Mind and machine

WHEN he was doing something—simulating on paper how a computer might solve one of Euclid’s theorems, say—Marvin Minsky often found himself improvising a nice little tune. He could only do that, though, if his hands were open. Why was that? Sometimes he could just sit down at the piano and play, out of his head, an original fugue. How? If he merely wanted to rearrange the imposing row of stuffed cows, dinosaurs and Ninja Turtles behind his sofa, his brain would whirr through millions of anecdotes, analogies, histories, possibilities and smidgens of common sense before settling on the line-up he perceived as “best”.

Nothing was so fascinating as human intelligence. Physics, which he excelled at, was quite profound, but intelligence struck him as “hopelessly” so, an immense realm of the imperfectly understood. Neuroscientists scratched at bits of it, psychologists at others, but he had little time for them; as a trained, if easily bored, mathematician, he started from the proposition that logic governed the whole thing. Every thought, hence all intelligence, was the result of cascades of pulses rippling through networks of semi-autonomous agents, the neurons, each connected to countless others. Since men thought like machines, it should be perfectly possible to build, with simulated neurons, a machine that could think like a man.

Other dabblers in artificial intelligence wanted it to do things humans found hard, such as calculus or chess, but Mr Minsky was more interested in apparently going backwards. He dreamed of programming a machine which, in reasoning by analogy and learning from experience, would approximately reach the level of a three-year-old child; for that was much more difficult.

First steps were slow. For his doctorate at Princeton in 1954 he analysed a “learning machine” he had built as an undergraduate to simulate the neural networks of the brain: primitive information in, primitive information out. When he went to MIT to teach in 1958 he set up the Artificial Intelligence Group, later Laboratory, which became the live core of all AI research, though it was still focused on building machines as much as on using computers. (He remained keen on building, especially on the merits of giving children Tinker Toys rather than Lego, which enabled them to build triangles and think flexibly in threes, rather than in Lego’s relentless rectangles.)

Even as computing power increased, though, he could not build networks beyond a few hundred neurons in a single layer: a micro-fraction of the brain’s complexity. And, not least because real neurons were arranged in columns of apparently hierarchical layers, he had to find a way of working top-down as well as bottom-up. He needed to find ways of saying what sort of high-level aims an artificial intelligence might have, and towards what goals it might be programmed. The search for that soon obsessed him.

Disrupting Mozart

He also, almost by the way, did other things, such as inventing a confocal scanning microscope and robotic “seeing hands” for surgery. His own intelligence continually leapt between postulations and speculations, all delivered with an endearing smile: what a thinking machine would have to notice when it drove down the highway, whether robots could be made tiny enough to beat up aphids or dexterous enough to put a pillow in a pillowcase, what would happen if you wrote “Eine Kleine Nachtmusik” to a different rhythm. Students flocked to his evening classes, never quite knowing what mental challenge he would toss out next.

His searching flowed in two main directions, and these were complementary. As he built his thinking machine, he hoped the machine would illuminate how the human brain worked. Mystery annoyed him, but he’d gradually clear it up. Emotions, for example (as he explained in “The Emotion Machine” in 2006), were just as mechanistic as any other action of the brain. They could all be reduced to defence mechanisms, information retrieval, and so on. Even consciousness was the result of “possibly 16” processes like those. As for pleasure, that too was a piece of machinery, and one you had to learn to turn off before you got addicted.

What about his own pleasures? Questioned about his “love” of this or that, he wondered what that meant. Science fiction he enjoyed, because it was full of human transformations. Literary novels, though, were “all the same”. His life was his own ideas and those of intellectual strivers even brighter than he was, from Bronx Science high school onwards. No time, in all that fizzing seriousness, for “social stuff”.

He lived just long enough to see the re-emergence of his theory of neural networks, as the data-crunching capabilities and “deep learning” of modern computers began at last to approximate to the workings of the brain. In the very week he died, a computer beat a human at the ancient, infinitely complicated game of Go. The key to making a machine that could think like a man (or, as he hoped, far better) evidently lay in one of his “third ways”: a combination of the top-down, symbolic approach and the bottom-up simulation of the ever-pulsing neurons. It would still be a devil of a job, though, to make a machine that thought like him.