Towards Debiased Generalized Category Discovery

Pengcheng Guo, Boyu Wang, Yonghong Song

Abstract:

Generalized Category Discovery (GCD) aims at
classifying unlabeled training data coming from old
and novel classes by leveraging the information of
partially labeled old classes. In this paper, we reveal
that existing methods often suffer from competition
between new and old classes, where the focus
on learning new classes often results in a notable
performance degradation on the old classes. Moreover,
we delve into the reason behind this problem:
the GCD classifier can be overconfident and
biased towards the new class. With this insight, we
propose Debiased GCD (DeGCD), a simple but effective
approach that mitigates the bias caused by
the overconfidence from new categories by a debiased
head. Specifically, we first propose semantic
calibration loss that aids the GCD classifier in debiasing
by enforcing neighborhood prediction consistency
with the latent representation of the debiased
head. Furthermore, a debiased contrastive objective
is proposed to refine the similarity matrix
from the GCD classifier and the debiased classifier,
suppressing the overconfidence in new classes
in unlabeled data. In addition, an alignment constraint
loss is designed to prevent damaging the distribution
of the old categories caused by overconfidence
in the new categories. Experiments on various
datasets shows DeGCD achieves state-of-theart
performance and maintains a good balance between
new and old classes. In addition, this method
can be seamlessly adapted to other GCD methods,
not only to achieve further performance gains but
also to effectively balance the performance of the
new class with that of the old class.