MUSIC AND THE MIND

With a computer and some nursery tunes, a professor attempts to model the brain.

One of the stars in the film classic "2001: A Space Odyssey" is a computer named HAL. Its soft-spoken manner is reassuringly human. But its very humanity becomes a fatal flaw when the computer starts to think for itself. "I'm sorry, Dave," HAL responds to an order from the space ship's commander, "I can't let you do that." When Dave dismantles the recalcitrant computer, it regresses through its "childhood," finally singing a tired "Bicycle Built for Two."

The experiment is one of many attempts among scientists to see if computers can mimic the human brain. Researchers like Bharucha hope that their discoverwith ies will enable them to answer the original brain-teaser: how the mindthought, awareness of the environment, memory arises from the physical mass of the brain.

The mind-brain question, which has stymied thinkers at least as far back as Hippocrates, has given rise to a relatively new discipline: cognitive psychology. A field that regards the brain as a vastly complex information-processing network, cognitive psych draws on data from such diverse disciplines as biology, computer science, Unguistics, tics, philosophy, anthropology, and traditional psychology. Jamshed Bharucha was attracted by this eclecticism. After majoring in biopsychology and minoring in philosophy at Vassar as well as picking up a philosophy master's at Yale he went to Harvard for a Ph.D. in cognitive psychology. "It allowed me to combine my interest in the mind with my interest in science," he says of the discipline. "It's the best field for somebody who wants to bring the two together."

Between classes, Bharucha chips away at the mind-brain problem by making use of one of cognitive psychology's greatest research tools, the computerized neural network an electronic system roughly modeled after the brain. Although the first such networks date back to the 19505, they were ignored for many years as researchers concentrated on producing programmable serial computers that process pieces of information one at a time, the way a personal computer running a word-processing program does. Neural networks, on the other hand, spread their data out in brainlike fashion. Thousands sands of processors, each serving as an approximation of a neuron, or brain cell, form links in a constantly changing web. While serial computers must be given precise instructions to carry out a task, neural nets are capable of something remarkable: they can learn. They adjust their own connections as they adapt prior experience to novel situations. Their creators give them general rules of learning, not a set of specific knowledge. Thus a neural net's behavior is at times unpredictable and at times far more sophisticated than its better-known serial counterpart.

The machines are at the heart of a debate over how the mind works. Connectionists, as the researchers who create neural nets call themselves, base their computer models on what neuroscientists know about the human brain. Decades ago, scientists discovered that the brain's neurons—all 100 billion of them are interconnected by trillions of synapses, or gaps, between the cells. Messages speed back and forth across the gaps by means of electrochemical impulses. When a neuron becomes excited, it in turn sends impulses to other neurons. Memory, perception of the environment, thinking all are physically caused by fired-up brain cells, according to cognitive psychologists. Yet the very physicality of the brain limits the ways scientists can study it. There is only so much experimentation on the brain dubbed "wetware" by connec tionists that is both technically possible and ethically acceptable. Instead, connectionists attempt to simulate the brain in their computers.

The manmade neural networks use electronic junctions, or nodes, in place of the brain's neurons. Each node is wired either physically or through software to several other nodes, which in turn have their own webs of connections. Nodes have two jobs: they accumulate electronic signals that come to them from other nodes, and they send signals to yet more nodes. Repetition of a particular signal strengthens the connection between cells. Working on a principle of "use it or lose it," connections that are made only infrequently remain relatively weak. Eventually a neural net can discriminate criminate between expected and unexpected patterns. In this way, the network "learns" through experience.

But critics wonder whether the brain actually works the way the neural nets do. Does learning from experience account for all knowledge that people have? Or are they born with a set of rules to help process the jumble of data that surrounds them? In the 1960s the philosopher and language theorist Noam Chomsky postulated that as different as the world's languages are from each other, each has a common underlying grammar; subject and verb, he said, always form the basic structure of communication. Chomsky went on to infer that this universal grammar must be hard-wired into the human mind.

Ray Jackendoff, one of Chomsky's followers, later applied this same reasoning to music. He noted that, like language, there are musical elements common to every culture. People all over the world recognize octavesnotes spaced at regular intervals that sound like each other and can detect certain tones that are more important than others. Perhaps, Jackendoff proposed, everyone is born with a musical "grammar."

Jamshed Bharucha is skeptical. While he concedes that aspects of language grammar might indeed be innate as studies on children suggest he says this doesn't necessarily apply to music, which is less a language than an experience. He points out that octaves occur in nature: pluck a guitar string and then press your finger on the middle of it, and you will produce a note an octave higher. Even within a person's voice there are several octaves present at once. Common experience rather than a genetic blueprint may account for the world-wide awareness of octaves. Bharucha insists we can't claim that something is innate until we first confirm that it cannot be accounted for by learning.

To test the theory, Bharucha fed synthesized music converted into mathematical symbols to the Kiewit computer. And he did something that, to the uninitiated, seems even stranger: like other connectionists, he created a system that comes to expect future experiences. To him, expressions like "cultural bias" and "intuition" are technical terms, part of a vocabulary of the cloudy frontier between mind and machine, between innateness and experience, between what's hard-wired and what's acquired.

Bharucha seems in many ways an ideal person to conduct such an experiment. Having spent all 33 years of his life bridging cultures and disciplines, he is well used to making connections. The son of an American musician mother and an Indian father, he grew up in India studying classical violin and dabbling with the Indian sitar and tabla drum. At 17 he left Bombay for Poughkeepsie, New York, to program computers for IBM. Shortly after that he entered Vassar, which is where he first began seriously exploring the connections between the mind and the brain. Now, at Dartmouth, he has made the Kiewit computer undergo, in a simplified way, his own east-west musical background.

A decade and a half ago, a joke went around the profession about the grandmother neuron. If there are brain cells that fire up at the sight of a table, is there a cell sitting sweetly in a quiet corner of your brain that responds only to your grandmother? Are there, at any rate, feature detectors devoted to various parts of your grandmother her hand, her nose, her eyes, the little handbag with her initials on it? "Well," says Bharucha cautiously, "at least there might be a cell that registers handbags in general. I don't know about your grandmother's. Feature detectors probably respond only to generic objects." Researchers experimenting on monkeys have already found neurons that specifically respond to monkey faces and hands.

Connectionists think this network between specialized higher-order and lower-order neurons may allow the brain to detect objects that are partly hidden. Your brain probably contains neurons for, say, the word "sweater." Cells in the retina fire up at the lines and curves of type, which in turn excite cells that specialize in individual letters, which in their turn excite cells that specialize in the word "sweater." The processing works both ways: after the "sweater" neurons get excited, they send pulses back down the cell hierarchy, firing up the lower-order neurons even more. If a letter is missing from a word, or a piece is missing from a letter, the brain may actually perceive the whole word, thanks to the neurons firing up the cells that they are connected with. Thus you may not notice that we deliberatly misspelled a word in this sentence.

Bharucha's music-recognizing neural net functions in much the same way. He programmed nodes within the computer to recognize individual notes like first-order "cells" in the human neural hierarchy. As digitized music was played to the computer from a synthesizer, the network gradually organized itself around higherand higher-order nodes. As chords were played repeatedly to the computer, nodes specializing in the constituent notes strengthened the connections between them. Nodes excited and inhibited each other to the point where individual ones started to recognize whole chords. Eventually, some nodes even specialized in chord progressions. For example, because in western music the "C" chord typically follows the "G" chord, the network began to expect the one to follow the other. Without any prompting from Bharucha, the node for the "C" chord began to get excited even before the computer heard the "G" chord in a series of chords. Bharucha thinks a similar system within the human brain may help it work quickly, allowing it to begin processing information even before the information arrives. This, he says, accounts for the "intuition" musicians feel about their workexpectation of a future event (the next guitar riff, the perfect flatted blue note), based on prior experience.

To see if human parallels to the neural net really do exist, Bharucha played chord progressions to human volunteers, asking them to press one of two computer keys to say whether or not the chords were in tune. He found that his subjects hit the keys a fraction of a second faster when the chords were expected; they took longer to respond to an unexpected chord. Bharucha thinks this is because their brains were already attuned from years of exposure to music to hear the proper chord; an unusual cluster of notes required the brain to process the information more or less from scratch. Remarkably, the response was similar regardless of the amount of formal musical training his subjects had had. Bharucha contends that the musical environment that is part of our everyday lives the nursery songs we hear as youngsters, the constant tune barrage from car radios, the Muzak in elevators and offices and shopping malls have formed interconnections in our brains that are like a map of common chords and chord progressions.

At least, that is what happened to the computer. It similarly happened when Bharucha set up a different neural network and played it music from his native India. Unlike Western tunes, the traditional Indian musical form called raga does not contain chords as we know them; instead, the raga consists of a complex pattern of individual notes. Again, as the Indian music was played to the network over and over, the network came to expect one note to follow another. And tests on Indians matched these simulations.

Bharucha has taken the research yet another step, playing Indian music to Western students. If there is a musical "grammar" innate to everyone, the students' responses to Indian music should be similar to those of Indians to Indian music. Instead, they showed their lack of Eastern culture, thus helping to confirm Bharucha's theory that musical structures are learned rather than innate. He conducted the same cross-cultural experiment on the computer, playing Western and Indian scales to a network that was biased toward the East and one with a Western bias. The responses were similar to those of the human subjects.

Bias. It's a term that makes many people uneasy. But consider this. Can cultural bias in a computer tell us something about human prejudice? If, like the computer, the neural networks in our brains are geared to generalize from specific information, can a person look at a stranger of another race without automatically conjuring up a stereotype? Bharucha thinks that neural nets may indeed come to help us understand such social phenomena.

The professor also thinks neural nets offer implications for education. He asserts that repetition is crucial in the learning process. But traditionalists who want to bring back rote learning should beware. Bharucha and his colleagues maintain that context is also important: rehearsing a lesson in a slightly different way or location is necessary in order for the lesson to be "retrieved" by the human data-processing system at different times and places. Spacing also seems to matter: rehearsing something over time rather than in one shot makes for better memorization. As generations of students have learned to their dismay, cramming is not a good way to commit something to long-term memory.

It remains to be seen whether neuroscientists will ultimately show the brain to behave like computer neural nets. The few thousand connections in a network like Bharucha's is to the human brain's trillion synapses what a kiddie pool is to an ocean. Even the most advanced computers have come nowhere near approaching the complexity plexity of the nervous system of a fruit fly or a garden slug. And the connectionists are not at all certain that future experiments on the brain's wetware will resemble today's software. But, they Bharucha included have already shown that it is possible to make a machine act like a human: they can make it learn.

Turning a Kiewit computer into a primitive replica of the human brain, cognitive psychologist Jamshed Bharucha plays it simple tunes and watches what happens to electronic "brain cells" on the screen.

"UNTITLED": COLLECTION OF MARILYN BOGHOSIAN

"THE MUSIC MAKER": COLLECTION OF MARILYN BOGHOSIAN

Jay Heinrichs is the editor, and Karen Endicott is the faculty editor, of this magazine.