External Event - 9th Sep 2021
Visual Perception of ShapeThis online event takes place on the 9th of September with Philip J. Kellman.
How we perceive and represent shape involves many of the deepest issues in understanding visual perception. In this talk, Philip considers recent progress related to 3 of these issues. 1) Abstraction in perception and representation: He describes experiments that show the abstract nature of human shape representations and the time course of their formation. 2) The relation of human perception and deep learning: He contrast abstract shape in human perception with the most successful AI approaches through experiments suggesting that deep convolutional neural networks (DCNNs), although successful in object classification, have no access to global object shape. 3) The transition from subsymbolic to symbolic encoding in vision: He will argue that understanding abstraction in human visual shape perception, and its absence in deep learning systems, require some account of a poorly understood transition: How, in visual processing, might symbolic representations be obtained from initially subsymbolic encodings? He will present experimental and modeling results that suggest an answer in perception and representation of contour shape. Evidence suggests that we encode contour shape in terms of segments of approximately constant curvature and that these initial symbolic descriptions may be based on the outputs of neural mechanisms that extract constant curvature via sets of linked orientation-sensitive units at different turning angles and scales.
