To Make Computers Think Like Brains, We Need Circuits That Forget
Every year, computers get closer to "thinking" like humans. Instead of following commands by rote, artificial intelligence can learn. Instead of processing one task at a time, neural networks use synthetic neurons to do many things at once. And soon, instead of overwriting old data every time they learn something new, the circuits of the future could gradually forget the information they no longer need.
Remember To Forget
You might think that if learning is essential to our survival, the fact that we forget information is a bug in our biological software. But imagine if every time you recalled something new, you had to contend with all the old information it made obsolete. What if when you had to find your brand new car in the parking lot, your brain had to sort through every car you've ever owned to find the right one? It would take ages to do anything. Instead, it just takes you a couple of strange experiences looking for your old car before your brain gets the hang of looking for your new one, and then you eventually "forget" to look for your old car altogether.
But artificial intelligence doesn't work like that. Right now, it's simple enough to train an AI to perform a task, but pretty difficult to re-train it to perform a new task. Depending on the technology it uses, the system has a tendency to keep everything it's ever learned (as if every car you've owned has the same level of importance). Or the system might accidentally forget old, but important information (I own a sedan. What is a hatchback?). If it could incorporate new information as it comes up while gradually forgetting the old stuff that it no longer encounters, AI could be a lot more efficient and practical. That's the goal behind a new technology from Purdue University that the researchers call "organismoids."
Breathe In, Breathe Out
As described in a Nature Communications paper in August 2017, the researchers created organismoids out of a ceramic "quantum material" called samarium nickelate. When the researchers exposed that material to hydrogen gas — or as co-author Shriram Ramanathan puts it, when it "breathes in hydrogen" — the electron from each hydrogen atom breaks off and attaches to the nickel within the material. That dampens the material's ability to conduct electricity, but only temporarily. Take the hydrogen away, and the material conducts electricity again. By carefully tuning the way the material breathes the hydrogen gas, researchers can change the way it accesses data. They can help it learn and forget like a brain.
"If I see certain information on a regular basis, I get habituated, retaining memory of it. But if I haven't seen such information over a long time then it slowly starts decaying," co-author Kaushik Roy said. "So the behavior of conductance going up and down in exponential fashion can be used to create a new computing model that will incrementally learn and at same time forget things in a proper way."