RECAP: Local Hebbian Prototype Learning as a Self-Organizing Readout for Reservoir Dynamics

Researcher Heng Zhang has introduced RECAP (Reservoir Computing with Hebbian Co-Activation Prototypes), a novel, bio-inspired machine learning strategy for robust image classification. The core innovation is its departure from standard backpropagation, instead coupling an untrained, high-dimensional 'reservoir' of neurons with a self-organizing readout mechanism. This readout uses a local Hebbian-like rule—'neurons that fire together, wire together'—to incrementally build class-specific prototype matrices from co-activation patterns in the reservoir's response. The system operates online and performs inference through simple prototype matching.

RECAP's performance was validated on the MNIST-C benchmark, a dataset designed to test robustness against common image corruptions like blur, noise, and fog. Remarkably, the model maintained high accuracy across these diverse corruptions without ever being exposed to corrupted data during training. This demonstrates an emergent robustness derived from its architecture, not from exhaustive data augmentation. The approach aligns more closely with hypothesized brain mechanisms, where local plasticity and high-dimensional population coding contribute to stable perception, offering a compelling alternative to the global error signals used in most modern deep learning.