AlphaGo’s victory: Has AI overcome human limitations and ushered in a new era?

In 2016, AlphaGo’s victory over Lee Sedol 9 demonstrated the incredible advances in artificial intelligence. Utilizing deep learning and convolutional neural networks (CNNs), AlphaGo’s achievement proved that AI can mimic human intuition to solve complex problems.

 

The biggest story of 2016 was the battle between Lee Sedol 9 and Google’s DeepMind. The confidence of the world’s strongest Go player, Lee Sedol 9, and the fact that the computer was built by a subsidiary of one of the world’s most famous tech companies, Google, created no shortage of buzz before the match. The traditional and highly intriguing “computer vs. human” conflict also contributed to the buzz. The match lasted five games, with AlphaGo winning by an overwhelming score of 4:1. After the match ended, many people praised Lee Sedol IX for winning even one game, saying that it is almost impossible for a human to beat a computer that computes all the cases, but if AlphaGo was simply a computer that computes all the cases, the victory would not have been surprising, but simply the natural outcome of a computer that is computing faster than before. But AlphaGo’s victory didn’t come from a leap forward in hardware, but from a revolutionary advance in its internal algorithms. And these algorithms are about to make a huge difference in our lives.
To understand the astonishment of the algorithm used in AlphaGo, it’s important to understand the Minimax algorithm, an algorithm previously used by computers playing board games like chess and Go. Unlike Go, as we know, the world champion of chess became a computer quite a while ago. The Minimax algorithm is the basis of the algorithm used in this case. The basic idea behind the Minimax algorithm comes from the self-evident idea that the ideal way to win a board game is to consider all the possible moves, look as far away as possible, and choose the best one. The algorithm gets its name from the idea of maximizing the minimum, and in the context of chess, it’s a way to make sure that any move you make minimizes the probability that you will lose given your opponent’s moves. The Minimax algorithm is a guaranteed winner, but it’s hard to apply it to Go. This is because the number of possible moves is much larger in Go than in chess. Whereas the chess board is 8×8 and the pieces are limited in their movements, Go is 19×19 and it doesn’t matter where you place the stones, so the number of possibilities increases dramatically. If you were to compute a hex in Go, you would have to consider 22 trillion cases, roughly 361^6, which would take 700,000 years, even if each case takes one second. Since it is impossible to compute all possible cases, even in chess, it is a result of using various techniques to shorten the time, so it seemed impossible for a computer to beat a human in Go.
AlphaGo, which beat a human in Go, also follows the steps of the Minimax algorithm. However, Go is a very tricky game to calculate the probability of winning, so a simple application of the Minimax algorithm could not beat a human. AlphaGo was able to achieve its remarkable results by using something akin to human intuition to drastically reduce the computational effort. This intuition comes from the deep learning techniques used in AlphaGo.
Deep learning is a technique that mimics human neural networks. Just like in the brain, where neurons send electrical signals to each other, deep learning techniques put artificial neurons in different layers. Each of these neurons contains data, and the data travels from one neuron to the next through a specific computational process. Deep learning techniques update the processing in the intermediate neurons so that the information traveling through the neurons from the input produces the desired result at the output. The update uses a method called Gradient Descent. This is a way to change the values in the existing processing along a gradient, moving towards a better result. After enough of these updates, a neural network is built that goes from input to output, which instills intuition in the computer and overwhelmingly reduces computation time. Another advantage of deep learning techniques is that they can be computed in parallel, since everything is done with matrix operations. The fact that GPUs (graphics cards) are optimized for matrix computation allows for parallel computation, which means that you can combine multiple pieces of hardware to get good performance even if the hardware doesn’t get better, making it very efficient.
AlphaGo uses a Convolutional Neural Network (CNN), which is one of the newest techniques in deep learning. CNNs were originally a breakthrough algorithm for processing images. CNNs excel at identifying and categorizing objects in a photo, something that traditional computers cannot do. CNNs have specialized regions of neurons where computation is bundled from multiple neurons into a single neuron. This makes it very good at extracting features from images and classifying images. AlphaGo used this CNN to correspond each piece of the checkerboard to a pixel in the image, and the different features that occur in a number to the RGB data in the image. This CNN allowed AlphaGo to effectively calculate the probability of winning, which is difficult to calculate using traditional methods. Another amazing aspect is that the network can update itself and learn from itself, allowing it to grow. This is surprisingly close to what we think of as artificial intelligence.
Deep learning techniques have the potential to dramatically improve human decision-making. Problems that require thought, such as categorizing photos, used to be something that only humans could do, but learning through neural networks has allowed computers to do it as well, making them more efficient. Here’s a huge possibility: when deep learning becomes widespread enough that it can be trained based on a person’s behavior, computers will know what to do in the current situation and will make their own judgments, doing the best things and preventing us from making mistakes. Our lives will be incredibly enriched.