An Analog VLSI System for Stereoscopic Vision investigates the interaction of the physical
medium and the computation in both biological and analog VLSI systems by synthesizing a
functional neuromorphic system in silicon. In both the synthesis and analysis of the system a
point of view from within the system is adopted rather than that of an omniscient designer
drawing a blueprint. This perspective projects the design and the designer into a living
landscape. The motivation for a machine-centered perspective is explained in the first chapter.
The second chapter describes the evolution of the silicon retina. The retina accurately encodes
visual information over orders of magnitude of ambient illumination using mismatched
components that are calibrated as part of the encoding process. The visual abstraction created
by the retina is suitable for transmission through a limited bandwidth channel. The third
chapter introduces a general method for interchip communication the address-event
representation which is used for transmission of retinal data. The address-event
representation takes advantage of the speed of CMOS relative to biological neurons to preserve
the information of biological action potentials using digital circuitry in place of axons. The
fourth chapter describes a collective circuit that computes stereodisparity. In this circuit
the processing that corrects for imperfections in the hardware compensates for inherent
ambiguity in the environment. The fifth chapter demonstrates a primitive working stereovision
system. An Analog VLSI System for Stereoscopic Vision contributes to both computer engineering
and neuroscience at a concrete level. Through the construction of a working analog of
biological vision subsystems new circuits for building brain-style analog computers have been
developed. Specific neuropysiological and psychophysical results in terms of underlying
electronic mechanisms are explained.These examples demonstrate the utility of using biological
principles for building brain-style computers and the significance of building brain-style
computers for understanding the nervous system.
