Abstract
Generalized category discovery (GCD) is a problem setting where the goal is to discover novel categories within an unlabelled dataset using the knowledge learned from a set of labelled samples. Recent works in GCD argue that a non-parametric classifier formed using semi-supervised $k$-means can outperform strong baselines which use parametric classifiers as it can alleviate the over-fitting to seen categories in the labelled set. In this paper, we revisit the reason that makes previous parametric classifiers fail to recognise new classes for GCD. By investigating the design choices of parametric classifiers from the perspective of model architecture, representation learning, and classifier learning, we conclude that the less discriminative representations and unreliable pseudo-labelling strategy are key factors that make parametric classifiers lag behind non-parametric ones. Motivated by our investigation, we present a simple yet effective parametric classification baseline that outperforms the previous best methods by a large margin on multiple popular GCD benchmarks. We hope the investigations and the simple baseline can serve as a cornerstone to facilitate future studies. Our code is available at: this https URL.
Abstract (translated)
URL
https://arxiv.org/abs/2211.11727