Retinomorphic vision systems embody these retinal principles. Their function is determined by top-down constraints, based on information theory, and their structure is determined by bottom-up constraints, based on VLSI design principles---the same forces that, no doubt, shaped the highly evolved piece of tissue that lines the eyeball. Paying attention to structure as well as function results in energy- and area-efficient sensor designs that are ideally suited to perceptive human-made systems, because, unlike video cameras, these retinomorphic designs go beyond simply reproducing the scene, to extracting salient information in real time.
I will talk about vision chips that I designed to model the structure and function of the vertebrate retina. I have measured and analyzed the spatiotemporal dynamics of these retinomorphic chips, and compared the results with measurements from the biological counterparts. By combining synthesis and analysis, I have identified tradeoffs in sensory-system design. These tradeoffs yield insights into how the retina simultaneously optimizes structure and function, and provide a unified view of outer and inner retinal processing. My work has also advanced the state of the art in focal-plane image processing. I will conclude my talk by describing how we can incorporate retinomorphic chips into multichip neuromorphic systems; I will illustrate this procedure by showing a video demonstration of a two-dimensional stereo vision system.
HOSTS: Profs. J. Guttag and J. Shapiro
|
Modified: Jun 24, 1997|
Current events|
Your comments
and inquiries are welcome.