Understanding the role of feedback in vision systems.
Abstract
Feedback is a ubiquitous anatomical feature of the brain but its function remains largely unknown. Visual neuroscience has been dominated by work on feedforward transformations, raising the critical question: what is the role of feedback in visual computation? Here, we pursue the hypothesis that feedback serves to implement a generative model of the visual world, incorporating priors that make it possible to resolve visual ambiguity. We propose experiments to tackle the computational role of feedback in visual recognition of a particular class of objects, faces. The macaque face patch network constitutes the currently best-understood system for high-level object representation, and the unique spatial organization of this system allows us to target specific feedback pathways. In Aim 1, we will build a Convolutional Neural Network with Feedback that augments a feedforward CNN with a feedback generative network. This network will provide the computational framework for the experimental Aims. In Aim 2, we will record responses of neurons in face patches ML and AM to faces degraded in various ways. In Aim 3, we will use cutting-edge tools to target specific populations of feedback neurons for electrophysiological recording. Finally, in Aim 4, we will inactivate specific feedback projections in the face patchsystem and observe consequences on face representation as well as on a monkeys ability to recognize degraded faces. We believe these experiments will lead to new insights into the function of feedback in the primate visual system, as well as development of a new state-of-art artificial system for visual inference.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Oct 19, 2020
- Source ID
- N000142012786
Entities
People
- Doris Tsao
Organizations
- California Institute of Technology
- Office of Naval Research
- United States Navy