This is an article I wrote for the school magazine.
“Ok Google, what’s the weather today?”
Today, artificial intelligence and machine learning are heavily utilized into our daily lives. By applying artificial intelligence and machine learning into mobile phones and refrigerators, developers create software that help people operate more efficiently.
In the past, humans had to write individual lines of code in order to represent all types of actions. For example, say you wanted to translate Korean into Chinese. Without machine learning, developers would have to code all combinations of text in order to fit the each and every need of customers. With the introduction of machine learning, however, developers are able to make astonishing improving using customers suggestions and feedback. From these suggestions, developers are able to alter their algorithms or modify their code so that the machines designed to translate are able to “learn” with every feedback. With every repetition, the machines develop more and more. This is what allows big companies such as Google and Facebook to build accurate software. More data, more accuracy.
Since the 1970s, computer scientists have been training computers to play chess. One of the earliest computers that could play on a competitive level was Belle. Belle was a chess computer developed at Bell Labs. It could play chess at master-level with a USCF (United States Chess Federation) rating of 2250. Out of 50000 approximate ranked players, only 2500 are above master level. Belle went on to win the ACM North American Computer Chess Championship five times and the 1980 World Computer Chess Championship.
Fast forward to March 2016. Google DeepMind’s AlphaGo defeats South Korean professional Go player Lee Sedol 4 to 1. Lee Sedol, who has the second highest victories in world, faced AlphaGo at Four Seasons Hotel in Seoul, South Korea. While AlphaGo ran on servers in the United States, Aja Huang, a DeepMind team member, placed stones on the Go board based on AlphaGo’s calculations. It was reported that the AlphaGo version played against Lee Sedol used 1920 CPUs and 280 GPUs. Compared to computer chess, computer Go is significantly harder to mimic because there are just so many probabilities and combinations. In chess, the first player has 20 possible moves while, in Go, the first player has 361 moves. Given an average of 200 moves throughout the games, in order to evaluate every possible position 8 moves ahead, a computer would have to compute more than 512 quintillion (5.12 * 10^20) possible combinations. It would take roughly 4 hours for the Tianhe-2, the fastest supercomputer in the world as of 2014, to run through each combination and find the best move.
Well now you ask, “If it takes the world’s best supercomputer 4 hours to calculate a single move, how come AlphaGo placed its stones in a matter of seconds?”
Here is the part where machine learning comes in. Say you are solving a trivia problem that goes “What is the capital city of the United States?” and you have 4 choices. 1. Berlin, 2. New York, 3. Washington D.C., and 4. London. Based on your background knowledge, and if you passed middle school geography, you should immediately get rid of Berlin and London because they are capital cities of Germany and London. However, some might be confused between New York and Washington D.C. After all, they are both famous cities and are in the United States. So which one to choose? However, with enough background knowledge in geography or have just visited the United States, one can eliminate New York and choose Washington D.C. Similarly, AlphaGo amassed a huge amount of background knowledge to have accurate precision in Go. DeepMind states that, through the KGS Go Server, AlphaGo learned 30 million different moves from expert Go players. Having this vast knowledge in mind, AlphaGo uses a Monte Carlo tree search (a search algorithm for decision making) to process positions of stones on the Go board and in turn generates the best move to make. By using machine learning (feed in data –> results change in real time), AlphaGo does not have to individually calculate the 512 quintillion possible moves. If you want to read more, here is a Nature article explaining AlphaGo in detail.
Now, you may wonder how AlphaGo and search algorithms have anything to do with our lives. After all, you can’t see AlphaGo or these algorithms in real life. However, peeking into the daily lives of people up close, these things impact us significantly. From Wifi to passwords and from traffic signals to video compression, algorithms are heavily implemented into our daily lives. Through artificial intelligence and machine learning, algorithms are able to be used effectively extensively. Yet we can’t sit here bewildered by its marvelousness. Some are calling the evolution of artificial intelligence the next Industrial Revolution where everything in machines will be connected to the Internet. From minimizing food waste to self-creating music, the application of artificial intelligence in the future is mysterious and exciting.
Sang Won (Peter) Park 