A Manifold Approach to Understanding Visual Systems

Abstract:

The data horizon in neuroscience is expanding rapidly. Large scale neural recordings are now commonplace, which raises the question of how the activity of thousands of neurons should be organized to reveal function. We study this question in the visual system, and introduce an inference procedure that yields a manifold in which each point is a neuron, and neighboring neurons respond similarly (in time) to similar aspects of a stimulus ensemble. The procedure can also be applied to artificial neural networks. Since deep networks were modeled loosely after biological visual systems, the resultant topology of the manifolds has implications for neuroAI.