Continual Learning using Evolution Strategies

Abstract: Continual Learning (CL) aims to train neural networks on sequences of tasks without triggering catastrophic forgetting. Existing approaches typically rely on gradient-based optimization, which breaks down in exemplar-free settings where data and therefore gradients from past tasks are unavailable. To overcome this limitation, we propose EvoCL, a gradient-free method that employs an evolutionary strategy to optimize neural network using a surrogate loss constructed by an adapter network. The adapter maps stored latent features of previous classes into the current task's embedding space, enabling joint training of the feature extractor and adapter without access to past data or gradients. This reframes CL as an optimization problem that does not require gradient information. Experiments on multiple benchmarks demonstrate that EvoCL achieves strong performance under tight parameter budgets, highlighting it as a promising direction for gradient-free, data-free CL. The code to reproduce these results is available at (omitted for the review, we enclose it in the supplementary material).

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Yingbin_Liang1

Submission Number: 8659