Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Continual Learning for Adaptive AI Systems

Md Hasibul Amin, Tamzid Tanvi Alam

TL;DR

This work tackles catastrophic forgetting in continual learning by proposing Cluster-Aware Replay (CAR), a hybrid method that couples a small, class-balanced replay buffer with a feature-space regularization term called Inter-Cluster Fitness ($L_{ ext{ICF}}$). CAR explicitly shapes the latent space to promote inter-task separation, rather than relying solely on parameter or logit constraints. On the standard 5-task Split CIFAR-10 benchmark with a ResNet-18 backbone, CAR demonstrates improved retention of earlier tasks compared to naive fine-tuning and aligns with basic regularization baselines, while highlighting the need for adaptive regularization as task complexity grows. The results suggest that geometry-aware feature-space regularization is a promising direction for mitigating forgetting in continual learning, meriting further exploration and scaling to larger benchmarks.

Abstract

Continual learning the ability of a neural network to learn multiple sequential tasks without catastrophic forgetting remains a central challenge in developing adaptive artificial intelligence systems. While deep learning models achieve state-of-the-art performance across domains, they remain limited by overfitting and forgetting. This paper introduces Cluster-Aware Replay (CAR), a hybrid continual learning framework that integrates a small, class-balanced replay buffer with a regularization term based on Inter-Cluster Fitness (ICF) in the feature space. The ICF loss penalizes overlapping feature representations between new and previously learned tasks, encouraging geometric separation in the latent space and reducing interference. Using the standard five-task Split CIFAR-10 benchmark with a ResNet-18 backbone, initial experiments demonstrate that CAR better preserves earlier task performance compared to fine-tuning alone. These findings are preliminary but highlight feature-space regularization as a promising direction for mitigating catastrophic forgetting.

Continual Learning for Adaptive AI Systems

TL;DR

This work tackles catastrophic forgetting in continual learning by proposing Cluster-Aware Replay (CAR), a hybrid method that couples a small, class-balanced replay buffer with a feature-space regularization term called Inter-Cluster Fitness (). CAR explicitly shapes the latent space to promote inter-task separation, rather than relying solely on parameter or logit constraints. On the standard 5-task Split CIFAR-10 benchmark with a ResNet-18 backbone, CAR demonstrates improved retention of earlier tasks compared to naive fine-tuning and aligns with basic regularization baselines, while highlighting the need for adaptive regularization as task complexity grows. The results suggest that geometry-aware feature-space regularization is a promising direction for mitigating forgetting in continual learning, meriting further exploration and scaling to larger benchmarks.

Abstract

Continual learning the ability of a neural network to learn multiple sequential tasks without catastrophic forgetting remains a central challenge in developing adaptive artificial intelligence systems. While deep learning models achieve state-of-the-art performance across domains, they remain limited by overfitting and forgetting. This paper introduces Cluster-Aware Replay (CAR), a hybrid continual learning framework that integrates a small, class-balanced replay buffer with a regularization term based on Inter-Cluster Fitness (ICF) in the feature space. The ICF loss penalizes overlapping feature representations between new and previously learned tasks, encouraging geometric separation in the latent space and reducing interference. Using the standard five-task Split CIFAR-10 benchmark with a ResNet-18 backbone, initial experiments demonstrate that CAR better preserves earlier task performance compared to fine-tuning alone. These findings are preliminary but highlight feature-space regularization as a promising direction for mitigating catastrophic forgetting.

Paper Structure

This paper contains 17 sections, 3 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Literature Review
  3. Regularization-based Approaches
  4. Rehearsal (Replay-based) Methods
  5. Generative Replay
  6. Parameter Isolation and Dynamic Architectures
  7. Geometry-aware Approaches
  8. Proposed Methodology
  9. Architectural Backbone
  10. Inter-Cluster Fitness Function ($L_{\mathrm{ICF}}$)
  11. Overall Loss
  12. Experiments
  13. Setup
  14. Results and Discussion
  15. Visualization
  16. ...and 2 more sections

Figures (4)

  • Figure 1: Final per-task accuracy after all five tasks (ours) with a typical fine-tuning reference level.
  • Figure 2: Average accuracy over seen tasks after each training step $T_k$ (from Table \ref{['tab:results_over_time']}).
  • Figure 3: Task-wise forgetting: peak accuracy per task minus its final accuracy after $T_5$.
  • Figure 4: Representative training convergence on one task.