Quantized Spiking Neural Networks on FPGA: An Application to Retinal Prosthetics

Marwan Besrour, Samuel Bouchard, Romit Mohanty, Jacob Lavoie, Takwa Omrani, Berhtié Gouin-Ferland, Konin Koua, Audrey C. Therrien, Mounir Boukadoum, Réjean Fontaine

Abstract: We present an embedded Spiking Neural Network (SNN) in a Field-Programmable Gate Array (FPGA) for retinal prosthetic applications. Our primary goal is to minimize resource utilization, making the solution suitable for edge AI systems with neural-machine interfaces. The proposed quantized SNN minimizes computational resources for a power-efficient system. The PRANAS open-source software, which emulates retinal ganglion cells (RGCs), is used to generate a spiking dataset from popular MNIST database and the developed SNN is trained on it by the error backpropagation and surrogate gradient technique. We subsequently applied weight quantization on the trained SNN’s weights, using both 4-bit and 8-bit precision, to create a compact and power-efficient version of the SNN. The resulting accuracies after the 4-bit and 8-bit post-training quantizations are 83.2% and 87.2%, respectively. They show the potential of FPGA-based implementations of quantized SNNs to offer high-performance and energy-efficient solutions for retinal implants, thus answering their inherent hard power and area constraints.