Lambda-based Threshold Adaptation for Accuracy Improvement in Spiking Neural Networks

초록

Spiking Neural Networks (SNNs) are promising for neuromorphic hardware due to their event-driven and energyefficient computation. However, fixed thresholds often cause excessive spiking and reduced accuracy. We propose a lambda (?)- based adaptive threshold mechanism that dynamically adjusts neuron thresholds during inference: when a neuron fires, its threshold offset increases by a small constant ? to suppress redundant spikes and then gradually decays to restore sensitivity. The model is implemented in Verilog HDL with fixed-point arithmetic using a compact 784-300-10 architecture of two fully connected and two LIF layers. Experiments on the MNIST dataset demonstrate that γ-adaptation improves accuracy and energy efficiency over fixed-threshold baselines, significantly reducing spike activity. Parameter sweeps of γ, initial threshold, and decay constants reveal clear trade-offs between accuracy, latency, and energy. Overall, ?-based threshold adaptation provides a simple yet effective approach for improving both accuracy and power efficiency in hardware-friendly SNN classifiers suitable for resource-constrained systems. © 2026 IEEE.

키워드