Computational and theoretical understanding of regulatory mechanisms shaping natural vision

Project Summary/Abstract How is our visual system capable of making sense out of the complex pattern of light received in the retina? Machine vision has recently been successful at solving complex tasks on natural images such as object classification. Thus, we can use these models to infer computations which transform the pattern of light intensities driven by a natural image all the way up to the behavioral output such as the category of objects in the image. However, the models of vision offered by machine vision don’t provide insight into the cellular and molecular mechanisms in the visual system. On the other hand, classical models of visual system offer biologically faithful accounts for multiple phases the visual system goes through during development, from before eye-opening to adulthood. However, their scope is often very limited and they cannot provide functional accounts for the cellular and molecular mechanism they model. In this proposal, my aim is to put large-scale models of object recognition under biological constraints to be able to understand the computational and functional roles of those constraints. To this end, I focus on a theoretically- tractable aspect of optimization problems that both natural vision and machine vision face: regularization. Regularization refers to the parts of optimization goal that subject the mapping between input and output to some constraints usually related to resources e.g., energy efficiency in biological systems or robustness to input noise in machine vision. I hypothesize that regulatory mechanisms in natural vision have a fundamental computational role in shaping the visual cortex rather than the mere maintenance and stability roles they are often attributed to. To test this hypothesis, I aim to conduct a series of computational, theoretical and eventually experimental steps (in collaboration with experimental labs). In Aim 1, I build large-scale models of the visual cortex and I train them under different regularization terms, notably cellular-level regularization constraints mimicking neuronal self- regulatory processes. I then assess those models under a battery of functional and brain similarity measures. In Aim 2, I build a theoretical framework to gain a fundamental understanding of how these regulatory mechanisms relate to each other. Finally, in Aim 3, I develop experimental protocols to validate the predictions made by Aim 1 and Aim 2. Based on my preliminary results, I hypothesize that retinal spontaneous activity can play a significant regulatory role with functional implications for natural vision and I envision a new use for retinal prosthetic devices as valuable experimental tools to study visual development.