This is a blog post from
For ages, human beings have held the upper hand over computers when it comes to one thing — skill in the game of Go, which calls for intuitive acumen rather than brute-force computation.
In late January this year, therefore, computer scientists were ebullient that a piece of software, called AlphaGo, had defeated a human Grandmaster.
Go is an ancient two-player board game that’s popular in China, Japan and Korea and also loved by mathematics aficionados (John Nash is shown playing the game in A Beautiful Mind ) and programmers. It is similar to chess but played out on a board with movable black and white buttons and involves the war-like capturing of ‘territories’. It’s harder to determine an eventual winner than in chess because of the much-bigger board. Unlike in chess, where capturing the king is the purpose of the game, the winner in Go is the one who captures more area on the board.
Twenty years ago, chess titan Garry Kasparov overcame Deep Blue, a formidable machine developed by IBM that was for all practical purposes designed to overcome the quasi-unassailable Soviet (Russian?) grandmaster. The victory of the temperamental Kasparov over the feeling-less Deep Blue in that six-game tournament was a reassurance that ‘insight’ — an attribute we humans arrogate to ourselves in order to distinguish us from other organisms with a brain — triumphed over the formidable calculating-prowess and the vast memory of Deep Blue; Kasparov could evaluate three chess positions per second to Deep Blue’s 200 million.
Next year, in a rematch, Kasparov lost to a new and improved Deep Blue. The scores were thus even: Machine ~ 1, Humans ~ 1.
Children learn — a sport, swimming or languages — not necessarily always through the rules being read out (or programmed into them) but by beginning to play, figuring out patterns and imbibing strategies and learning from mistakes.
AlphaGo’s victory, >reported in Nature , is paradoxically an important victory for humans — the program is a product of an alternative approach called Artificial Intelligence, which endows machines with the capacity to think, like a human being. Though it might look like our computers are amazing multi-taskers — streaming movies, playing music, reading email — their circuitry-innards are arranged linearly and it goes from one task to another in a strictly sequential manner. This has worked brilliantly so far is because, for half a century, it’s been possible to double computing power annually without increasing costs.
AI basically mimics so-called ‘neural networks’. This is where you arrange the several million processors in a machine to interact with one another in the same way that neurons connect in a human brain. Neurons, unlike integrated circuits of chips, can be ‘taught’ to respond to certain signals and ignore others.
Children learn a sport or a language not necessarily always through the rules being read out (or programmed into them) but by beginning to play, figuring out patterns and imbibing strategies and learning from mistakes. London-based R&D firm DeepMind used a similar strategy, employing neural networks to teach their brainchild AlphaGo.
For a long time — before the British company became famous after being acquired by Google — their programmers spent a fair amount of time in the ticklishly-metaphysical activity of teaching their several softwares to win at videogames. AlphaGo hasn’t yet met Lee Sedol — the Garry Kasparov of Go — but that’s scheduled for later this year. What’s undeniably impressive, though, is not only that AlphaGo has bested the European Go champ Fan Hui, but that it has picked the game up the way humans do — watching, blundering and then improving.
In 1997, after Deep Blue won, IBM was politely derisive about artificial intelligence. “Artificial Intelligence is more successful in science fiction than it is here on earth, and you don't have to be Isaac Asimov to know why it's hard to design a machine to mimic a process we don't understand very well to begin with” is how IBM's website explained it. Beyond chess, the capabilities of Deep Blue were to be applied to the problems of evaluating stocks, scheduling flights and travel.
Tough problems — those that were beyond the calculating ken of humans — were believed to be well within the capabilities of AI, which would home in on solutions using the glut of low-cost processing power that Intel was churning out.
But this was before stock picks recommended by those efficient, mathematically-rational models running on supercomputing behemoths turned out to be duds that wiped away wealth. And the economics Nobel Prize, for debunking the assumption that humans were always rational while making economic decisions, had not yet been awarded.
Today, servers everywhere are obese with data and the problem — as Google and Amazon are realising — is how to make sense of it all. It is like being stuck in dimly-lit library which stocks every book in the world. The breakthrough algorithms that DeepMind seems to have discovered lies in being able strike a match in that library. Rather than trying to make machines superhuman, several AI researchers are in fact trying to better understand what it means to be human.
Driverless cars and algorithms that can predict newsfeeds — and hopefully, one day, fire-hazards and dengue outbreaks — are now first taught to sift through trillions of cat videos and learn for itself that ‘This is a cat’. Broadly, these new algorithms and such training are said to be in a class of research called ‘Deep Learning’, but nobody can yet comprehensively explain how these neural networks learn. Kasparov was prescient. On losing to Deep Blue, he told Time : “I had played a lot of computers but had never experienced anything like this. I could feel — I could smell — a new kind of intelligence across the table.”
