In this thesis, I develop, test, and understand neural network models of recurrent circuits for V1 and early visual areas in general in a series of three studies. In the first study, I trained a Boltzmann machine using 3D natural scene data and established the consistency of the learned connections of such a recurrent network to the functional connectivity of neurons measured in neurophysiological experiments, and thus showing that the recurrent connectivity of the brain reflects statistical priors of the natural scenes. In the second study, I compared feed-forward convolutional neural networks (CNNs) with other popular models for predicting V1 neural responses and isolated the key components underlying the superior performance of CNN models. In the third study, I introduced recurrent circuits to the CNN and showed that recurrent models provided better predictive performance and were more data efficient compared to feed-forward models of comparable configurations. The learned recurrent models could reproduce a variety of contextual modulation effects observed in the visual cortex. To understand the computational advantage of the recurrent models, I proposed a new conceptualization of the recurrent network as a multi-path ensemble model and I established that compared to feed-forward models, multi-path ensembles as implemented by recurrent models can be more flexible and data efficient in learning and approximating complex computations as those in the brain.

186 pages

Thesis Committee:
Tai Sing Lee (Chair)
Abhinav Gupta
Robert E. Kass
Alan L. Yuille (Johns Hopkins University)

Srinivasan Seshan, Head, Computer Science Department
Martial Hebert, Dean, School of Computer Science

Return to: SCS Technical Report Collection
School of Computer Science

This page maintained by reports@cs.cmu.edu