Notes

Mead, Carver. "Neuromorphic electronic systems." Proceedings of the IEEE 78.10 (1990): 1629-1636.

meadNeuromorphicElectronicSystems1990

Biological information-processing systems employ different principles compared to present computational technologies. For many problems, the brain is a billion times more efficient than our current digital solutions. Although transistor development will likely reduce energy consumption becoming comparable to the neuronal unit, the nervous system would remain a factor of ten million more efficient. Such a difference in energy efficiency is attributable to the organisation of devices in the system. In this paper, I argue that confining ourselves to a digital representation of the information made of bits (0 or 1) and elementary operations such as AND, OR, and NOT is inefficient. Instead, letting the physical properties of the transistors define our elementary operations provides a rich set of computational primitives, which allow for greater system efficiency. A representation of the information via analog signals takes advantage of the inherent capabilities of the device, such as the ability to generate exponentials, integrate with respect to time, and perform zero-cost addition using Kirchhoff’s law. I discuss an implementation of such computational primitives in analog integrated-circuit technology, the Mahowald retina. Such a ”neuromorphic” device shows adaptive capabilities that mitigate the effects of component differences and naturally lead to a self-organising system that learns about the environment. I argue that adaptive analog systems use silicon more efficiently, consume less power than conventional systems and show robustness to component failure. Furthermore, the two-dimensional limitation of silicon technology does not restrain large-scale implementations. Thus, this adaptive analog technology would exploit the full potential of wafer-scale silicon fabrication.