An international collaboration of researchers from UMass Amherst, HP, and the Air Force have built a proof-of-concept memristor that could lead to real-world neuromorphic chips. The memristor is made of a silicon-oxygen-nitrogen material laced with clumps of silver nanoparticles at the electrical terminals. When current is applied across the memristor, the silver nanoparticles shuffle around within their parent oxynitride matrix, to line up within a lightning-bolt-like path of least electrical resistance. The electrical field exerts enough force on particles within a given radius of that path to scoot them subtly into line in the memristor’s ON state. The heat released by the electricity both permits and enables the movement.
If the memristor has been activated within a few seconds, it will still be in its ON state, with the silver nanoparticles still helpfully arranged into a nano-wire. This mimics short-term potentiation, which is an important kind of neuronal conditioning that we depend on for learning and memory. After a bunch of activations within a short period of time, silver drifts slowly toward one pole of the memristor and away from the other, mimicking desensitization (another form of classical conditioning) as calcium is depleted within the “tapped-out” upstream neuron. When the field is released and the memristor allowed to cool down and “rest,” those silver nanoparticles relax back to where they originally were, in discrete clumps around each electrode.
How does this all happen? It depends on a physics principle called Ostwald ripening. Ostwald ripening is a poetic five-dollar term for the tendency of two immiscible things to move so they exclude one another, leaving the smallest possible shared surface area. (The surface area matters because high relative surface area is energetically expensive.)
Shake a bottle of Italian dressing and then sit there and watch it separate. You’ll see both the aqueous part and the oil part make small bubbles and droplets, which slowly join back together, forcing oil out of water and vice versa as the dressing relaxes back to its lower-energy state of being neatly separated along one plane. It’s energetically disfavored for oil and water to combine. Adding energy to the system by shaking the bottle can push the fluids closer to being evenly dispersed. But it can’t really overcome the fundamental tendency for the fluids to draw back together, away from the immiscible other.
The same principle is at work on the silver nanoparticles embedded in the slice of semiconductor. When the applied current imparts energy, the nanoparticles can get caught up in the current and be pulled away from their low-energy residence near the electrical contacts into the form of the more highly ordered nanowire. As soon as the circuit is closed, this process starts. But a single pulse doesn’t apply enough energy to assemble the conductive nanowire; repeated activation forges the bridge, and makes the resistance across the memristor plummet. Break the circuit and, after a short lag, resistance across the memristor begins to rise back to its original level. It’s a nanomachine that self-assembles the silver wire under the right voltage conditions and then self-disassembles when conditions revert.
The time the memristor takes to relax and recover its original resistance is a function of “voltage pulse parameters, operation history, Ag concentration, host lattice, device geometry, humidity and other factors,” the researchers wrote, which could be used alone or in combination to tune the performance of a neuromorphic chip based on memristors of this type.
Neurons have a refractory period after they fire off an action potential, because firing sucks up ions from the fluid bath that then have to be replenished by diffusion, which is slow. But the way these memristors work mimics how neurons handle calcium ions with respect to time. Calcium ion movement is important not just for the firing of an action potential, but because it controls the release of neurotransmitters into the synaptic cleft.
Memristors are a way to abstract ideas about how the brain works. This is important because when building an interface between neurons and silicon, we’re going to need a Rosetta stone: a translating interface, a layer of abstraction that permits easy discourse between these two very different environments. While a memristor that mimics how neurons handle calcium won’t be sufficient to bring forth a fully functional neuromorphic architecture, it’s easy to see how this principle could be extended to other things that happen at synapses. Making that work in the physical world, though, would depend on having a solid, comprehensive model of the way neurons work, which is yet to be articulated.
Now read: How neuromorphic ‘brain chips’ will begin the next era in computing