Press enter to see results or esc to cancel.

It’s Go time in the AI world!

Ever since ancient times, people have always thought of building artificial life with human-like abilities. This has always seemed to be a distant dream. Until now.

But before I get into the details of why you should fear the Terminator, let me give you a very brief history of AI.

Let’s go back to the early 50’s during the time of Alan Turing, the great computer scientist. In his seminal paper, Computing Memory and Intelligence, Alan asks, “Can Machines Think?” He proposed a test of a machine’s ability to mimic intelligent human behavior.

The way it works is as follows. Imagine you are a judge in a strange contest. There are two contestants A and B hidden from your view. One of them is a computer, and the other a human being. You are to determine which player is which based on their written responses to a set of questions. If you cannot reliably do so, then the computer has passed the Turing test.

The test is not without its criticisms, but it is, nevertheless, an important concept. Variants of this test abound.

The Turing Test
The Turing Test

Around this time, AI researchers hit upon the idea of having computers attempt to beat the best humans at mind-bending games like chess.  The reasoning was that since playing games like these is such a human endeavor, it would be a perfect barometer of measuring AI’s progress. During the 1950s and 60s, computer chess programs got better and better until they were able to challenge decent amateur players. But beating world class players remained a distant dream. Also, looking at the games played between the computer programs and humans, you could easily tell which player was the computer, and which was the human. In other words, these programs failed the “Turing test” for chess programs!

This was the status quo until the late 1980s. In 1988, Deep Thought, a chess-playing computer shared first place with a British Grandmaster Tony Miles, in a chess tournament. It beat several strong players on the way.

However, when it faced off against the World Champion, Garry Kasparov, it stumbled. It was humiliated 2-0, and Garry Kasparov went so far as to proclaim that he would never be beaten by a computer in his lifetime.

In 1997, he was forced to eat his words as the impossible happened! IBM’s Deep Blue, in a highly publicized six game match, beat him with two wins to one and three draws. This was headline news all over the world! In fact, I remember this moment vividly. I was in class when our teacher broke the news. Everyone was shocked. And what was even stranger was that in a few games, the computer’s play was indistinguishable from a human. It had passed the Turing test! Could we now become slaves to our own creation?

After the dust had settled, people analyzed the result to examine what it really meant. People realized that it was not so much as intelligence driving the program, it was its “calculative” power that gave it the edge. It calculated around 200 million moves per second! Of course, this is a mere drop in the pond compared to the number of possible chess positions. But the point is that it was basically like a calculator. You don’t feel too bad if your calculator performs arithmetic faster than you, do you? And Deep Blue couldn’t tie its shoelaces together. Also, it was prone to situations such as these:

Deep Blue loses in one move!
Deep Blue loses in one move!

But there still was that eerie feeling of “What have we unleashed?” This feeling intensified in 2011 when the “intellectual” successor of Deep Blue, IBM Watson, beat the best human champions at Jeopardy! Jeopardy! is a fast paced general knowledge game with a twist. Participants are given a clue in the form of answers, and they have to form the question. It was thought to be beyond the realm of AI for a while because the game uses puns and quirks of human speech in the clues. But Watson was up to the task as it used the latest in Natural Language Processing to beat the humans. The result shook up the community for a while, but people rationalized the result away just as they did with Deep Blue. Also, some of the answers Watson gave was quite ridiculous. In other words, it didn’t pass “the Turing test” for Jeopardy! Also, Watson still couldn’t tie its shoelaces.  (Perhaps what would be real news, is something like the following.) 

Watson wins a date!
Watson wins a date!

Let’s now fast forward to the present day. When the news of the potential creation of the Terminator broke.

Just this past week, researchers at Google DeepMind announced that their creation, AlphaGo, beat the reigning European Go Champion, Fan Hui, 5-0 at the ancient game of Go. This is an astonishing result! To really appreciate the significance, you must first understand why Go is such a tough nut to crack for computers.

Go is played on a 19 by 19 grid board. In each turn, there is an average of 200 moves possible. The estimated number of possible of positions is on the order of 10^180 – far more than that of the number of atomic particles in the universe. Far more than that of chess even, which stands at around 10^100.

You can see that a brute force technique is out of the question. So AlphaGo used the latest in AI techniques to achieve this awe-inspiring result. To beat Fan Hui, it used deep neural networks: networks that contain millions of connections. Just like the human brain.

The catch is that these networks would require a large training set. Researchers had the program go over many thousands of games played by human experts. They also had it play multiple games against itself. This gave the program “a feel” for the game. This knowledge enabled it to beat Fan Hui. What’s more, the program passed the Turing test! It was impossible to tell, from the moves alone, which was the program and which was the human.

But the true test of this program’s strength is yet to come. In March, AlphaGo will face off against Lee Sedol, the reigning World Champion.  If it beats Lee Sedol, it would probably feel like a punch to the stomach. As impactful as Kasparov’s loss to Deep Blue.

In fact, even more so, as the techniques AlphaGo uses is very generic and can be applied to a variety of situations. What’s more, AlphaGo goes through far fewer positions per second than Deep Blue. Pretty soon, the program’s descendants may become our overlords!

But what comforts me is that, in the end, no matter what, AlphaGo cannot tie its shoelaces together. However, this robot scares me.

Comments

Leave a Comment

IBM

About Author

Bharat Ramakrishna

Blogger. Part-time mathematics enthusiast. Loves esoteric and quirky things. Bibliophile. Has a wide range of interests including playing chess and pool, juggling and creating puzzles of fiendish difficulty. Grammar Nazi.