Silicon approximation to biological neuron
V. A. Gorelik · 2004
This paper presents a new approach to simulate the behavior of a biological neuron in silicon. The proposed devices have the ability to mimic a variety of structures and interconnect architectures commonly found in biological neural nets. The proposed device may be fabricated in polysilicon rather than in a single crystal substrate and thus permits multi-layer architectures. Unlike MOS and BJT-based structures, the device utilizes significantly different operating principles that are more closely related to transport mechanisms in biological neurons. The device is well suited to simulate axodendritic, dendrodendritic, axoaxonic, and reciprocal synapses. A network of such devices can be constructed to perform both spatial and/or temporal processing. Basic principles underlying the design allow multi-layered, almost zero-power neural networks on a single silicon die. One possible implementation, utilizing temporal neural circuitry for extraction and production of atomic auditory elements - phonemes in Broca's and Wernicke's areas of the cortex, is also shown.