Abstract: Perception is the process of inferring properties
of the world from sensory information. Noise and ambiguities in
the sensory signal typically lead to uncertainty in interpreting this
information. Perceptual systems, either biological or artificial,
require efficient strategies to deal with this uncertainty.Focusing
on the perception of visual motion, I will describe an analog network
architecture that can robustly compute the visual motion of an object
despite sensory noise and the ambiguities imposed by the aperture
problem. The network’s dynamics guarantee an asymptotic approximation
of the statistically optimal solution to the problem, which is
continuously updated for new spatiotemporal visual input. Finally, I
will demonstrate an implementation of this network architecture in
standard CMOS semiconductor technology, resulting in a sensor for
real-world real-time applications that shows perceptual characteristics
similar to human visual motion perception.
- This event has passed.