Speaker: Prof. Heiko Neumann
         CS Department, Neuroinformatik,
         Ulm University, Germany
         and Center for Adaptive Systems, BU CNS Department
         heiko.neumann@uni-ulm.de

Date:    Monday, February 27, 2006
Time:    4:00 pm
Place:   Room MCS 135, 111 Cummington Street
         (for directions, see www.cs.bu.edu/colloquium)

Title:   A Neural Model of Motion Computation:
         From Features to Moving Surfaces


Abstract:

Local motion detection is ambiguous due to, e.g., the aperture problem
for intrinsically 1-dimensional luminance patterns. Localized
intrinsically 2-dimensional structures often contribute robust
features to track. However, when multiple objects move and partially
occlude each other, junctions generated by intrinsic as well as
extrinsic surface boundaries suggest incorrect movement directions.

We propose a neural model of motion detection and integration that
solves the motion aperture problem by disambiguating normal flow at
extended boundaries through cortical reentry. In the motion pathway
mechanisms of modulatory feedback realize a filling-in process of
motion responses along surface outlines. In the form pathway junction
configurations of arbitrary order will be processed and represented by
multiple boundary mechanism acting on different scales. Segmentation
of motion signals at occlusions is controlled by interactions between
these form representations and the motion pathway to generate
representations of coherently moving forms. The mechanisms replicate
neuroscience data and are also demonstrated to process real-world
data.

Host: Margrit Betke