Abstract
Purpose: :
Develop of an asynchronous model of the retina reproducing biological timing accuracy for a majority of ganglion cell types.
Methods: :
An asynchronous light sensor was used for light acquisition. Each pixel of this sensor generates an event as soon as the incident light reaches a particular threshold independently of the other pixels. The threshold follows a logarithmic adaptation and mimics photoreceptors gain control. It allows a sparse coding and a reduction of the redundancy present in frame-based acquisition. From this signal, we used a linear/non-linear model with an integrate and fire spike generation mechanism to reproduce retinal spikes trains. We relied on physiological data from the literature to fit the parameters of our model. The stimulus consisted into a 1s square of light sequentially translated through the receptive field of the cells. To probe spiking statistics, we used a spatially uniform random stimulation and quantified the timing accuracy as well as the variability in the number of spikes.
Results: :
By using a fast asynchronous light acquisition sensor, we showed that we could reproduce the physiological responses of the majority of ganglion cell types in the rabbit retina (ON beta, OFF beta, ON parasol, OFF parasol, ON delta, OFF delta, ON bistratified, OFF coupled, ON-OFF direction selective and local edge detector). More importantly, we were able to reproduce the millisecond precision critical in retinal coding. The standard deviation of the spike timings ranged from 2ms to 8ms in close agreement with physiological findings. Finally, we quantified the spiking statistics by computing the mean spike count versus its variability. We found a Fano factor of 0.195, compatible with previous observations in mammals.
Conclusions: :
We showed that the millisecond precision of retinal spike trains could be reached by using an asynchronous light acquisition device. We therefore provide an alternative to computationally intensive image processing in the design of artificial retinas. In particular, our approach is well suited for real time applications of retinal models as retinal prosthetics and optogenetic rehabilitation in blind patients.
Keywords: retinal connections, networks, circuitry