Semantic-Enhanced Prototypical Network for Universal Novel Category Discovery

Haonan Lin, Wenbin An, Yan Chen, Feng Tian, Wei Ding, QianYing Wang, Ping Chen, Yaqiang Wu, mingxiang cai, Guang Dai

Keywords: Prototype Refinement strategy, Semantic enrichment, Prototypical Network, Patch-Entropy Balance

Abstract: We address the challenging task of Universal Novel Category Discovery (UniNCD) in image classification, where models must distinguish between common and novel categories while avoiding the misclassification of novel categories as private-known ones. Previous prototype-based approaches face two major challenges: first, they significantly increase the negative transfer risk by often misaligning novel categories with private-known categories; second, they lead to sub-optimal prototypes because traditional prototype learning ignores diverse object characteristics of images, resulting in insufficient semantic guidance when optimizing instance representations using only instance-level prototypical distributions. To tackle these challenges, we present a Semantic-Enhanced Prototypical Network, dubbed SEPNet. This prototypical network is enhanced by refined prototypes and enriched semantics to learn better representations and avoid negative transfer, including three key ideas: (1) we design a Prototype Refinement (PR) strategy that can decouple common, private-known, and novel categories from unlabeled data, which can exclude misaligned prototypes to avoid negative transfer; (2) we attach prototypical distribution to each patch of images, which embed enhanced semantic information to prototypes and guide prototypical contrastive learning and, (3) we design a patch-entropy balance (PEB) method to encourage sparser patch-level prototypical distributions while maintaining the uniformity of dense distributions, sparsity emphasizes dominant category characteristics, and uniformity avoids the misguidance of irrelevant disturbance, thereby enhancing the distinctiveness of instances to the prototypes. Our method demonstrates superior performance on the UniNCD task across three benchmark datasets, outperforming existing state-of-the-art approaches by approximately 3.4% in terms of accuracy. We will release our code for reproduction.

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/2024/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: 7270