V1-LSM: Biologically Structured Liquid State Machines for Event-Based Vision

Liquid state machines (LSMs) are well suited to event-based vision because they process time-varying spike streams through recurrent spiking dynamics. However, standard LSM reservoirs are usually random and homogeneous, unlike the structured organization of the primary visual cortex (V1). We propose V1-LSM, a modular reservoir for N-MNIST classification that introduces coarse cortical structure through retinotopic hypercolumns, layered intra-column processing, and sparse lateral coupling between neighboring columns. Each column implements a layered processing pathway: Layer 4 for input gating, Layer 2/3 as the recurrent liquid, Layer 5 for stable state extraction, and Layer 6 for gain modulation. A linear readout is trained on the resulting column-level features, while the reservoir itself remains fixed. In matched-neuron experiments on the N-MNIST dataset, the V1-inspired model improves test accuracy from 86.45\% to 91.67\% at 512 neurons and from 88.23\% to 92.37\% at 1024 neurons. These results suggest that coarse cortical organization can provide a practical inductive bias for event-based reservoir computing.