Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Enhancing Generative Class Incremental Learning Performance with Model Forgetting Approach

Taro Togo, Ren Togo, Keisuke Maeda, Takahiro Ogawa, Miki Haseyama

TL;DR

This work tackles Generative Class Incremental Learning (GCIL) under streaming data with limited capacity by introducing a forgetting mechanism inspired by human memory. The method first applies Selective Amnesia (SA) to forget a designated class $c_f$, producing a forgetting model $M_f$ via Generative Replay, and then trains a final model $M_{final}$ to learn a new class $c_{new}$ from $D_{new}$ while leveraging retained information from $D_r$. It combines SA-based forgetting with Elastic Weight Consolidation (EWC) to balance the retention of old knowledge and acquisition of new information, leveraging a diagonal Fisher Information Matrix for efficiency. Across MNIST and Fashion-MNIST with a one-hot VAE base, the approach yields improved $c_{new}$ generation accuracy under forgetting in both EWC and fine-tuning regimes, and it reveals the role of class characteristics in the forgetting process, underscoring the practical impact for memory-constrained continual generative learning.

Abstract

This study presents a novel approach to Generative Class Incremental Learning (GCIL) by introducing the forgetting mechanism, aimed at dynamically managing class information for better adaptation to streaming data. GCIL is one of the hot topics in the field of computer vision, and this is considered one of the crucial tasks in society, specifically the continual learning of generative models. The ability to forget is a crucial brain function that facilitates continual learning by selectively discarding less relevant information for humans. However, in the field of machine learning models, the concept of intentionally forgetting has not been extensively investigated. In this study we aim to bridge this gap by incorporating the forgetting mechanisms into GCIL, thereby examining their impact on the models' ability to learn in continual learning. Through our experiments, we have found that integrating the forgetting mechanisms significantly enhances the models' performance in acquiring new knowledge, underscoring the positive role that strategic forgetting plays in the process of continual learning.

Enhancing Generative Class Incremental Learning Performance with Model Forgetting Approach

TL;DR

This work tackles Generative Class Incremental Learning (GCIL) under streaming data with limited capacity by introducing a forgetting mechanism inspired by human memory. The method first applies Selective Amnesia (SA) to forget a designated class , producing a forgetting model via Generative Replay, and then trains a final model to learn a new class from while leveraging retained information from . It combines SA-based forgetting with Elastic Weight Consolidation (EWC) to balance the retention of old knowledge and acquisition of new information, leveraging a diagonal Fisher Information Matrix for efficiency. Across MNIST and Fashion-MNIST with a one-hot VAE base, the approach yields improved generation accuracy under forgetting in both EWC and fine-tuning regimes, and it reveals the role of class characteristics in the forgetting process, underscoring the practical impact for memory-constrained continual generative learning.

Abstract

This study presents a novel approach to Generative Class Incremental Learning (GCIL) by introducing the forgetting mechanism, aimed at dynamically managing class information for better adaptation to streaming data. GCIL is one of the hot topics in the field of computer vision, and this is considered one of the crucial tasks in society, specifically the continual learning of generative models. The ability to forget is a crucial brain function that facilitates continual learning by selectively discarding less relevant information for humans. However, in the field of machine learning models, the concept of intentionally forgetting has not been extensively investigated. In this study we aim to bridge this gap by incorporating the forgetting mechanisms into GCIL, thereby examining their impact on the models' ability to learn in continual learning. Through our experiments, we have found that integrating the forgetting mechanisms significantly enhances the models' performance in acquiring new knowledge, underscoring the positive role that strategic forgetting plays in the process of continual learning.
Paper Structure (13 sections, 7 equations, 3 figures, 2 tables)

This paper contains 13 sections, 7 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Continual Learning
  4. Class Incremental Learning
  5. Machine Unlearning
  6. Generative Class Incremental Learning with Forgetting
  7. Overall Architecture
  8. Introducing Forgetting Mechanism
  9. Generative Class Incremental Learning
  10. Experiments
  11. Settings
  12. Results
  13. Conclusion

Figures (3)

  • Figure 1: Overall architecture and our problem setting of generative class incremental learning with forgetting.
  • Figure 2: The changes in the probability of the class are influenced by the integration of the PM (white noise) into different learning techniques when new classes are learned in various datasets. (a) and (b) show the results for the MNIST dataset, while (c) and (d) show the results for the Fashion-MNIST dataset. The vertical axis indicates the classes subject to forgetting, whereas the horizontal axis represents the newly acquired classes. Blue, red, and black colors signify positive impact, negative impact, and absence on the data, respectively.
  • Figure 3: The qualitative differences in images generated for all classes in the MNIST((a)-(d)) and Fashion-MNIST dataset ((e)-(h)) when the proposed method is applied to EWC and fine-tuning strategies. Specifically, the forgotten class $c_{\rm f}$, denoted by the tenth class (MNIST: '1', Fashion-MNIST: 'Ankle Boot'), is highlighted with a red border, and the newly learned $c_{\rm new}$, denoted by the second class, is highlighted with a blue border. These classes are displayed in the second row from the top and the bottom row, respectively.