Sample Compression for Continual Learning

2025-03-13

AI Summary

This work addresses catastrophic forgetting under non-stationary data distributions in continual learning. We propose Continual Pick-to-Learn (CoP2L), the first framework to systematically integrate sample compression theory into continual learning. CoP2L constructs task-adaptive compression sets—compact, representative subsets of each task’s data—that provably bound generalization error while enabling efficient storage and incremental updates. The method unifies three key components: (i) a Pick-to-Learn mechanism for selecting informative samples, (ii) dynamic construction of compression sets across tasks, and (iii) optimized experience replay with per-iteration quantification of generalization loss. Evaluated on multiple standard benchmarks, CoP2L consistently outperforms conventional experience replay baselines, achieving both strong theoretical guarantees—via formal generalization bounds—and superior empirical performance in accuracy, memory efficiency, and adaptability.

Technology Category

• Machine Learning: Life-Long and Continual Learning
• Search and Optimization: Learning to Search
• Constraint Satisfaction and Optimization: Constraint Learning and Acquisition

Application Category

• Economics, Online Markets and Human Computation: Social networks and social learning
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Practical large-scale studies of user experience

Abstract

Continual learning algorithms aim to learn from a sequence of tasks, making the training distribution non-stationary. The majority of existing continual learning approaches in the literature rely on heuristics and do not provide learning guarantees for the continual learning setup. In this paper, we present a new method called 'Continual Pick-to-Learn' (CoP2L), which is able to retain the most representative samples for each task in an efficient way. The algorithm is adapted from the Pick-to-Learn algorithm, rooted in the sample compression theory. This allows us to provide high-confidence upper bounds on the generalization loss of the learned predictors, numerically computable after every update of the learned model. We also empirically show on several standard continual learning benchmarks that our algorithm is able to outperform standard experience replay, significantly mitigating catastrophic forgetting.

Problem

• Addresses non-stationary training distribution in continual learning
• Provides learning guarantees for continual learning algorithms
• Mitigates catastrophic forgetting in continual learning benchmarks

Innovation

• Efficient sample retention for tasks
• High-confidence generalization loss bounds
• Outperforms standard experience replay