An Event-driven Recurrent Spiking Neural Network Architecture for Efficient Inference on FPGA

Spiking Neural Network (SNN) architectures are promising candidates for executing machine intelligence at the edge while meeting strict energy and cost reduction constraints in several application areas. To this end, we propose a new digital architecture compatible with Recurrent Spiking Neural Networks (RSNNs) trained using the PyTorch framework and Back-Propagation-Through-Time (BPTT) for optimizing the weights and the neuron's parameters. Our architecture offers high software-to-hardware fidelity, providing high accuracy and a low number of spikes, and it targets efficient and low-cost implementations in Field Programmable Gate Arrays (FPGAs). We introduce a new time-discretization technique that uses request-acknowledge cycles between layers to allow the layer's time execution to depend only upon the number of spikes. As a result, we achieve between 1.7x and 30x lower resource utilization and between 11x and 61x fewer spikes per inference than previous SNN implementations in FPGAs that rely on on-chip memory to store spike-time information and weight values. We demonstrate our approach using two benchmarks: MNIST digit recognition and a realistic radar and image sensory fusion for cropland classifications. Our results demonstrate that we can exploit the trade-off between accuracy, latency, and resource utilization at design time. Moreover, the use of low-cost FPGA platforms enables the deployment of several applications by satisfying the strict constraints of edge machine learning devices.