Recurrent models of orientation selectivity enable robust early-vision processing in mixed-signal neuromorphic hardware

Abstract

Mixed signal analog/digital neuromorphic circuits represent an ideal medium for reproducing the dynamics of biological neural systems in real-time with bio-physically realistic dynamics. However, similar to their biological counterparts, these circuits have limited resolution and are affected by a high degree of variability. Considering this, we developed a recurrent spiking neural network that implements a faithful model of the retinocortical visual pathway to reliably produce Gabor-like receptive fields tuned to visual stimuli with specific orientation and spatial frequency properties. Specifically, we developed a neuromorphic visual system comprising a Dynamic Vision Sensor that emulates the transient pathway of real retinas and a mixed-signal Dynamic Neuromorphic Asynchronous Processor with adaptive exponential integrate-and-fire neurons and dynamic synapses and mapped the recurrent network model on it to produces the desired orientation and spatial frequency tuning responses. Compared to alternative feed-forward schemes, the model developed gives rise to robust highly structured Gabor-like receptive fields of any phase symmetry, optimizing the hardware resources available in terms of synaptic connections. We present experimental results using both synthetic and natural images validating the model with its hardware implementations.