Biologically motivated analog VLSI systems for optomotor tasks

Abstract

For machines to interact with their environment, they have to solve problems which have traditionally been "difficult" for von Neumann based machines. On the other hand, biological organisms manifest solutions to these problems in their everyday existence. Therefore, the benefits of biological computational models must be incorporated with high speed electronics to produce the computers and robots of the future. Furthermore, analog VLSI offers an efficient and compact substrate for these computational systems. Vision provides enormous information about the environment. However, to extract this information, vast amounts of computations are required. Biological visual systems offer a method to reduce the computational band-width through hierarchization using computational sensors and parallel distributed processing. Therefore, a computational sensor for image motion extraction and parallel processing using a general purpose analog neural computer is presented. Furthermore, applications of these system to real world problems requiring real-time solutions are also discussed. Many image understanding algorithms require the a priori knowledge of the image motion. However, image motion determination is a difficult problem. A sensor which computes motion at the imaging plane would greatly improve the speed of these algorithms. Traditional approaches for VLSI motion detection have been mostly unsuccessful due to the incompatibility of these techniques with compact VLSI circuits. Hence, a new image motion algorithm designed for analog VLSI implementation is presented. The primary contributions of this dissertation are: (1) the amalgamation of biological and computational models to develop an image motion estimation algorithm for focal plane VLSI implementation, (2) the consideration of hardware limitations in the development of the algorithm makes it implementable in VLSI, (3) as a result of the hardware restrictions, a compact software implementation is also possible, (4) the implementation of a motion computational sensor, which has multiple processing layers similar to biological retinas, (5) the application of the motion sensor to a single chip target tracking system based on the smooth pursuit mechanism, and (6) the development of a general purpose analog neural computer and its application to cortical motion estimation using oriented spatiotemporal filters. The design, analysis and results for all these systems are presented.