Mimicking neurones in a chip

A phase-change memory chip that learns to recognize handwritten numbers by simulating a network of neurons is tested in IBM’s Almaden Research Center near San Jose, California. Credit: IBM

IBM announced a year ago Synapse, a chip that could mimic one million neurones and their connectivity (in a much limited scale) using over 5 billion transistors. It was tested in some activities that are typically well performed by a brain and poorly executed by a computer and indeed it showed very interesting results, including a two orders of magnitude faster analyses of image content using 100,000 times less power than an equivalent, classical, chip performing the same activity.

However, Synapses is missing a crucial property of a brain. It does not change over time: it is like having a "frozen" brain, very good in what it is good but unable to learn. If you want to change its behaviour you need to reprogram it.

Now IBM research centre in conjunction with Pohang University in South Korea has announced a much less powerful chip in terms of simulated neurones, mimicking 913 neurones with 165,000 synapses based on phase change memories, but much better in other respects.

The overall architecture is completely different from the Synapses chip but the most striking difference is that the use of phase change memories allows the chip to "learn" as it is stimulated by incoming data (I am intentionally using the word "stimulated" rather than processing because it no longer performs a mechanical processing).

The researchers stimulated the chip with 5,000 hand-written numbers and that trained the chip to recognise hand-written numbers. They tested the chip by showing other hand-written numbers and they got a very good recognition rate (over 80%). That rate can be increased to 99% by a bit of tweaking of the chip. This would be similar, or even better, than what our brain can do (ever tried to read your doctor's prescriptions? Not sure how doctors around the world are writing their prescriptions, in Italy they are close to be unreadable....).

This feat is possible by leveraging the properties of phase change memories. These memories are particularly good for neuromorphic computers since they can store data  so densely that can support more complex synaptic connections. They are also easier to reprogram and this can make learning by the chip possibile: the chip itself can adjust its processing as data flow through it.

It is not the first time that researchers are trying to use phase change memories in neuromorphic computers but so far the number of synaptic connections implemented was below a 100. The IBM and Pohang University chip has 1,000 more synapses and this was possible because the researchers found a way to compensate for errors that occur as data are stored in such dense memory grid. This also takes place in our smartphones but the error rate, and thus the recovery issue, is much lower.

It is good to open the new year with this news of progress in neuromorphic computing, a further step towards the "singularity".

Author - Roberto Saracco

© 2010-2018 EIT Digital IVZW. All rights reserved. Legal notice