Machine learning algorithms of the visual system
Dr. Zucker's group has built a bridge between theoretical models of information processing and machine learning, computational vision and computational neuroscience. Although his career began with a focus in computer vision, he is now working entirely on understanding how brains could solve vision problems. His group now works in two main areas, from early vision to intermediate levels bordering on cognitive vision, and on the mouse and the primate as physiological models. The early vision project is motivated by an analogy between differential geometry and visual cortex. It is a collaborative effort with two experimental neuroscience groups, the Field lab at Duke, which works on retina, and the Stryker lab at UCSF, which works on visual cortex. The goal is to understand the functional organization of neural circuits from the retina to the cortex, and to use these functional models to inspire direct experimental tests. For analysis we are developing novel machine learning algorithms. The result is a manifold. Each point is a neuron, and, on the manifold, nearby neurons respond similarly in time to an ensemble of flow and artificial stimuli. Such modeling, analysis, and collaboration is relevant to the Center for Neurocomputation and Machine Intelligence. Associated models of learning address how brains can infer latent variable models using Hebbian plasticity, which is also relevant to the Center for Neurodevelopment and Plasticity. The intermediate level project relates to how humans could infer shape (3D information) from image information across lighting, material, and viewing changes. Flows are again involved, but the novel aspect is using topological methods to understand the rapid, generic, and qualitative aspects of shape perception. Color perception, and its geometric properties, and implicated in material inference.