Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Exploiting Fine-Grained Prototype Distribution for Boosting Unsupervised Class Incremental Learning

Jiaming Liu, Hongyuan Liu, Zhili Qin, Wei Han, Yulu Fan, Qinli Yang, Junming Shao

TL;DR

This work tackles unsupervised class incremental learning (UCIL) by recognizing labels may be unavailable in open-world sequences. It introduces a three-component framework: (1) distribution modeling via fine-grained Gaussian prototypes to capture detailed feature sub-clusters, (2) unsupervised class discovery guided by granular alignment that maximizes mutual information between fine-grained prototypes and coarse class assignments, and (3) knowledge preservation through overlap reduction that reuses memorized prototype statistics to prevent forgetting. The model uses a frozen Vision Transformer encoder and learns only the prototype and classifier components, achieving substantial gains over state-of-the-art methods across five datasets, including notable improvements of around 9% in challenging five-step splits. This approach offers a practical, scalable pathway for robust unsupervised continual learning in dynamic environments, with strong empirical support and clear avenues for future refinement.”

Abstract

The dynamic nature of open-world scenarios has attracted more attention to class incremental learning (CIL). However, existing CIL methods typically presume the availability of complete ground-truth labels throughout the training process, an assumption rarely met in practical applications. Consequently, this paper explores a more challenging problem of unsupervised class incremental learning (UCIL). The essence of addressing this problem lies in effectively capturing comprehensive feature representations and discovering unknown novel classes. To achieve this, we first model the knowledge of class distribution by exploiting fine-grained prototypes. Subsequently, a granularity alignment technique is introduced to enhance the unsupervised class discovery. Additionally, we proposed a strategy to minimize overlap between novel and existing classes, thereby preserving historical knowledge and mitigating the phenomenon of catastrophic forgetting. Extensive experiments on the five datasets demonstrate that our approach significantly outperforms current state-of-the-art methods, indicating the effectiveness of the proposed method.

Exploiting Fine-Grained Prototype Distribution for Boosting Unsupervised Class Incremental Learning

TL;DR

This work tackles unsupervised class incremental learning (UCIL) by recognizing labels may be unavailable in open-world sequences. It introduces a three-component framework: (1) distribution modeling via fine-grained Gaussian prototypes to capture detailed feature sub-clusters, (2) unsupervised class discovery guided by granular alignment that maximizes mutual information between fine-grained prototypes and coarse class assignments, and (3) knowledge preservation through overlap reduction that reuses memorized prototype statistics to prevent forgetting. The model uses a frozen Vision Transformer encoder and learns only the prototype and classifier components, achieving substantial gains over state-of-the-art methods across five datasets, including notable improvements of around 9% in challenging five-step splits. This approach offers a practical, scalable pathway for robust unsupervised continual learning in dynamic environments, with strong empirical support and clear avenues for future refinement.”

Abstract

The dynamic nature of open-world scenarios has attracted more attention to class incremental learning (CIL). However, existing CIL methods typically presume the availability of complete ground-truth labels throughout the training process, an assumption rarely met in practical applications. Consequently, this paper explores a more challenging problem of unsupervised class incremental learning (UCIL). The essence of addressing this problem lies in effectively capturing comprehensive feature representations and discovering unknown novel classes. To achieve this, we first model the knowledge of class distribution by exploiting fine-grained prototypes. Subsequently, a granularity alignment technique is introduced to enhance the unsupervised class discovery. Additionally, we proposed a strategy to minimize overlap between novel and existing classes, thereby preserving historical knowledge and mitigating the phenomenon of catastrophic forgetting. Extensive experiments on the five datasets demonstrate that our approach significantly outperforms current state-of-the-art methods, indicating the effectiveness of the proposed method.
Paper Structure (15 sections, 1 theorem, 14 equations, 7 figures, 7 tables, 1 algorithm)

This paper contains 15 sections, 1 theorem, 14 equations, 7 figures, 7 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Works
  3. Proposed Method
  4. Problem Definition
  5. Framework Overview
  6. Distribution Modeling by Fine-grained Prototypes
  7. Unsupervised Class Discovery by Granular Alignment
  8. Knowledge Preservation by Overlap Reduction
  9. Training Procedure
  10. Experiments
  11. Experimental Settings
  12. Comparison with Leading Approaches
  13. Case Study for Each Component
  14. Hyper-Parameter Analysis
  15. Conclusion and Discussion

Key Result

Proposition 3.1

$p(w, y)$ is equal to the matrix multiply by $W^{T} \cdot Y \in \mathbb{R}^{r\times k}$.

Figures (7)

  • Figure 1: An illustration of the proposed method, including three parts: (a) Fine-grained modeling, which explores fine-grained prototype distributions to capture more detailed information in the feature space. (b) Unsupervised class discovery, which guides the training of the classifier by granularity alignment. (c) Knowledge preservation, which extracts historical class distribution and decreases overlaps between new and existing classes.
  • Figure 2: The visualization of fine-grained Gaussian prototype distribution within feature space.
  • Figure 3: Hyper-parameter analysis on the number of prototypes $pNum$.
  • Figure 4: Hyper-Parameter analysis on (a) the weight of old loss $\lambda_{old}$, (b) the weight of the granular alignment regularization term $\lambda_{GA}$, (c) temperature score $\tau$, and (d) output dimension of the projector.
  • Figure : The visualization of fine-grained Gaussian prototype distribution within feature space.
  • ...and 2 more figures

Theorems & Definitions (1)

  • Proposition 3.1