Deep Learning

For a while, in the mid-1970’s, I was obsessed with chess, caught up in the fever generated by Bobby Fisher. I spent many hours playing and studying the game. Among competitive players, I was just average. I didn’t have whatever it takes to play at a high level.

But I could beat any computer in the world. In those days, any reasonably competent player could. Computers just weren’t any good at chess, and experts predicted they never would be.

In a chess game there are literally millions of potential positions after just 3 moves each, and hundreds of billions of potential positions after only 4 moves. The number of potential combinations in a typical chess game exceeds the number of atoms in the universe. The reason early computers weren’t any good at chess was because they had to analyze every potential move, whereas a human player, based on experience and pattern recognition, could automatically exclude the vast majority of potential moves and focus only on the few that were sensible in that situation. Playing chess well required a degree of abstract thought that computers weren’t capable of.

Not any more. Nowadays chess computers can beat any player in the world. Cheap devices found in toy departments play at the grandmaster level. The human mind simply cannot match human-created computers when it comes to playing chess.

Go is exceedingly more complex than chess. The number of potential combinations in Go is far greater than that in chess. Yet this year a computer called AlphaGo defeated the world Go champion. As with chess, humans have created a machine that plays Go better than its creators–better even than any potential human creator.

These computer achievements are due to a process called “Deep Learning,” which uses “deep neural networks” that enable computers to mine Big Data and teach themselves, essentially imitating the human learning process, but at lightning speed. The implications of this technology, which is developing at an astonishing pace, are mind-boggling. I urge anyone who has the time to read this article, provocatively titled: “Deep Learning is Going to Teach Us All the Lesson of Our Lives: Jobs are for Machines.”

I’ve blogged often about the effects automation and robotics will likely have on human work. I won’t go into that again this morning, but the linked article addresses it well and raises some fascinating questions and concerns.

But aside from what Deep Learning means for the human job market, I wonder what it will mean for human intelligence?

Over the course of history our appreciation for what it means to be intelligent has evolved. We don’t typically consider mere literacy an indication of advanced intelligence any more, for example. When I was in school rote memorization was still being taught. The ability to memorize and recite a poem or a passage from Shakespeare was considered a sign of intelligence. Nowadays that might be a neat bar trick but hardly anyone would measure intelligence by how much stuff a person has memorized. More recently the ability to do calculus would be an indication of intelligence, for example. But if an inexpensive device most people carry in their pockets can do it better and more quickly, what’s the point of knowing how to do it? Why is it any different from being a chess grandmaster, who would be routinely defeated by an inexpensive computer?

Over the last hundred years or so humans have devised ways to measure human intelligence, and hundreds of studies have proven that intelligence correlates to nearly every measure of human well-being: education level, income, longevity, criminality, marital stability, health, etc. It would be wrong, I think, to assume that in a world of Deep Learning computers human intelligence becomes entirely irrelevant.

But other than as a measure of human capacity for well-being, what are we to do with our “intelligence” once we don’t need it for the things we’ve traditionally used it for? Are we at a point in history when in the time it would take for a human to acquire the skills, experience and education to find a cure for cancer (for example), Deep Learning machines could have already solved the issue in a small fraction of the time? Every time I think of things like this the question Wendell Berry asked many years ago rattles around in my brain, “What are people for?”

I go on about this for hours, but for now I’ll close with another interesting thought. What if in 25 years or so, what we now identify as human intelligence is no longer important, because artificial intelligence will have so far surpassed it as to make it practically irrelevant? I heard a commentator predict that when that happens, we will begin to objectively measure and assess human intellect by reference to “emotional intelligence,” rather than whether our neurons fire in ways that make us good at chess, calculus or memorizing Shakespeare.

We surely live in a fascinating time.