TAI-SING LEE
Associate Professor, Computer Science and Neural Basis of Cognition
www

Research in my laboratory seeks to elucidate the computational principlesand neural mechanisms underlying visual perception.  We use computational,mathematical and neurophysiological experimental techniques to address thefollowing fundamental issues in the fields of computational and biologicalvision.

Statistical and ecological approaches to higher order neural codes. The study of neural representation should start with a rigorous study ofthe statistical regularities in the visual environment. I have adopted theGibsonian ecological approach in our quest to understand how the brainrepresents and computes 2D curves and 3D surfaces. We developed databasesof 2D and 3D natural scenes and applied statistical and machine learningtechniques to discover the statistical structures in the data. This hasallowed us to predict plausible neural representations based on theprinciple of efficient coding.  We are currently carrying outneurophysiological experiments to evaluate the various possible candidatesfor the neural representations of mid-level vision (2D shapes and 3Dsurfaces). We have developed methods to discover how neurons encodeinformation and to decode visual stimuli based on neural responses. We areexperimenting with the implantation of microelectrode arrays that canrecord from hundreds of neurons simultaneously. Our long term goal is todecode and reconstruct computationally the mental images that arerepresented in our visual cortex.

Learning, adaptation and development in neural systems. Learning and adaptation are what make biological systems so much morerobust and powerful than current man-made vision systems. My currentresearch explores the principles underlying adaptation and learning in thevisual system at different time scales and at the level of neurons and ofneural systems.  We have studied theoretically how neurons adaptdynamically to the statistical context of the visual stimuli. We havedetermined biophysical features in spiking neurons that make them adapt,and have isolated the statistical features in natural stimuli that driveneuronal adaptation.  In our neurophysiological studies, we havedemonstrated that the neural machinery of perceptual processing is veryflexible and subject to modification by behavioral experience. We arecurrently undertaking computational and neurophysiological studies on thetime course and dynamical processes of neural plasticity and visualdevelopment. We hope these studies will provide insights to the designprinciples underlying the adaptation of adult visual systems and thedevelopment of infant vision.

Principles and algorithms of hierarchical perceptual inference. Perceptual inference is an active and creative process that constructs aninternal interpretation of the outside world in our mind. Bottom-upinformation from the retina is only a clue that starts the inferenceprocess, which is affected by various global scene contextual factors andperceptual experience. We have carried out a series of neurophysiologicalexperiments to demonstrate the influence of a variety of contextualfactors in shaping visual processing in the early visual areas.  We havedemonstrated experimentally that visual processing likely involves theentire hierarchical circuit interactively.  We are developing acomputational framework for perceptual inference based on hierarchicalBayesian inference to elucidate the rules of recurrent feedback incortical circuits, and to understand the computational algorithmsunderlying the inference of shapes and surfaces of visual objects in ourmind.

For further details of research in my laboratory, please visitActive Perception Laboratory web page ( http://www.cnbc.cmu.edu/apl ).