SCIENCE & ENGINEERING NEWS
Princeton, N.J. – Concluding an unusual intellectual contest, a Princeton scientist has revealed the principles behind a computer model of a mouse brain capable of recognizing spoken words.
Neuroscientist John Hopfield created the brain simulation several months ago based on a theory he developed about how the brain interprets sensory perceptions, from touch to hearing. Hopfield did not, however, publish his insights immediately. Instead, he made the simulation available on a Web site and, in September, issued a challenge to colleagues to deduce the principle behind it.
“We wanted to provoke neuroscientists into thinking about thinking,” said Hopfield, who developed the brain simulation and contest in collaboration with Carlos Brody of New York University.
Hopfield and Brody brought the contest to a conclusion on Dec. 14, announcing a first-place winner from Cambridge University and a second-place winner from the California Institute of Technology. The two described the principle in detail in a paper to be published later this winter in the Proceedings of the National Academy of Sciences. An uncorrected proof of the paper is being made available to journalists early to coincide with the end of the contest.
The question behind Hopfield’s challenge is a critical one for neuroscience: How does the brain recognize patterns in the sensory inputs it receives. The problem is particularly difficult for inputs that arrive over a period of time, such as spoken words or the sensation of touching a familiar object.
Based on years of investigation, Hopfield concluded that performing such feats requires brain cells to be very sensitive to the timing with which they fire off electrical signals to one another. The conventional view has been that networks of neurons respond only to broad differences in firing patterns – a slow series of electric spikes means one thing and a quick burst means another. Hopfield noted such a system is time consuming – it requires several signals to make a good reading just as it takes 10 or 15 seconds to take a person’s pulse.
If neurons could respond to the timing of individual spikes rather than just averages of many spikes, they could perform faster and more sophisticated calculations, he reasoned.
Hopfield and Brody used this insight to create a computer program that models sensory perceptions in the brain. Their highly simplified organ contains only 800 neurons compared the billions in a human or mouse brain. Yet after a single instance of saying the word “one” into a microphone, the model was able to make reliable distinctions between subsequent readings of the same word and samples of other words, even closely related words such as “ton.” The device was able to recognize the word “one” even when it was masked by extraneous noise or when it was said faster or slower.
Although there is no direct evidence that Hopfield’s model depicts what actually happens in human brains or those of other animals, he believes it is likely to be very close. “I think we are going to find that the way this simulated brain computes is an example of the way a whole lot of real organisms compute,” he said.
For one thing, he said, all the “parts” of the simulated brain are modeled after actual biological entities observed in real brains. “It doesn’t require any new hardware and it doesn’t require any gimmicks in the hardware,” he said.
Also, the system replicates and explains results from previous experiments on human subjects monitored by electroencephlogram (EEG). Scientists had observed that when subjects performed certain types of decision making, unique “signatures” appeared in the EEG readouts. The same signatures would emerge from a brain that is wired using Hopfield’s principles, he said.
Hopfield dubbed the simulated organism the Mus silicium or “sand mouse” and posted it to the Web at http://neuron.princeton.edu/~moment . Visitors to the site had access to a host of “experimental” data about the organism’s anatomy and electrophysiology. They could even run their own experiments by submitting sound files and observing the results.
The first researchers to deduce the principles behind the simulation were in a group led by David MacKay of Cambridge University. Second place went to Benjamin Rahn, a graduate student at the California Institute of Technology. “Both of them clearly showed how a deductive process could lead to the answer, and only to this answer,” said Brody.
Hopfield and Brody will award each winner $500 and a Handspring Visor handheld computer signed by Jeff Hawkins, inventor of the Palm Pilot and founder of both Palm Computing Inc. and Handspring Inc. Hawkins has long been interested in the brain, and when he learned about the competition, he volunteered to fund the prizes. The runners-up each receive $200 and a Visor.
Hopfield emphasized that beyond the fun of the contest, he and Brody had a serious objective in issuing their challenge. Too few biologists, he said, attempt to deduce broad biological principles from the experimental facts at hand. Indeed the constant quest for new data can distract researchers from the important job of fitting the facts together in a coherent picture.
“When you are trying to answer a question, you either think about the facts you have or you go out and get new ones,” he said. “The temptation always is to go out and get new facts, but you may have what you need to solve the problem right away.”
If the response to the contest is any measure, Hopfield and Brody appear to have made their point. When a description of Hopfield’s challenge appeared in the Oct. 3 New York Times, the site had hits from 17,000 different visitors over the following week, with 4,000 people downloading the research paper describing the sand mouse experiments.
“We are delighted,” said Hopfield. “We had no idea that so many people would be interested in a scientific paper.” “We expected to chiefly engage computational neuroscientists,” said Brody, “but judging by the Web hits and the e-mail we have received, we engaged a much more diverse group-mathematicians, computer scientists, physicians, even high-school students. All these people have now been drawn into thinking hard about neurons and the brain.”