Article
Open Access
Energy-efficient spiking neural network implementation for a retinal prosthesis
1 Department of Electrical Engineering and Computer Science, Université de Sherbrooke, Sherbrooke, Canada
2 3IT Interdisciplinary Institute for Technological Innovation, Université de Sherbrooke, Sherbrooke, Canada
3 Département d’informatique, Université du Québec à Montréal, Montréal, Canada
  • Volume
  • Citation
    Besrour M, Lavoie J, Omrani T, Ménard J, Ranjbar-Koleibi E, et al. Energy-efficient spiking neural network implementation for a retinal prosthesis. Neuroelectronics 2025(1):0003, https://doi.org/10.55092/neuroelectronics20250003. 
  • DOI
    10.55092/neuroelectronics20250003
  • Copyright
    Copyright2025 by the authors. Published by ELSP.
Abstract

The quest for visual rehabilitation via retinal implant technology remains a complex, multi-dimensional endeavor. Retinal implants are biomedical devices surgically introduced into the ocular region. They propose a novel treatment for degenerative retinal pathologies. One of the most challenging aspects is replicating the retina’s natural scene encoding. Many investigative teams have consistently pursued strategies to overcome this intricate challenge. Still, no one has effectively generated a device mirroring 20/20 vision. This paper presents a proof-of-concept framework that models the computational circuitry of the human retina using spiking neural networks (SNNs) and implements the model on a mixed-signal application-specific integrated circuit (ASIC) employing leaky integrate-and-fire (LIF) neurons. The proposed neuromorphic system on chip (NeuroSoC) demonstrates energy and area efficiency, achieving an energy consumption of approximately 25 pJ per synaptic operation and operating within an active silicon area of about 1 mm square. A detailed characterization of the chip is provided, including measurements of energy consumption per inference, an inference time of 1.264 ms, and validation over process, voltage, and temperature variations. Additionally, a Python-based emulation framework derived from chip measurements is developed and integrated with machine learning techniques, yielding a post-training quantization accuracy of 80.3% over 4-bit resolution on a synthetic retinal dataset modeled after Parasol ON retinal ganglion cells using the CIFAR-10 dataset. These quantitative performance metrics underscore the effectiveness and impact of our approach for potential retinal implant applications.

Keywords

neuromorphic circuit; spiking analog CMOS neuron; machine learning; spiking neural networks (SNN); edge ML; biomedical retinal implant; retinal ganglion cells

Preview