Implementation of Visual Motion Detection in Analog "Neuromorphic" Circuitry-A Case Study of the Issue of Circuit Precision

Many animals rely on visual motion to infer their relation to the surrounding environment while moving around in it. Motion processing in biological nervous systems, particularly in the insects, has been a subject of active research for over 50 years. With the advent of interest in "neuromorphic" integrated circuits in the late 1980s, this mode of sensory processing has also been the subject of various analog silicon modeling efforts. The author discusses the background of certain models for motion detection in insects, and then the implementation of a particular model in analog integrated circuitry. The paper does not focus on such circuits themselves, however, but on the issue of computational precision in the analog domain: it is argued that insufficient attention to this issue has been an impediment to the development of neuromorphic circuits that are practically useful in engineering applications. A statistical approach is described for assessing one of the designs discussed herein, and the degradation of its performance due to variations in the electrical properties of its constituent devices, with an eye toward suitability for application in autonomous robotics. This approach relies on tools and models routinely used in conventional analog/mixed signal design. The author argues that adoption of such techniques, along with a deeper consideration of computational precision, i.e., a shift in the neuromorphic design philosophy, would be an important step in moving the field forward.