To the cynic, an electronic component that mimics the human brain’s biological circuitry doesn’t exactly sound like a great leap forward. But the latest advance in neuromorphic engineering could eventually lead to computers that quickly and accurately perform complex perceptual tasks, such as recognizing objects by sight.
Developed jointly by researchers at Lucent Technologies Inc.’s Bell Labs and the Massachusetts Institute of Technology (MIT), the circuit is designed to replicate — on a very modest scale — the brain’s hybrid analogue/digital processing structure. Based on a silicon chip, the circuit consists of 17 transistor-based neurons (compared to billions of neurons inside a brain) that communicate with each other via artificial synapse pathways. When simultaneous electric currents are applied to two artificial neurons, the circuit can be configured to respond to only one stimulus and suppress its response to the other.
“It’s like when you decide to bite into an apple rather than into your hand,” says Rahul Sarpeshkar, assistant professor of electrical engineering and computer science at MIT. As in the brain, there is no single element in the circuit that decides which stimulus to suppress. The decision is the result of a consensus among a majority of the neurons.
Sarpeshkar, who led the research team with H. Sebatian Seung, an MIT assistant professor of computational neuroscience, points out that getting an electronic circuit to act like a brain is no small feat. “Like a digital computer, our brains make a yes/no decision on whether we are looking at an apple or a hand. Yet the decision is based on a stream of analogue information about the object’s size, shape and spatial relations to other objects.” It’s by receiving the information and accepting or rejecting specific inputs that the brain (or a neuromorphic circuit) can arrive at the logical, and usually correct, choice.
Although neuromorphic research is still at an experimental stage, Sarpeshkar foresees a day when such circuits will help computers, mobile phones, cameras and a wide array of other devices better cope with sounds, images and other real-world information. “Anything that utilizes sensory data, such as speech recognition, phone noise reduction and biometric security products, could benefit from this technology,” he says.
Sarpeshkar believes neuromorphic circuits are destined to replace digital-signal processors and other conventional ways of rapidly processing sensory data.
“People often deride the way humans think,” he says, “yet there’s actually a lot that we can learn from the process.”
