Multi-Scale Visual Features Improve Continual Learning in Developmental AI

A new study introduces multi-scale structural features for visual recognition within a developmental, gradient-free learning framework. The framework, which learns discrete topological models through local variation and selection, inherently supports continual learning without replay buffers or task boundaries. Previous work on shape recognition was limited by feature expressivity; the new encoding captures edges, contours, and spatial relations across scales, improving accuracy while maintaining comprehensibility and knowledge reuse. The paper is available on arXiv.