Overcoming device unreliability with continuous learning in a population coding based computing system

Published

September 25, 2018

Author(s)

Alice C. Mizrahi, Julie Grollier, Damien Querlioz, Mark D. Stiles

Abstract

Nanodevices have promising features for novel computing implementations, but also drawbacks like unreliability. The brain offers an example of a computing system working with unreliable components: neurons and synapses exhibit stochastic behavior and die regularly. Computing schemes inspired from biology may be able to take advantage of the positive features of nanodevices while allowing for their drawbacks. In previous work we have proposed a such a system called population coding based on a particular artificial neuron, the superparamagnetic tunnel junction. Here we show that this system is resilient to the catastrophic loss of neurons, as well as to the loss of synaptic information. Furthermore, we show that using unreliable components can lead to an interesting trade-off between power consumption and precision.

Citation

Journal of Applied Physics

Volume

124

Issue

15

Pub Type

Journals

Download Paper

https://doi.org/10.1063/1.5042250

Keywords

spintronics , neuromorphic computing , magnetic tunnel junction , population coding , stochastic computing

Citation

Mizrahi, A. , Grollier, J. , Querlioz, D. and Stiles, M. (2018), Overcoming device unreliability with continuous learning in a population coding based computing system, Journal of Applied Physics, [online], https://doi.org/10.1063/1.5042250 (Accessed January 2, 2025)