Domain-specific processing - building a digital/analog neural processing system for control of intelligent robotics

Electronic neural architectures have been shown to be particularly well suited for tasks requiring high-dimensionality, low-precision and regular computational architecture. While these limitations are not serious for many sensory applications, they do not permit neural architectures as a panacea for complex computation and control. This paper discusses a “domain specific” architecture based upon a combination of specialized neural architectures combined with digital logic and computers for less structured tasks. A specific example is discussed involving a proposed system for high-speed sensory processing and control of autonomous aerial vehicles. The discussion includes the impacts of three-dimensional integration on reducing form factor and power while increasing overall performance