TransFusion: Contrastive Learning with Transformers

This paper proposes a novel framework, TransFusion, designed to make the process of contrastive learning more analytical and explainable. TransFusion consists of attention blocks whose softmax being replaced by ReLU, and its final block's weighted-sum operation is truncated to leave the adjacency matrix as the output. The model is trained by minimizing the Jensen-Shannon Divergence between its output and the target affinity matrix, which indicates whether each pair of samples belongs to the same or different classes. The main contribution of TransFusion lies in defining a theoretical limit for answering two fundamental questions in the field: the maximum level of data augmentation and the minimum batch size required for effective contrastive learning. Furthermore, experimental results indicate that TransFusion successfully extracts features that isolate clusters from complex real-world data, leading to improved classification accuracy in downstream tasks.

Contrastive Learning with Orthonormal Anchors (CLOA)

This study focuses on addressing the instability issues prevalent in contrastive learning, specifically examining the InfoNCE loss function and its derivatives. We reveal a critical observation that these loss functions exhibit a restrictive behavior, leading to a convergence phenomenon where embeddings tend to merge into a singular point. This "over-fusion" effect detrimentally affects classification accuracy in subsequent supervised-learning tasks. Through theoretical analysis, we demonstrate that embeddings, when equalized or confined to a rank-1 linear subspace, represent a local minimum for InfoNCE. In response to this challenge, our research introduces an innovative strategy that leverages the same or fewer labeled data than typically used in the fine-tuning phase. The loss we proposed, Orthonormal Anchor Regression Loss, is designed to disentangle embedding clusters, significantly enhancing the distinctiveness of each embedding while simultaneously ensuring their aggregation into dense, well-defined clusters. Our method demonstrates remarkable improvements with just a fraction of the conventional label requirements, as evidenced by our results on CIFAR10 and CIFAR100 datasets.