Breakthrough in neuromorphic computing

A prototype chip with large arrays of phase-change devices that store the state of artificial neuronal populations in their atomic configuration. The devices are accessed via an array of probes in this prototype to allow for characterization and testing. The tiny squares are contact pads used to access the nanometer-scale phase-change cells (inset). Each set of probes can access a population of 100 cells. There are thousands to millions of these cells on one chip and IBM accesses them (in this particular photograph) by means of the sharp needles (probe card). Credit: IBM Research

Our neurones are not "digital machines" they flip from one state to another when certain thresholds are reached but these are not fixed, as they are in a computer chip. In a way neurones are not predictable at micro level. They are more likely comparable to analogue devices then to digital ones. Besides a bit in a computer chip flips to 1 from 0 whenever an appropriate signal requesting it to flip to 1 is received. In the case of a neurone ... it depends on its previous history. A neurone is both a micro storage of the past events and a "state" participating to a global "state".

Of course, the boundaries between a digital and an analogue device are somewhat "fuzzy". Even in a digital device like a computer chip, if we were to go down to electrons level, we would see that sometimes a "zero" will flip into a "one" when -say- one million electrons have moved, whilst another time the flip occurred after one thousands electrons moved. However, in a digital device the thresholds are much better defined than in an analogue device.

At the neurone level this fuzziness gets even more pronounced by the fact that stimuli (signals) are coming from other neurones connected by a tangled web. The fuzziness of a single neurone gets compounded by the ones of all other connected neurones.

Simulation of neurones through digital chips is therefore just an approximation.  Researchers at IBM Zurich have made a leap in the approximation of a neurone behaviour by using a phase changing material chip, a memristor, that can simulate the way a real neurone spikes. The goal is to be able to create chips mimicking the neural circuits of a brain, with the same energy efficiency and the same computation power.

Not necessarily such chips will be comparable to a brain in terms of "intelligence". Scientists are envisioning smart sensors that can provide a better understanding of their environment, smarter IoT.

IBM researchers have used phase changing materials, like the ones used in BlueRay disc, that similarly to real neurones have the property of "remembering" the past and process new signals taking the past experience into account.

The idea is to combine these artificial neurones with artificial synapses to create processing and storage units that would operate differently from present computers, based on a von Neumann architecture. The IBM researchers have already managed to assemble hundreds of these artificial neurones to process complex signals.

Neuromorphic computing is one of the possible "futures" of processing in the next decade, along with quantum, molecular computing and massively distributed micro computations.

Author - Roberto Saracco

© 2010-2019 EIT Digital IVZW. All rights reserved. Legal notice. Privacy Policy.

EIT Digital supported by the EIT