular cosine similarity, can we design a better similarity function? We provide answers to both questions by generalizing the KL-based mutual information to the ff-Mutual Information in Contrastive Learning (ff-MICL) using the ff-divergences. To answer the first question, we provide a wide range of ff-MICL objectives which share the nice properties of InfoNCE (e.g., alignment and uniformity), and meanwhile result in similar or even superior performance. For the second question, assuming that the joint feature distribution is proportional to the Gaussian kernel, we derive an ff-Gaussian similarity with better interpretability and empirical performance. Finally, we identify close relationships between the ff-MICL objective and several popular InfoNCE-based objectives. Using benchmark tasks from both vision and natural language, we empirically evaluate ff-MICL with different ff-divergences on various architectures (SimCLR, MoCo, and MoCo v3) and datasets. We observe that ff-MICL generally outperforms the benchmarks and the best-performing ff-divergence is task and dataset dependent.

View on arXiv
Comments on this paper