AttentionNCE: Contrastive Learning with Instance Attention

Bin Liu, Xiaohong Liu, Ziqiao Shang, Lin Ma, Zhaoyu Li, Fei Teng, Tianrui Li

Keywords: Contrastive Learning, Image Classification

TL;DR: This paper constructs a variational lower bound for the contrastive loss and correspondingly proposes a new contrastive loss.

Abstract:

Contrastive learning has found extensive applications in computer vision, natural language processing, and information retrieval, significantly advancing the frontier of self-supervised learning. However, the limited availability of labels poses challenges in contrastive learning, as the positive and negative samples can be noisy, adversely affecting model training. To address this, we introduce instance-wise attention into the variational lower bound of contrastive loss, and proposing the AttentionNCE loss accordingly. AttentioNCE incorporates two key components that enhance contrastive learning performance: First, it replaces instance-level contrast with attention-based sample prototype contrast, helping to mitigate noise disturbances. Second, it introduces a flexible hard sample mining mechanism, guiding the model to focus on high-quality, informative samples. Theoretically, we demonstrate that optimizing AttentionNCE is equivalent to optimizing the variational lower bound of contrastive loss, offering a worst-case guarantee for maximum likelihood estimation under noisy conditions. Empirically, we apply AttentionNCE to popular contrastive learning frameworks and validate its effectiveness. The code is released at:
\url{https://anonymous.4open.science/r/AttentioNCE-55EB}

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: 4392