Stochastic Approximation to Contrastive Learning
Keywords: contrastive learning, self-supervised learning, unsupervised learning
TL;DR: New contrastive learning method which works well with one negative sample.
Abstract: Contrastive learning is a powerful paradigm that has been crucial for self-supervised representation learning. While there is evidence for its effectiveness, these methods typically rely on arbitrary definitions of positive and negative pairs. Most existing contrastive learning methods require large batch sizes during training due to their rigid control over the tradeoff between the two contrastive terms. Consequences are that, substantial computational resources are wasted on negative pairs that provide minimal learning signals. To address this issue, this work present a novel method. We reformulate contrastive learning as a matrix approximation problem using I-divergence, a non-normalized form of Kullback-Leibler divergence. Our proposed objective function is decomposable across instance pairs, enabling the development of efficient stochastic approximation algorithms from neighbor embeddings which perform well with fewer negative samples. Additionally, we generalize the scaling factor beyond normalization, allowing it to adaptively emphasize positive pairs that carry more learning signals, thereby reducing the computational waste associated with negative pairs. Experimental results on visual representation learning benchmark datasets such as CIFAR and ImageNet demonstrate major improvements over other contrastive learning methods, particularly when using small batches and with only one negative pair.
Supplementary Material: zip
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 8017
Loading