Prompt-SSLC: A Unified Framework for Dual Prompt-Augmented Semi-Supervised Sequential Leader Clustering in On-the-Fly Category Discovery

Hon Lam Li, Jie Li, Shaozhe Hao, Peng Wang, Kai Han

Keywords: On-the-Fly Category Discovery, Dual Prompt Learning, Semi-Supervised Clustering, Sequential Leader Clustering, Open-Set Recognition, Streaming Data

TL;DR: Prompt-SSLC integrates SSLC, dual prompting, and open-set-aware classification for real-time category discovery, dynamically updating prototypes and features to achieve state-of-the-art performance on streaming data.

Abstract: On-the-fly Category Discovery (OCD) enables intelligent systems to perform real-time predictions while adapting to emerging categories in dynamic environments. We present Prompt-SSLC, a unified framework that integrates three synergistic components to balance stability and adaptability in streaming data scenarios. First, Semi-Supervised Sequential Leader Clustering (SSLC) dynamically updates prototypes to accommodate incoming data streams, ensuring flexibility in clustering. To enhance discriminability and mitigate prototype overlap, SSLC incorporates a Distance-Aware (DA) update mechanism that optimizes prototype distributions, maintaining inter-class separation as new data arrive. Second, dual prompting augments the foundation model: a *Task* Prompt guides category discovery, while an *Instance* Prompt dynamically recalibrates features to prevent drift toward previously learned classes without requiring retraining. Third, an Open-Set-Aware (OSA) classifier employs uncertainty estimation to identify and filter ambiguous samples, ensuring robust prototype updates. This cohesive integration of streaming clustering, feature recalibration, and uncertainty-aware filtering establishes a robust framework for OCD. Extensive experiments on generic and fine-grained benchmarks demonstrate that Prompt-SSLC achieves significant performance improvements, setting a new state-of-the-art for OCD.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Submission Number: 8473