Neuromorphic Models of the Visual System for Multichannel, Spike based Encoding and Processing
Abstract
The mammalian retina is a complex neural tissue composed by hundreds of millions of specialized neurons. Even though its architecture is simpler than that of the brain cortex, it processes a variety of visual signals by optimally filtering and extracting relevant environmental information through complex computing. Visual signals processed by the retina are then transmitted to the brain in the form of spikes through several parallel channels -the neuronal axons- forming the optic nerve, where they are further processed in a parallel and hierarchical way, at increasing levels of complexity. Recent research has shown that the number of information channels emerging from the retina, defined by the number of different Retinal Ganglion Cells types, is much larger than previously thought. While functionally characterized, they are still not fully understood, and defining why so many distinct channels are present is still an open question. Building upon a set of artificial neural networks (ANN) models accounting for the encoding process of nearly 40 types of Retinal Ganglion Cells (RGCs), we studied the response of each ANN model -we termed Neuromorphic Neural Circuits (NNCs)- to different images. In doing so, we characterized how much and what information is shared between these NNCs representing the variety of RGCs. We found that despite our NNCs being trained with artificial images, it is capable of replicating some of the expected behavior of the retina when stimulated with natural images. Even though there is considerable evidence supporting the notion that the mammalian visual system is evolutively tuned to optimally encode natural images, our work is the first one to perform a direct comparison between artificial and natural images using the complete set of retinal outputs.
Document Details
- Document Type
- Technical Report
- Publication Date
- Apr 22, 2021
- Accession Number
- AD1137268
Entities
People
- Tomas Perez-Acle