Last October, while you were probably busy studying for midterms or putting together a Halloween costume, Google was quietly running a new, self-teaching artificial intelligence (AI) program – the largest step in AI of 2015. On January 27, Google revealed that its Google AI, AlphaGo, had beaten a world-class master at the complex game of Go, a popular board game originating in East Asia. This discovery made waves across the world of technology immediately, spelling potential new steps in advancing AI performance.
Computers and our brain
This news carries great weight because it challenges our understanding of the fundamental differences between humans and computers. Perhaps you may not have noticed, but compared to a computer, you are not very reliable. For example, you are unlikely to successfully perform the same minute task repeatedly. In contrast, the actions of a computer are not very adaptable. Of course, it is clear that humans have found a mind-blowing range of tasks that computers can do, but any given program could only ever do exactly what itís instructed to. Thus a distinction between computers and humans can be made: computers are good at completing repetitive tasks, while humans are good at creative problem solving.
This difference is at the very core of what a computer is. Underneath numerous layers of computational architecture, computers are simply machines that hold, access, and manipulate binary data in a series of 1’s and 0’s. In fact, the majority of historical advances in computer hardware have, more or less, involved allowing machines to perform more of these functions and at faster rates, or performing simultaneous functions. This holds true from silicon technology to miniaturization.
That being said, viewing computers as just a jumble of 1’s and 0’s is clearly reductive. Computers’ software and machinery are very complex. Many of us struggle with McGillís various computer interfaces, yet no one would be willing to entirely forgo them: they are clearly necessary. Among the many impressive accomplishments that these 1’s and 0’s have under their collective belt is the ability to consistently beat any human at chess. Go had however, until recently, been seen as an entirely different ball game.
[There are] an astounding 2.08×10170 [possible moves] on a standard Go board.To put this in context, there are only an estimated 1029 stars in the entire universe.
The ancient game of Go
Go is an immensely popular board game that has deceptively simple rules. Players aim to control as much of the board as possible by using their own pieces to surround those of their opponents. The complexity lies in the sheer quantity of different moves available to each player: an astounding 2.08 x 10170 on a standard Go board. To put this in context, there are only an estimated 1029 stars in the entire universe. Obviously, this number is so large that the human brain can only rely on metaphors in order to conceive of it. We simply have no frame of reference.
Intuition and the art of Go
The interplay between deceptive simplicity and inward complexity is perhaps part of the reason that Go has commanded the passion of billions for over two millennia. It is also perhaps why Chinese and Japanese thought often describes the playing of Go as an art.
Google, for its part in the announcement earlier this year, stated that intuition is the key for success in Go. This cuts right to the heart of the question that AlphaGo’s victory represents. While nebulously defined, intuition clearly falls on the human side of the spectrum of abilities. In fact, it could be argued that intuition is the marker for the human end of that spectrum, because it represents the elucidation of knowledge or understanding obtained without inference or the use of reason. There are no 1’s and 0’s. Therefore, if AlphaGo used a form of machine intuition, this could be one of the first instances where artificial intelligence has crossed over the barrier toward human intelligence.
This begs the question, however: how did AlphaGo beat a professional Go player? Essentially, it did two things. It trained against recreations of moves and games that had happened in real life, using state of the art machine-learning techniques to improve its game. Specifically, it used an algorithm known as a Monte Carlo Tree Search (MCTS) which, at every turn, selects many moves at random. It then follows the implications of these moves through an entire simulated game, using the results of those decisions to decide how to move in the actual turn at hand. This is more or less the process that Deep Blue, the AI that beat the then world’s champion in chess – Garry Kasparov – in 1997, underwent. This process builds up a store of possible moves within possible contexts that the AI can draw upon during any given game. Sophisticated as these algorithms are, the massive number of possible moves in Go rules out their use as the central strategy for an AI. Even a computer can’t be realistically expected to learn a number of moves so great that the numbers that describe them, over a googol (1 x 10100), sound more at home in children’s games than in science.
AlphaGo also relied on neural networks to further improve its gameplay. These networks [were] inspired by the current understanding of the human brain.
In order to surmount this obstacle, AlphaGo also relied on neural networks to further improve its gameplay. These networks, inspired by the current understanding of the human brain, simultaneously select the next move from the repertoire the machine has built up. In addition, they also predict who the winner of the game will be. These predictions are updated every turn and are used recursively to select future moves.
This innovative approach is a step toward where Google claims that a mechanical intuition may be created. By analyzing an uncertain future, AlphaGo informs its present. This is something that we do every day without realizing it. We dress warmly in February because experience has taught us to expect brutal winter chills. We binge-watch Netflix because we have learned that we can get away with it. But is this really what we mean when we talk about intuition? Probably not.
The essence of intuition
Prediction can inform intuition and vice versa, but they are not one and the same. I can predict that tomorrow I will drink a cup of coffee, but it is intuition that tells me I will be tired and need a cup of coffee tomorrow morning. In this sense, intuition in a machine has yet to be achieved.
The future still holds much excitement for AlphaGo and its team. In March, they will face off against Lee Se-dol, who is considered the greatest living Go player. This will be a fascinating test of the extremely advanced and cutting edge AI technology that Google has created. It will not, however, drastically change what AlphaGo has to offer the world and will leave an important question unanswered – the uncertain future path of AI.
Artificial intelligence is an uncharted field that will continue marching forward. What is unclear is if this progress will be made by replicating the human mind, as a machine capable of intuition would do, or if it will follow the trajectory of its own unique and heretofore unknown kind of intelligence.