QMTS: Fixed-point Quantization for Multiple-timescale Spiking Neural Networks

Spiking Neural Networks (SNNs) represent a promising solution for streaming applications at the edge that have strict performance and energy requirements. However, implementing SNNs efficiently at the edge requires model quantization to reduce memory and compute requirements. In this paper, we provide methods to quantize a prominent neuron model for temporally rich problems, the parameterized Adaptive Leaky-Integrate-and-Fire (p-ALIF). p-ALIF neurons combine the computational simplicity of Integrate-and-Fire neurons, with accurate learning at multiple timescales, activation sparsity, and increased dynamic range, due to adaptation and heterogeneity. p-ALIF neurons have shown state-of-the-art (SoTA) performance on temporal tasks such as speech recognition and health monitoring. Our method, QMTS, separates SNN quantization into two stages, allowing one to explore different quantization levels efficiently. QMTS search heuristics are tailored for leaky heterogeneous neurons. We demonstrate QMTS on several temporal benchmarks, showing up to 40x memory reduction and 4x sparser synaptic operations with little accuracy loss, compared to 32-bit float.