Lessons from the AlphaGo shock
Last week, a new chapter was added to the artificial intelligence (AI) textbook. The game of Go was considered a unique territory of human intelligence, but a computer showed it could beat the best human Go player. Strategy board games requiring high intelligence, such as chess, Go and poker, have been studied by artificial intelligence developers. Set aside the philosophical debate over what is intelligence and what is not, if a computer can display capabilities on par with the highest levels of human intelligence in a game that requires advanced strategies, it could be engineering proof of intelligence.
AlphaGo is especially significant because no Go player was involved in developing the algorithm. It adopted a textbook AI algorithm and optimized the ability by studying historical games between human players that can be easily downloaded online.
AlphaGo uses old technologies that have already been published. Monte Carlo tree search, which AlphaGo uses to randomly search numerous possibilities and determines the move, is a standard used in nearly all AI Go software for 10 years. Self-learning through self-match was first used in an algorithm for backgammon in 1992.
Understanding the overall situation affects the efficiency of the search algorithm, and AlphaGo accomplished a great improvement in capability here. It applied deep neural network technology, which is a big trend in machine learning today. Neural network technology was developed in the 1990s.
The neural network used in AlphaGo is a convolutional deep neural network that mimics the calculation happening in the cortex of vision in highly intelligent animal brains and perceives objects in images. In other words, AlphaGo looks at the Go board as an image of black and white pixels, just as a human player intuitively understands the image. AlphaGo uses the search strategies using the computer’s computing speed and deep neural network mimicking human intuition and performed beyond the highest capability of human players.
In fact, I wasn’t comfortable with this match. Google DeepMind is scouting competent AI researchers around the world, so if AlphaGo is defeated, AI research itself may lose momentum. I thought the so-called AI winter - the period when funding and interest in artificial intelligence research was reduced in the 1980s - could return.
The Go matches got great public attention and could affect the future of the AI field. Dr. David Silver, who led the technological development of AlphaGo, studied AI Go since graduate school. I think he planned the event because he was confident from the experience of playing against a professional player in 2008. I believe that the matches were not a triumph for AI but a victory of David Silver’s devotion and passion.
I believe that the following are significant in AlphaGo’s match against Lee Se-dol.
First, let’s not look at it as a match between computer and humanity. Computers have become convenient and powerful enough to solve these complicated problems. The purpose of a computer is to solve challenging problems that humans find too complicated or time-consuming, so artificial intelligence is an inevitable product. For example, it is up to humanity to use nuclear engineering on nuclear energy or nuclear weapons. We are the ones who utilize advanced technologies. Technologies will continue to progress, and we need a social foundation to use them wisely.
Second, the outcome of the matches do not show “completion of AI.” The technologies used in AlphaGo are not universally-used AI. It contains an algorithm specially designed for the game of Go. Computer chess, for instance, does not use Monte Carlo tree search or deep neural networks, the core technologies of AlphaGo. AlphaGo learned through winning and losing numerous matches. So the self-learning algorithm cannot be used on unmanned vehicles or robots, because even a simple mistake could lead to serious damage or casualties. We still have a long way to go to achieve the grand goal of AI.
Third, AI has become a key piece of knowledge for the future generation. The Go matches have highlighted the importance of computer programing education, and in the future, people need to be able to freely handle AI technologies. I am guessing that the key source code personally written by DeepMind researchers for AlphaGo is about 1,000 lines. This is a workload a computer science student can do in a few days. But if one doesn’t have core AI knowledge, one can code for his entire life without making a product.
JoongAng Ilbo, March 15, Page 33
*The author is a computer science professor at the Korea Advanced Institute of Science and Technology.
by Kim Kee-eung