Is Our Brain the Secret to Better AI?

The human brain is the crowning achievement of humanity’s evolution. Literally. Without our brains, modern society and all of its bells and whistles would never exist. You wouldn’t be reading Dogtown Media’s news on your phone or your computer. And I wouldn’t be writing articles about researchers taking a page out of biology’s book to improve AI. Yet, that’s exactly where we find ourselves right now.

New research out of MIT this week suggests a new chip will bring us closer to computers that work like the human brain.

The Right Medium for AI “Thoughts”

Although machine learning and AI improved at record paces in the last few years, computer processors have stagnated in innovation. In short, the computing power is struggling to keep up with the technological demand.

Companies like Google, Qualcomm, and Huawei have been re-designing mobile AI chips in an effort to improve AI’s speed and capabilities. AI’s design is increasingly informed by neuroscience, and this new chip design from MIT is a testament to that statement.

Previously, trying to track down the right medium for the necessarily varying electrical signals has largely been unsuccessful; the current spreads out everywhere, risking a short circuit of the motherboard.

Brain Power

So the Boston developers used crystalline silicon and germanium that form microscopic lattices. This facilitates the movement of electricity much better, and signals are more stable and predictable. The chips, dubbed “neuromorphic computing” chips, can save up to 1,000 times more energy while running more complex machine learning and AI tasks.

What makes these chips like our brain is their elasticity in power processing. For example, your computer’s processor is processing data in a digital fashion, and neuromorphic chips process data in an analog manner. So instead of on/off switch-type programming, the intensity of the signal can fluidly vary. As a result, more information can be stored in each signal, reducing energy consumption and computing power.

We could finally have a seamless verbal translation service with this type of power. Image recognition would become commonplace, even given out as open-source software for aspiring AI developers to play around with.

Promising Potential

“This is the most uniform device we could achieve, which is the key to demonstrating artificial neural networks,” says Jeehwan Kim, lead researcher of the chip.

Kim and his team trained a neural network that recognized handwriting with a 95% accuracy rate. The “handwriting test” is standard for training new AI. Though the baseline is 97% for existing technology, the neuromorphic chip’s results are very promising.

So for now, we’ll have to continue relying on our brain’s computing power to produce these hardware and software improvements in AI. It’s safe to say AI still needs human brains right now to advance the field. But with these types of innovations, that may not last for long.