What Does Google's AI AlphaGo Beating a 9-dan Go Player Mean?

A lively discussion these two days concerns the artificial intelligence launched by Google called AlphaGo, which played Go against Korean 9-dan player Lee Sedol, winning two consecutive games on March 9 and today, the 10th.

Ever since IBM's Deep Blue defeated the world chess champion at chess, Go has long been considered an area difficult for computers to challenge because of its enormous computational load.

However, since the emergence of deep neural network technology, the field of artificial intelligence has developed enormously. People generally find that in recent years, whether handwriting input or speech recognition, there have been enormous improvements.

The main improvement of deep neural networks is that they greatly enhance the extraction of deep logic. Through massive data training, they can continuously identify, abstract, and extract more complex logical rules, then save the abstracted results as data structures for recognizing things.

For systems like AlphaGo the principle is the same—but what truly brings the greatest improvement is that it can play against itself, constantly summarizing better results, and thus continuously optimize.

In past neural network systems, a problem was that when recognizing data, it was easy to get halfway through recognition and have the entire network distribution scatter, after which recognition could not continue. Explaining this in detail would involve deep learning, which we will not go into here.

Simply put, the achievement AlphaGo has reached in Go already surpasses most players. This means that in the domain of Go, even professional 9-dan masters have fallen before it.

Many say that Go has nothing left to play for because humans have no power to fight back against machines. This claim is clearly wrong. The solution set of Go is enormous; the number of possible Go positions is astronomical. The Go rules and playing methods AlphaGo has extracted so far are still learned from game records, with continuous self-computation. However no matter how it computes, it cannot cover all possibilities; it has only found an optimal solution within a local scope. The solving result of this local scope is very likely wrong from another angle. If there are enough game examples to analyze AlphaGo's playing characteristics, then defeating AlphaGo is not difficult for humans. In some subtle recognitions, artificial intelligence algorithms will have great difficulty.

For example, in character recognition, if you write a "2" somewhat ambiguously, then to machine recognition it might be a "2," or a Chinese character "yi," or a twisted "S," or something else. If there is context that produces misleading ambiguity, then recognizing it becomes very difficult. Although theoretically a machine can also process with reference to context, in reality it is not so easy—for instance, in machine learning, the data provided usually needs to be in standardized formats; if the content is messy and needs organizing, learning on new content becomes a complete mess.

Any system is bound to have bugs. I remember playing with intelligent chess-playing systems before: if they played step by step according to convention, the difficulty was quite high; but if you made some obviously unreasonable bizarre moves, often the system itself would be finished—they would also play some inexplicable moves. Systems like AlphaGo necessarily have this kind of problem too; it is just unclear where it manifests.

Moreover, since AlphaGo can optimize by playing against itself, this shows that when artificial intelligence plays against itself, there is always one side winning and one losing; without that, optimization would be impossible.

In other words, a solution for defeating AlphaGo in its current state certainly exists—but how these solutions can be applied in play, and how exactly to win, can only await the players' thinking.

Speaking of artificial intelligence, AlphaGo's limitation is this: the degree of intelligence it currently achieves is indeed very high, but it manifests only in the single aspect of playing Go. As an analogy, AlphaGo is essentially a hardworking idiot—because he works so hard, he can attain extremely profound mastery in one area, but in other areas he is completely useless.

An artificial intelligence that can recognize and learn everything is general artificial intelligence. Currently no such system has appeared. AlphaGo likewise cannot be called a good beginning in this direction, because its performance lies in deep specialization in one area, not in others. Of course this is still very meaningful—using computers as tools can let us better study and research some profound knowledge.

If we are somewhat humble, what illumination does AlphaGo's data training method give us in real life? It means that if a person's innate endowment is not too poor, hard postnatal effort can at the very least enable someone to achieve something in a certain field.