I forgot —

Memristor that forgets makes a good model neuron

New device mimics how neurons alter behavior once they're activated.

Memristor characterization silicon at Hewlett Packard Labs.
Enlarge / Memristor characterization silicon at Hewlett Packard Labs.
Hewlett-Packard Enterprise

While the human brain isn't particularly quick at handling complex calculations, it performs a variety of tasks, such as image analysis, far more accurately than any computer. And, when it comes to energy efficiency, a brain beats a traditional computer with ease. Which is why it's somewhat ironic that all the original attempts to model a brain were forced to use software running on traditional computers, since that was all we had.

Recently, however, there have been a variety of attempts to build hardware that acts more like a collection of neurons than anything that Apple or Intel have designed. While some of that uses traditional silicon-based transistors, other efforts have explored a relatively newer development, the memristor. Now, an international collaboration that includes everyone from Hewlett Packard Enterprise to the Air Force has designed a memristor that behaves a bit more like a neuron: its recent activity influences how it responds. This is accomplished by allowing metal to diffuse within the solid memristor.

The idea behind the new design was apparently inspired by our understanding of neurons. In many cases, the activity of a neuron isn't only set by the signals it's receiving right at that instant. Instead, it has the biochemical equivalent of short-term memory. If it's received signals in the recent past, it's easier to activate that neuron by an additional signal. Over time, that memory fades, and the neuron goes back to its normal level of responsiveness.

Now, it's undoubtedly possible to provide an artificial neuron with a small bit of memory and circuitry to read it, all of which could provide similar behavior. But this team managed to do it using just a single memristor. The secret was a clever bit of materials science.

A memristor is a device with two states that it can be switched between. In one, it provides high resistance to current traveling through it; in the alternate state, it allows current to pass more freely. In this case, the authors fabricated a memristor using a silicon-oxygen-nitrogen material (the precise composition varied, and they also tested a hafnium oxide). To carry the current in the "on" state, they embedded some silver nanoparticles.

This led to the sort of behavior you'd expect from a memristor. When sufficient current was applied to the device, the heat generated by its resistance allowed the silver to diffuse within the solid, forming small "wires" that connected the two ends, allowing current to pass easily.

The intriguing thing is what happens once the current is switched off. Normally, you'd expect that things would stay in the low-resistance state. But something called (and I am not making this up) Ostwald ripening occurs.

When there's a bunch of small particles embedded in a solid, the energetics of that depends on how compatible the two materials are. If they're not very compatible, it becomes energetically favorable to minimize the contact between them. That means fewer, larger particles. And, because solids can (to a limited extent) diffuse within each other, that's exactly what happens. Smaller particles shrink out of existence, as their component atoms diffuse over and join larger ones.

For this memristor, this is precisely what happens to the silver. Instead of remaining in longer, wire-like structures, it quickly condenses back to a small collection of larger particles through Ostwald ripening (why yes, I will use that term as often as I possibly can).

This has a very specific effect on the device. If signals in the form of current have transited it within the last few seconds, it'll still be in the "on" state, and the current will travel with low resistance. Longer than that, however, and Ostwald ripening ensures that it goes back to the off state, meaning that any additional signals face high resistance. This also sets an activity threshold for the memristor, as it has to receive more than one signal in quick succession before it will become active—again, a bit like a neuron.

The authors of the paper describing this device spend a lot of time arguing that this behavior specifically mimics how neurons handle the calcium they use to help trigger activity. That's probably overstating it, as neurons are quite a bit more complex than a single electronic device. Still, there's no question that this device acts a lot more like a neuron than most typical circuitry does. And the authors have managed to make them using a variety of materials, so it's apparently robust behavior.

Whether this will end up useful enough to end up in a neural computer isn't clear at this point. But I'm always enthused at having more potential tools available to anyone who's trying to assemble something like that. It's impossible to know in advance which combination of hardware approaches is eventually going to prove compelling.

Nature Materials, 2016. DOI: 10.1038/NMAT4756  (About DOIs).

Channel Ars Technica