Title: Reassessing Continual Learning Through the Lens of Few-Shot Adaptation

Abstract:

Continual learning techniques are designed to optimize the balance between a model’s stability and its plasticity as it processes a series of tasks. Traditionally, stability—often equated with forgetting—is quantified by the model’s zero-shot performance on prior tasks, while plasticity is measured by its accuracy on the most recent task. However, this zero-shot benchmark fails to capture the full extent of a model’s capacity to preserve knowledge or rapidly adjust to new data, largely because it demands flawless recall across all encountered tasks. To address this limitation, we introduce few-shot evaluation as a more robust framework for assessing both stability and plasticity in continual learning systems. By applying this approach to task sequences in continual image classification, we uncover new perspectives on the effectiveness of established continual learning strategies. Furthermore, utilizing a new metric known as "per-shot plasticity," our few-shot evaluation reveals that incorporating "foresight" into continual learning—specifically through the meta-learning of brief sequences of future tasks—fosters a learning-to-learn dynamic throughout the task sequence.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC