Building a brain
UW prof constructs world’s largest simulation of a human brain
Technology developed by a University of Waterloo professor is behind a brain-like computer chip that could advance artificial intelligence and be used in applications such as controlling prosthetics.
Chris Eliasmith, a professor in philosophy, systems design engineering and computer science, has partnered with researchers at Stanford University to produce a neuromorphic chip.
Neuromorphic chips mimic the way human brains process information, and solve problems. Intel, IBM, HP and Qualcomm are all pursuing this technology.
Eliasmith, director of UW’s Centre for Theoretical Neuroscience, built the world’s biggest, functional model of a human brain. He calls it “Spaun.” It is a simulated network of 4.5 million neurons that imitates the way brain cells collect and process information.
Neural networks are behind leading edge technology such as driverless cars.
Each of the neuromorphic chips he’s making with Stanford will have the computing power of one million neurons. He wants to make the chips available to engineers to spark widespread innovation in what’s expected to be the next generation of computer hardware.
“We think that letting developers play with this kind of computation is important, to work it into their systems and understand how it works,” said Eliasmith.
That neuromorphic chip, like the human brain, will be energy-efficient. With 80 billion to 100 billion neurons, the human brain requires 15 to 20 watts of power to operate — about the same consumption as an efficient fluorescent bulb.
UW and Stanford’s neuromorphic chip uses about a thousand times less power than traditional graphics processing units (GPUs) that help a computer’s central processor stream videos and pictures, said Eliasmith.
“Which means that if you put them inside the skull they won’t burn the cortex it is on top of, which a normal chip would,” he said.
That’s essential if the chips are to be attached to human brains for controlling prosthetics.
Research and simulations show the technology works in principle, making it possible to control advanced prosthetics with the chips attached to brains.
“Then you want to program that chip to interpret the neural activity that it’s connected to in order to control a robot arm, or whatever you have that you are trying to control with this system.”
The neuromorphic chip should be available to developers and researchers in less than a year.
“Once we have the real chip, then we should be able to start actually closing the loop, using it with monkeys, showing that we can decode their neural activity,” said Eliasmith.
Even with 4.5 million neurons, a tiny fraction of the total in a human brain, his model demonstrates the huge potential of the technology.
One area the technology can be applied to is artificial intelligence. AI mainly uses advanced pattern recognition in order to do one task — play chess, recognize the face of a terrorist among airline passengers, weld vehicle frames in a factory, clean an office floor and carry materials from one point to another, as examples.
But neural networks learn on the fly, and adapt to changing environments.
Spaun recognizes a thousand different categories of objects. It mostly collects information through a camera. It draws pictures with an arm in response to questions.
“It can recognize a ton of different kinds of dogs, scenery and ships,” said Eliasmith.
“You say things like: ‘If you see a hockey puck, then draw the number eight.’ Then you show it stuff, it sees a hockey puck and it draws the number eight. … You can give it different instructions at different points in time. You just told it to do a new task, and then it can do that kind of task.”
With this ability, neural networks and neomorphic computing have potential to greatly expand the field of artificial intelligence.
“If we can understand how brains perform some computation, which we can’t build machines to do, maybe we can extract those principles and implement them in an artificial system,” Eliasmith said.
“The brain is performing a whole bunch of computations, and it does so in a very different way than our computers do.”
Eliasmith started building Spaun years ago, to better understand the normal functioning of the human brain, and the diseases that afflict it.
“The brain is a very poorly understood organ, so it makes it difficult to treat,” he said.
About five years ago, Spaun needed 2 ½ hours to do what the human brain computes in one second; today, it needs 12 seconds.
It’s a “big improvement,” said Eliasmith. “This let’s us add to it all the time, constantly making it bigger.”
The neuromorphic chips he is working on with Stanford will match the human brain for processing speed.
“And these chips are real time,” he said. “They are guaranteed real time. Those million neurons will run real time all the time, that’s all they can do.”
When many neuromorphic chips are used at the same time, computing power will be greatly increased. The advances will pave the way for robots so sophisticated the machines will be able to carry on a conversation with people, Eliasmith said.
His technology is based on mimicking brain-like functions. A neural network that truly mimics the human brain must have what Eliasmith calls spiky neurons, massive parallelism and asynchronous communication.
Brain cells communicate with muscles, glands and other brain cells by sending out pulses of electricity. Those electrical spikes are replicated in Spaun.
Brain cells send and receive those electrical signals whenever they are needed, and not according to a pre-set clock that synchronizes the entire system. That’s what makes Spaun, and human brain asynchronous, said Eliasmith.
With 80-to-100 billion neurons, the human brain can use a huge number of cells when processing information. So a neural network needs to use thousands or millions of computer cores running at the same time. This is what Eliasmith means by massive parallelism.
“There are definitely lots of commercial applications, if you can do computation a thousand times more efficiently,” he said.
Eliasmith co-founded a company called Applied Brain Research to commercialize technology breakthroughs coming out of Spaun. In a basement laboratory on the UW campus is a new kind of robot arm the startup is developing that is using the same motor system as Spaun.
It is programmed to do repetitive tasks. But it also gathers new information through a camera, processes the information and adapts to changes in real time. When it runs into a person or another object it stops, because it knows there is something in the way.
Most robot arms in use today power through a task, hurting or damaging anyone or anything that gets in its way.