Quantization-Aware Training of Spiking Neural Networks for Energy-Efficient Spectrum Sensing on Loihi Chip

Abstract

Spectrum sensing is a technique used to identify idle/busy bandwidths in cognitive radio. Energy-efficient spectrum sensing is critical for multiple-input-multiple-output (MIMO) orthogonal-frequency-division multiplexing (OFDM) systems. In this paper, we propose the use of spiking neural networks (SNNs), which are more biologically plausible and energy-efficient than deep neural networks (DNNs), for spectrum sensing. The SNN models are implemented on the Loihi chip, which is better suited for SNNs than GPUs. Quantization is an effective technique to reduce the memory and energy consumption of SNNs. However, previous quantization methods for SNNs have suffered from accuracy degradation when compared to full-precision models. This degradation can be attributed to errors introduced by the coarse estimation of gradients in non-differentiable quantization layers. To address this issue, we introduce a quantization-aware training algorithm for SNNs running on Loihi. To mitigate errors caused by the poor estimation of gradients, we do not use a fixed configuration for the quantizer, as is common in existing SNN quantization methods. Instead, we make the scale parameters of the quantizer trainable. Furthermore, our proposed method adopts a probability-based scheme to selectively quantize individual layers within the network, rather than quantizing all layers simultaneously. Our experimental results demonstrate that high-performance and energy-efficient spectrum sensing can be achieved using Loihi.