Neocortical Architectures for Neuromorphic Computation

Abstract

Animal brains are much more effective in dealing with real-world tasks than even the most advanced computers. In vertebrates the neocortex is very likely the sub-system of the vertebrate brain most relevant for intelligent and effective interaction with the world, and is one region where we can hope to understand the relationship between neuronal architecture and the computation that it supports. Fortunately, the evidence shows that cortex has a surprisingly uniform architecture. Indeed, since the fundamental work of Gilbert and Wiesel on the neuronal circuits of visual cortex, it has seemed likely that the basic architecture and operation of cortex can be understood in terms of relatively few types of excitatory and inhibitory neurons. This simplification has encouraged neuromorphic engineers to develop systems that emulate some attributes of cortical processing in an electronic medium, hybrid CMOS VLSI technology. Successes include circuits that provide rapid solutions of constraint-satisfaction problems such as coherent motion and stereo-correspondence; networks composed of a variety of neurons and synapses; and methods for asynchronous event-based communication between analog computational nodes distributed across multiple chips. These working chips and systems, combined with the results of cortical research, are providing insights into novel methods of processing that use hybrid, asynchronous, event-driven computation; and co-localization of computation and memory.

About Rodney Douglas

Director of the Institute of Neuroinformatics at the ETH Zuerich,
http://www.ini.unizh.ch/

Current research interests include; experimental anatomy and physiology of visual cerebral cortex; theoretical analysis and simulation of cortical circuits; design and fabrication of neuromorphic systems; use of analog Very Large Scale Integration methods to construct electronic circuits that perform analogous signal processing and computational functions to biological neuronal networks; the development of neuromorphic robots that use analog VLSI chips for sensory, motor, and higher order processing; self-constructing simple behaving organisms; and self-constructing neural networks.