PromptCCD: Learning Gaussian Mixture Prompt Pool for Continual Category Discovery
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.
Keywords: Continuous Category Discovery, Generalized Category Discovery, Continual Learning, Incremental Learning, Open World
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: we propose PromptCCD, a simple yet effective approach that utilizes Gaussian mixture models as a prompting method for CCD task.
Abstract: In this paper, we address the challenging open-world learning problem of continual category discovery (CCD). Initially, a labelled dataset consisting of known categories is provided to the model. Subsequently, unlabelled data arrives continuously at different time steps, which may contain objects from known or novel categories. The primary objective of CCD is to automatically assign labels to unlabelled objects, regardless of whether they belong to seen or unseen categories. However, the crucial challenge in continual category discovery is to automatically discover new categories in the unlabelled stream without experiencing catastrophic forgetting, which remains an open problem even in conventional, fully supervised continual learning.
To address this challenge, we propose PromptCCD, a simple yet effective approach that utilizes Gaussian mixture model as a prompting method for CCD. At the core of PromptCCD is our proposed Gaussian Mixture Prompt Module (GMP), which acts as a dynamic pool updating over time to provide guidance for embedding data representation and avoid forgetting during continual category discovery.
Additionally, we introduce a GM-based category estimation module into PromptCCD, which enables it to discover categories in the unlabelled stream without prior knowledge of category numbers.
Finally, we extend the standard evaluation metric for generalized category discovery to CCD and benchmark state-of-the-art methods using different datasets. Our PromptCCD significantly outperforms other methods, demonstrating the effectiveness of our approach.
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: 1406
Loading