“Neuristor”: Memristors used to create a neuron-like behavior
The solid-state device has an output that looks like neural activity spikes.
Computing hardware is composed of a series of binary switches; they're either on or off. The other piece of computational hardware we're familiar with, the brain, doesn't work anything like that. Rather than being on or off, individual neurons exhibit brief spikes of activity, and encode information in the pattern and timing of these spikes. The differences between the two have made it difficult to model neurons using computer hardware. In fact, the recent, successful generation of a flexible neural system required that each neuron be modeled separately in software in order to get the sort of spiking behavior real neurons display.
But researchers may have figured out a way to create a chip that spikes. The people at HP labs who have been working on memristors have figured out a combination of memristors and capacitors that can create a spiking output pattern. Although these spikes appear to be more regular than the ones produced by actual neurons, it might be possible to create versions that are a bit more variable than this one. And, more significantly, it should be possible to fabricate them in large numbers, possibly right on a silicon chip.
The key to making the devices is something called a Mott insulator. These are materials that would normally be able to conduct electricity, but are unable to because of interactions among their electrons. Critically, these interactions weaken with elevated temperatures. So, by heating a Mott insulator, it's possible to turn it into a conductor. In the case of the material used here, NbO2, the heat is supplied by resistance itself. By applying a voltage to the NbO2 in the device, it becomes a resistor, heats up, and, when it reaches a critical temperature, turns into a conductor, allowing current to flow through. But, given the chance to cool off, the device will return to its resistive state. Formally, this behavior is described as a memristor.
To get the sort of spiking behavior seen in a neuron, the authors turned to a simplified model of neurons based on the proteins that allow them to transmit electrical signals. When a neuron fires, sodium channels open, allowing ions to rush into a nerve cell, and changing the relative charges inside and outside its membrane. In response to these changes, potassium channels then open, allowing different ions out, and restoring the charge balance. That shuts the whole thing down, and allows various pumps to start restoring the initial ion balance.
In the authors' circuit, there were two units, one representing the sodium channels, the other the potassium channels. Each unit consisted of a capacitor (to allow it to build up charge) in parallel to a memristor (which allowed the charge to be released suddenly. In the proper arrangement, the combination produces spikes of activity as soon as a given voltage threshold is exceeded. The authors have termed this device a "neuristor."
As it currently stands, the NbO2 neuristor uses too much power to put in large numbers on a chip. But there are other types of Mott resistors known, and the authors think that it should be possible to find one that's both low power and compatible with current chip-making techniques. They suggest there's a variety of ways the spiking behavior would be useful in existing applications. But I'm more intrigued with the idea that it might be possible to get more neuron-like behavior directly on a chip.