Computer Science Thesis Proposal

Monday, November 19, 2018 - 10:00am


Traffic21 Classroom 6501 Gates Hillman Centers



Modeling recurrent circuits of the primary visual cortex using neural network models

Speaker: Yimeng Zhang

Location: GHC 6501

Modeling recurrent circuits of the primary visual cortex using neural network models

There has been great interest in the primary visual cortex (V1) since pioneering studies decades ago. However, existing models cannot explain V1 neural responses to complex stimuli satisfactorily. One possible reason for this failure is the models' lack of recurrent connections, which form the bulk of synaptic connections in V1 and greatly contribute to the complexity of the visual system.

The goal of my thesis is to develop, test, and understand neural network models of recurrent circuits in V1 and the visual system in general. I have completed two studies and propose to finish an additional study.

My first study has demonstrated that the Boltzmann machine, a type of recurrent neural network, is useful for conceptualizing certain V1 recurrent computations.

My second study has demonstrated that the CNN's key components are crucial to its superior performance in explaining V1 data and are consistent with previous V1 studies. While not directly related to recurrent circuits, this project has demonstrated that neural response prediction is a useful metric for selecting models with high correspondence with biological reality. The metric and analysis methods here will be used in the proposed work.

In the final project, I propose to advance our understanding of recurrent circuits of V1 in a two-part investigation. First, I will find candidate models for V1 recurrent circuits, by designing models with recurrent computation components for predicting neural responses as well as predicting certain phenomena observed in V1 studies. The models to be explored will feature two new complementary designs: model architecture and training methodology. Preliminary results show that models with these designs can perform as well as state-of-the-art approaches using fewer parameters and less data.

Second, I will explain the role of recurrent connections in the candidate models for modeling V1 using various tools in machine learning as well as existing knowledge about V1.

Overall, this investigation will provide new neural network models with recurrent connections for explaining more V1 phenomena with higher accuracy, establish correspondence between model components and biological reality, and provide new insights about the roles of recurrent circuits in V1 and visual signal processing in general.

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

Copy of Proposal Document

For More Information, Contact:


Thesis Proposal