Multi-Scale Structural Features for Continual, Comprehensible Visual Recognition in a Developmental Learning Framework

This paper introduces a multi-scale visual representation for a developmental, gradient-free learning framework. The representation encodes edges, contours, and their spatial relations, and is combined with improved network refinement and readout mechanisms. On class-incremental MNIST, the authors report substantially higher accuracy than the earlier visual representation, matching or exceeding replay- and regularisation-based baselines at comparable storage. The method processes samples one at a time, stores no past data, requires no predefined task boundaries, and is designed to preserve responses to earlier classes while retaining human-interpretable structure. The reported evidence is limited to a controlled two-dimensional shape-recognition setting.