Active Learning for Continual Learning: Keeping the Past Alive in the Present

TL;DR

Active continual learning (ACL) aims to query informative unlabeled samples under labeling budgets while preserving past knowledge. The authors introduce AccuACL, which defines accumulated informativeness via Fisher information and uses a memory buffer to decouple past and new task information, supported by a diagonal Fisher information embedding and two theoretical properties that justify a greedy, scalable query strategy. The method balances past-forgetting prevention and rapid learning of new tasks through a distribution-based and magnitude-based scoring mechanism, achieving superior average accuracy and reduced forgetting across SplitCIFAR10/100/TinyImageNet benchmarks with favorable efficiency. The work advances ACL by providing a principled, scalable Fisher-information–based sampling framework and releasing implementation details for practical deployment.

Abstract

Continual learning (CL) enables deep neural networks to adapt to ever-changing data distributions. In practice, there may be scenarios where annotation is costly, leading to active continual learning (ACL), which performs active learning (AL) for the CL scenarios when reducing the labeling cost by selecting the most informative subset is preferable. However, conventional AL strategies are not suitable for ACL, as they focus solely on learning the new knowledge, leading to catastrophic forgetting of previously learned tasks. Therefore, ACL requires a new AL strategy that can balance the prevention of catastrophic forgetting and the ability to quickly learn new tasks. In this paper, we propose AccuACL, Accumulated informativeness-based Active Continual Learning, by the novel use of the Fisher information matrix as a criterion for sample selection, derived from a theoretical analysis of the Fisher-optimality preservation properties within the framework of ACL, while also addressing the scalability issue of Fisher information-based AL. Extensive experiments demonstrate that AccuACL significantly outperforms AL baselines across various CL algorithms, increasing the average accuracy and forgetting by 23.8% and 17.0%, respectively, on average.

Figures (5)

• Figure 1: Overview of AccuACL; (a) Unlike conventional AL strategies that only focus on the new task and cause catastrophic forgetting, AccuACL balances the prevention of catastrophic forgetting and the ability to learn new tasks quickly, by defining the accumulated informativeness; (b) shows the catastrophic forgetting of the conventional AL strategies on SplitCIFAR100.
• Figure 2: Comparison of AL strategies: (a) average accuracy of SplitCIFAR100 on ER for different labeling budget per task; (b) memory consumed for selecting 100 examples for a single task in SplitCIFAR10; (c) elapsed querying time for selecting 1000 examples for every task in SplitCIFAR10.
• Figure 3: Effect of the parameter $\lambda$ for SplitCIFAR10 on task 4.
• Figure 4: Effect of scoring measures.
• Figure 4: Comparison of AL strategies: (a) average accuracy of SplitCIFAR100 on ER throughout AL rounds; (b) forgetting of SplitCIFAR100 on ER throughout AL rounds.

Theorems & Definitions (13)

• Theorem 4.1
• proof
• Theorem 4.2
• proof
• Theorem 4.3
• proof
• proof
• Theorem B.1
• proof
• Lemma C.1