Author ORCID Identifier

Date Available


Year of Publication


Degree Name

Master of Science in Electrical Engineering (MSEE)

Document Type

Master's Thesis




Electrical and Computer Engineering

First Advisor

Dr. Luis Sanchez Giraldo


There are two encoding schema present in simple cells in the early visual system of vertebrates: the retinal simple cells activate highly when the receptive field contains a center surround stimulus, while the primary visual cortex’s (V1) simple cells activate highly when the receptive field contains visual edges. Work has been done in the past to enforce constraints on visual machine learning such that the retinal or V1 encoding is learned, but this work is often done to emulate retinal and V1 encoding in a vacuum. Recent work using convolutional neural networks focuses on anatomical constraints along with a supervised objective for training the network to explain the emergent representations of retina and V1 in vertebrates. The model dismisses observations made by other models of retinal processing where robustness to noise and coding efficiency are considered. Moreover, the use of a convolutional architecture explicitly enforce spatial equivariance in the features, which can limit the emergence of other relevant features. Here, we explore a more flexible model. We propose the EVSNet, a fully-connected neural network which learns retinal and V1 features. To analyze the representations learned with this network, we propose a measure called the orientedness to quantitatively discern expected retinal features from expected V1 features.

Digital Object Identifier (DOI)