Accuracy-Efficiency Trade-offs of Temporal Enrichment in Spiking Neural Networks

Abstract

Spiking neural networks (SNNs) promise energyefficient temporal processing, but the choice of neuron model affects both task accuracy and hardware cost. We compare three mechanisms for enriching temporal processing in SNNs- heterogeneity, adaptation, and synaptic delays-on auditory benchmarks (SHD, SSC) under controlled conditions, analyzing per-neuron arithmetic cost, state storage, and total energy. Our key finding is that constrained adaptive LIF (cAdLIF) neurons in a feedforward topology outperform recurrent heterogeneous LIF networks while eliminating recurrent connections, the dominant contributor to synaptic energy. Parameter-efficiency sweeps show that cAdLIF networks with $4 \times$ fewer parameters exceed the LIF accuracy ceiling at any size, producing 30% fewer total spikes at matched accuracy. Post-training quantization confirms that 10bit fixed-point incurs <1% accuracy loss on both benchmarks, and RTL synthesis (Yosys + SKY130) shows that, at matched 256$\times$2 size, feedforward cAdLIF is $2.4 \times$ smaller in silicon area than recurrent LIF, despite its higher per-neuron cost.