Jeremy Freeman
"Strange shapes of the unwarped primal world glided to and fro before his passive eyes."
-- Herman Melville
Open your eyes. You immediately become aware of a coherent world full of complex visual patterns. You see faces, letters, textures, animals. You do it easily and rapidly. How?
The visual image hitting the retina is initially decomposed into simple elementary features, akin to oriented lines and edges, which are detected in early visual cortex. What happens next? No one knows. Higher visual cortex is organized hierarchically, and we believe that some hierarchical computation integrates or transforms simple, local features into progressively complex representations. The computation must be both selective (we can recognize the letter A as different from the letter B) and invariant (we can recognize the letter A whether it's big small, upside down, or right side up). It's a highly demanding task, and the brain has developed a remarkable solution. We want to understand that solution. I study these topics from both physiological and computational perspectives. How does the brain analyze and interpret complex visual input? And can we build models that capture the computation that the brain is performing? In particular, what is the importance of hierarchy and feedback? I use a combination of behavioral, neuroimaging, and computational methods, including machine learning and Bayesian statistical modeling. I am particularly interested in applying techniques from machine learning to the analysis of neuroimaging data.
My recent focus is a phenomenon known as "crowding": If you view multiple objects (like letters) in the periphery, and they are close together, they get jumbled up; you can't recognize them. Try it! Fixate the +.
It is easy to identify the r on the left but impossible to identify the r on the right. The flankers, t and y, spoil recognition of the target. During crowding, the brain fails to integrate features into an object, so it is a window into the integration computation. We have studied the role of attention in crowding, and the relationship between crowding and reading rate. Most recently, we have measured neural responses to crowded and uncrowded letters at different levels of visual processing, and we have found that crowding disrupts interactions (correlations) between multiple levels of visual processing. This provides novel evidence for the hierarchical nature of the recognition computation.