Behavioral goals drive the requirements of a system. We use high throughput training and testing of visual behaviors that are challenging even for the most advanced machine vision systems. We then test the neural mechanisms underlying performance of these tasks.
Computational models are needed to generate quantitative predictions that are then subjected to experimental testing. The field of machine learning provides a broad hypothesis space for the nature of computation in the brain. By positing direct mappings between computational architectures and their putative neural substrates, we can efficiently test model predictions.
The sheer number of neurons and their interconnections in the cortical network can be daunting. Updating the classical neural encoding approach with modern experimental and analytical techniques for studying neural populations, we ask what is the format of information carried by identified neural subpopulations within a cortical stage and across stages of the cortical hierarchy. In this way, we hope to build, piecewise, a detailed picture of the information transferred across the cortical network.
Our research platform centers around the common marmoset, an emerging animal model in neuroscience. Our goal is to develop forward-looking infrastructure to help accelerate the overall progress of this animal model as part of the growing marmoset research community.