Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ImbaGCD: Imbalanced Generalized Category Discovery

Ziyun Li, Ben Dai, Furkan Simsek, Christoph Meinel, Haojin Yang

TL;DR

ImbaGCD addresses the realistic setting of imbalanced unlabeled data in generalized category discovery by proposing an optimal-transport–assisted EM framework that aligns the marginal class prior with the data. The method integrates Sinkhorn-based pseudo-labeling, prototype-driven likelihoods, and contrastive representation learning, together with moving-average class-prior estimation and momentum prototypes to combat head-class bias. Empirical results on CIFAR-100 and ImageNet-100 show consistent improvements over state-of-the-art GCD methods, particularly for novel/unknown classes under various imbalance factors, while remaining competitive in balanced scenarios. The work advances practical GCD by explicitly modeling and mitigating long-tail effects in unlabeled data, enabling more reliable discovery of unseen categories in real-world settings.

Abstract

Generalized class discovery (GCD) aims to infer known and unknown categories in an unlabeled dataset leveraging prior knowledge of a labeled set comprising known classes. Existing research implicitly/explicitly assumes that the frequency of occurrence for each category, whether known or unknown, is approximately the same in the unlabeled data. However, in nature, we are more likely to encounter known/common classes than unknown/uncommon ones, according to the long-tailed property of visual classes. Therefore, we present a challenging and practical problem, Imbalanced Generalized Category Discovery (ImbaGCD), where the distribution of unlabeled data is imbalanced, with known classes being more frequent than unknown ones. To address these issues, we propose ImbaGCD, A novel optimal transport-based expectation maximization framework that accomplishes generalized category discovery by aligning the marginal class prior distribution. ImbaGCD also incorporates a systematic mechanism for estimating the imbalanced class prior distribution under the GCD setup. Our comprehensive experiments reveal that ImbaGCD surpasses previous state-of-the-art GCD methods by achieving an improvement of approximately 2 - 4% on CIFAR-100 and 15 - 19% on ImageNet-100, indicating its superior effectiveness in solving the Imbalanced GCD problem.

ImbaGCD: Imbalanced Generalized Category Discovery

TL;DR

ImbaGCD addresses the realistic setting of imbalanced unlabeled data in generalized category discovery by proposing an optimal-transport–assisted EM framework that aligns the marginal class prior with the data. The method integrates Sinkhorn-based pseudo-labeling, prototype-driven likelihoods, and contrastive representation learning, together with moving-average class-prior estimation and momentum prototypes to combat head-class bias. Empirical results on CIFAR-100 and ImageNet-100 show consistent improvements over state-of-the-art GCD methods, particularly for novel/unknown classes under various imbalance factors, while remaining competitive in balanced scenarios. The work advances practical GCD by explicitly modeling and mitigating long-tail effects in unlabeled data, enabling more reliable discovery of unseen categories in real-world settings.

Abstract

Generalized class discovery (GCD) aims to infer known and unknown categories in an unlabeled dataset leveraging prior knowledge of a labeled set comprising known classes. Existing research implicitly/explicitly assumes that the frequency of occurrence for each category, whether known or unknown, is approximately the same in the unlabeled data. However, in nature, we are more likely to encounter known/common classes than unknown/uncommon ones, according to the long-tailed property of visual classes. Therefore, we present a challenging and practical problem, Imbalanced Generalized Category Discovery (ImbaGCD), where the distribution of unlabeled data is imbalanced, with known classes being more frequent than unknown ones. To address these issues, we propose ImbaGCD, A novel optimal transport-based expectation maximization framework that accomplishes generalized category discovery by aligning the marginal class prior distribution. ImbaGCD also incorporates a systematic mechanism for estimating the imbalanced class prior distribution under the GCD setup. Our comprehensive experiments reveal that ImbaGCD surpasses previous state-of-the-art GCD methods by achieving an improvement of approximately 2 - 4% on CIFAR-100 and 15 - 19% on ImageNet-100, indicating its superior effectiveness in solving the Imbalanced GCD problem.
Paper Structure (28 sections, 1 theorem, 16 equations, 3 figures, 3 tables, 1 algorithm)

This paper contains 28 sections, 1 theorem, 16 equations, 3 figures, 3 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. PRELIMINARIES
  3. Problem Setup
  4. Reminders on Optimal Transport
  5. Method
  6. Prototype Loss for Unlabeled Data
  7. E-step
  8. M-step
  9. Representation Improvement
  10. Overall Loss Objective
  11. Class Prior Estimation & Prototype Updates
  12. Moving-average distribution update
  13. Momentum prototypes update
  14. Experiments
  15. Setup
  16. ...and 13 more sections

Key Result

Lemma 1

For $\lambda > 0$, the solution $\mathbf{P}^{\lambda}$ is unique and can be represented as $\mathbf{P}^{\lambda} = \operatorname{diag}(\mathbf{\alpha}) \mathbf{K} \operatorname{diag}(\mathbf{\beta})$, where $\mathbf{\alpha}$ and $\mathbf{\beta}$ are two non-negative vectors uniquely determined up to

Figures (3)

  • Figure 1: Illustration of ImbaGCD. ImbaGCD attempts to identify known categories and discover new classes within a vast amount of unlabeled data from the real world. In the real world, unlabeled data includes both known class from the labeled set and unknown class, where known classes (e.g., cat, dog, rabbit) dominate the well-represented "head," while "unknown" classes (e.g., lion, rhino, panda) primarily reside in the underrepresented "tail" of the distribution.
  • Figure 2: Comparison of the predicted sample numbers for each class on the CIFAR-10 dataset, using an exponential decreasing strategy with an imbalance factor ($\rho$) of 5.
  • Figure 3: Comparison of the predicted sample numbers for each class on the CIFAR-10 dataset, using an step decreasing strategy with an imbalance factor ($\rho$) of 5.

Theorems & Definitions (2)

  • Definition 1: Generalized Category Discovery
  • Lemma 1