An effort towards a unified theory of high and low level perception

From Santa Fe Institute Events Wiki

Revision as of 18:29, 19 September 2007 by Mm (talk | contribs)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)

Much of cognition and perception can be reduced to two goals: discrimination and generalization. In this talk, a general account is presented that attempts to reduce all such tasks down to a single geometric framework . The approach describes the non-linearities of neurons from V1 to IT as a function of the manifolds of stimuli that produce a constant response. The approach attempts to explain such effects as end-stopping, cross orientation inhibition, and other non classical surround effects, as well as invariance effects shown with complex cells, and facilitory effects demonstrated with contour integration. It is argued that these effects are best described in terms the curvature of the response manifold (curvature in state-space). Each dimension of the neuron requires one generalized curvature parameter. It is argued that invariance and generalization in neural responses can be represented by positive curvature while hyper-selectivity and contrast normalization requires negative curvature. However, for most neurons, the vast majority of dimensions show no curvature (flat response). It will be argued that such an approach can provide insights into single neurons as well as more general questions of perceptual representation.

Back to Program