E-SKAN: Breaking the Efficiency-Accuracy Frontier in Neuromorphic Computing via Event-Driven Kolmogorov-Arnold Networks

Abstract

Spiking Neural Networks (SNNs) offer a promising path toward energy-efficient AI, but they tradition- ally require large parameter counts to match the accuracy of conventional networks. Kolmogorov-Arnold Networks (KANs) provide interpretable, parameter-efficient representations through learnable spline functions, yet their continuous computation requirements seem fundamentally incompatible with the discrete, sparse nature of SNNs. We introduce E-SKAN (Event-Driven Spiking Kolmogorov-Arnold Networks), a novel architecture that bridges this gap. Our key insight is that synaptic traces decay slowly, enabling us to skip redundant spline recomputations when trace changes fall below a threshold δ. This restores computational sparsity to the KAN framework. On MNIST, E-SKAN achieves 97.94% accuracy with 24% fewer parameters (179K vs 235K) compared to baseline SNN. On N-MNIST (neuromorphic event-based data), E-SKAN achieves 94.00% accuracy with 40% fewer parameters (375K vs 626K). Our validation confirms delta-gating witha mean trace change of 0.02, well below the δ = 0.05 threshold. E-SKAN represents the first architecture to simultaneously improve accuracy and parameter efficiency over standard SNNs on both static and event-based neuromorphic data.