Neurons for Robots


It’s no surprise, perhaps, that humans are still better at being humans than humanoid robots. Aside from the flesh and blood, the ability to grow, and that strange thing called consciousness, there’s one big reason we still beat bots at our own game: We’ve got synapses; they’ve got ones and zeros.

Until now, that is. Researchers at Stanford have created the Neurogrid, an array of analog chips that process information stochastically rather than with the precision of the usual digital computation. “The approach to making this work actually comes from the brain—the brain basically operates with analog systems,” says Kwabena Boahen, a bioengineering professor and the principal investigator for Stanford’s Brains in Silicon program. “We’re not relying on any one neuron, we’re distributing computation over thousands of neurons.”

Like the brain, all that analog computation needs a handful of watts. Where a desktop computer needs 100 watts to run a few gigaflops, the Neurogrid needs just a few watts to run teraflops. A digitally dependent robot that might perform with anything like human sophistication would need a crippling amount of juice. Take Honda’s Asimo. It could move, recognize objects, and react to them, but never for more than an hour before it needed a charge. For a robot, low power means freedom.

Asimo, a humanoid robot designed and developed by Honda. Image: Wikimedia Commons




Robots Teaching Humans

But the Neurogrid is more than a tool to give us less power-needy, more humanoid robots. It’s also likely to bring us a better understanding of how the humans that those robots emulate do their processing.

“Our neurons are analog systems that mimic neurons in [the] brain. They communicate with each other using spikes,” says Samir Menon, a Ph.D. student who runs the motor control aspects of the Neurogrid. Under certain conditions, when it gets to a communication threshold, a neuron creates a spike, a buzz that hits the neurons it’s next to. If we organize the neurons in a pattern that matches something we’re interested in then the neurons will do something useful.”

For Menon, that something useful means directing a robotic arm to move as efficiently and as humanly as possible. To understand the complexity of the problem, says Menon, imagine trying to simply pick up a cup from a table. After identifying the location of the cup in three coordinates, you move the 40 or 50 muscles in your arm to send your hand smoothly and precisely to that spot. “All of us would take the least energy to get the cup,” Menon says. “This is a major problem to solve.”

His solution uses the opposite of the brute force calculation that usually goes into computation. Taking measurements of the distance between hand and cup, he created an artificial spike stream sent to the neurons of the Neurogrid. “Many people, when studying neural networks, use complicated learning rules to force the network to do what they want it to do,” he says. “They keep fiddling with neurons until they do what they want them to do. But these learning rules are biologically implausible. They require neurons to communicate with other neurons very efficiently, and that doesn’t happen very often in the brain.”

Neurogrid emulates a million neurons across sixteen neurocores. Image:




Neuron’s the Limit

For a long time, researchers assumed that the incredible variation in size and shape of neurons in the brain presented a major problem that had to be somehow overcome. “It turns out this is not a problem at all, but one of the best features that allows us to do cool things,” says Menon. “Because there are so many neurons and they’re all different, by sheer chance some will be doing the right thing.” The trick is just identifying the neurons that are behaving correctly.

“We can take any position of the arm and convert it to how much each neuron likes it. But these neurons can have different thresholds; they can have different gains, different rates at which they fire,” says Boahen. “You can get this mixed bag of neurons. That’s how we represent information.”

As soon as they get their arm-moving model fully in action, the Neurogrid’s functionality will be extended to other parts of the body and mind. “At a fundamental level, we’re interested in engineering nervous systems with all the complexity found in biology,” says Menon. “And we think that will give us some insight into the problems the brain is solving.”

Michael Abrams is an independent writer.

Because there are so many neurons and they’re all different, by sheer chance some will be doing the right thing.

Samir Menon,
Stanford University


August 2014

by Michael Abrams,