To beat Go champion, Google's program needed a human army

(THE NEW YORK TIMES) Nearly 20 years ago, after a chess-playing computer calledDeep Blue beat the world grandmaster Garry Kasparov, I wrote an article about why humans would long remain the champions in the game of Go.

“It may be a hundred years before a computer beats humans at Go — maybe even longer,” Dr. Piet Hut, an astrophysicist and Go enthusiast at theInstitute for Advanced Studyin Princeton, N.J.,told me in 1997. “If a reasonably intelligent person learned to play Go, in a few months he could beat all existing computer programs. You don’t have to be a Kasparov.”

That was the prevailing wisdom. Last month, after a Google computer program called AlphaGodefeated the Go masterLee Se-dol, I asked Dr. Hut for his reaction. “I was way off, clearly, with my prediction,” he replied in an email. “It’s really stunning.”

At the time, his pessimism seemed well founded. While Deep Blue had been trained and programmed by IBM with some knowledge about chess, its advantage lay primarily in what computer scientists call brute-force searching. At each step of the game Deep Blue would rapidly look ahead, exploring a maze of hypothetical moves and countermoves and counter-countermoves. Then it would make the choice that its algorithms ranked as the best. No living brain could possibly move so fast.

But in Go, an ancient board game renowned for its complexity, the ever-forking space of possibilities is so much vaster that sheer electronic speed was not nearly enough. Capturing in a computer something closer to human intuition — the ability to seek and respond to meaningful patterns — seemed crucial and very far away.

Other seemingly distant goals included the ability to translate automatically between two languages or to recognize speech with enough accuracy to be useful outside the laboratory. Computer scientists had already spent decades trying to crack these problems.

For many, the aim was not just to make an artificial intelligence, but to understand deep principles of syntax, semantics and phonetics, and even what it means to think.

Now anyone with a smartphone or laptop (communing by Internet with a supercomputing cloud) can get a rough translation of text in many languages. They can dictate instead of type. Photo software can sort not just by date and location but by the faces of the subjects.

The results are imperfect and often clumsy, but they would have been mind-blowing in 1997. What happened between then and now?

Of course, computers became ever more powerful. But even today’s fastest aren’t able to anticipate all of the permutations of a situation like playing Go. Success on this and other fronts has come from harnessing speed in other ways.

The breakthrough in translation came from setting aside the question of what it means to understand a language and just finding a technology that works. The automated systems start with a text that has already been translated, by human brains. Then both versions are fed to a computer. By rapidly comparing the two, the machine compiles a thicket of statistical correlations, associating words and phrases with their likely foreign counterparts.

Similar approaches, more artificial than intelligent, have led to surprisingly rapid improvements in recognizing speech and facial images, as well as with playing championship Go.

In AlphaGo, learning algorithms, called deep neural nets, were trained using a database of millions of moves made in the past by human players. Then it refined this knowledge by playing one split-second game after another against itself.

Tweak by algorithmic tweak, it became ever more adept at the game. By combining this insensate learning, which amounts to many human lifetimes of experience, with a technique called Monte Carlo tree search, named for the ability to randomly sample a universe of possible moves, AlphaGo prevailed.

That was an enormous victory. But the glory goes not to the computer program but to the human brains that pulled it off. At the end of the tournament in Seoul, South Korea, 15 of them took the stage. They represented just a fraction of the number of people it took to invent and execute all of the technologies involved. Lee Se-dol was playing against an army.

Back in 1997 I wrote, “To play a decent game of Go, a computer must be endowed with the ability to recognize subtle, complex patterns and to draw on the kind of intuitive knowledge that is the hallmark of human intelligence.” Defeating a human Go champion, I wrote, “will be a sign that artificial intelligence is truly beginning to become as good as the real thing.”

That doesn’t seem so true anymore. Ingenious learning algorithms combined with “big data” have led to impressive accomplishments — whathas even been called bottled intuition. But artificial intelligence is far from rivaling the fluidity of the human mind.

“Humans can learn to recognize patterns on a Go board — and patterns related to faces and patterns in language — and even patterns of patterns,” said Melanie Mitchell, a computer scientist at Portland State University and theSanta Fe Institute. “This is what we do every second of every day. But AlphaGo only recognizes patterns related to Go boards and has no ability to generalize beyond that — even to games similar to Go but with different rules.

“Also, it takes millions of training examples for AlphaGo to learn to recognize patterns,” she continued, “whereas it only seems to take humans a few.”

Computer scientists are experimenting with programs thatcan generalize far more efficiently. But the squishy neural nets in our heads — shaped by half a billion years of evolution and given a training set as big as the world — can still hold their own against ultra-high-speed computers designed by teams of humans, programmed for a single purpose and given an enormous head start.

“It was a regrettable game, but I enjoyed it,” Mr. Lee said during the award ceremony. (Regret, enjoy — these words do not compute.) He added that the contest “clearly showed my weaknesses, but not the weakness of humanity.”

Picking up the plaque and bouquet he had been given as consolation prizes, he laughed nervously and stumbled from the stage. Several days later, he said he would like a rematch.