AlphaGo: The Essence & Implications

One company we all need to pay attention to right now is DeepMind. For those who may not know, Google spent a big sum, approximately $500 million, to acquire this small startup a few years ago. Last year, DeepMind’s AI program AlphaGo beat a grandmaster of a very ancient and, yet, extremely complex chess game “Go”. We would like to address some of technologies core to DeepMind’s success and all the wonderful applications it can have towards artificial intelligence.

Like our company Finaius, DeepMind has the vision of creating “TRUE” intelligence. What exactly is “TRUE” intelligence? Throughout history, many of our inventions are inspired by nature. By studying the birds and their aerodynamics, we invented the airplanes even though our inventions are not quite the same as birds. And, as computers have gotten “smarter” per se, we still have not been able to understand the smartest computer of all — our brain. Our brain holds the key for all the inventions throughout history. If we want to create something that is intelligent, our brain is the blueprint. Sure, maybe someday, we may discover a different type of intelligence that is better than human brain. But until that day, our only hope for creating artificial intelligence is to reverse-engineer the human brain. Just like how we studied birds to learn to design airplanes. We aim to understand the principles of intelligence, not completely copy the structure of the brain.

How close is the intelligence of AlphaGo to that of human brain. We would say that it is very far, but it is getting closer. In the highest level, AplhaGo works through the combination of efficient tree search and approximation of decision quality. Since “Go” is a very intuitive game, it has been a major challenge for machines to model the game. Think about it, with a 19x19 game board, the number of possible moves is more than the number of atoms in the universe. Even the fastest computers on the planet can’t model the game with brute force. When a human plays the game, he or she can’t possibility figure out all the steps in his head. Instead, based on many years of experiences, a player will do the following things:

· Have a sense who is winning or losing

· Have some good moves in mind, and possible some good moves after that. If that player is really smart, he or she might simulate more steps in his mind.

· Make the move that he or she thinks will lead to the ultimate victory.

Now, let’s have a computer model this process. Since it unrealistic to model all scenarios of the board positions by hand, we will use neural networks. Neural Networks are tools that can approximate almost any functions automatically, given enough inputs and outputs. For example, feed enough images of animals and tell them what animals are in there, the neural net will auto-associate pixels to figure out which is a dog or which is a cat. Based on the same idea, we can have a history of all the games and feed into the neural network to figure out two things:

• Who is wining?
• Given a board position, what are the likely moves after this position?

By viewing millions of games, the network shall learn these two things. Now, all what is left is to do the simulation of the game when playing. This is where the computer excels. Given a board position, since the computer now have a sense who is wining and losing, and also the likely moves at each steps, it can simulate way more steps than its human opponent, if not all the way to the end of game.

This is how AlphaGo works, through our extremely non-technical explanation above. To actually make it work, there is a lot more to it than that. AlphaGo has two networks. The first one is policy network, which is to select best moves. The second one is value network, which gives a value between 0–1 to show which is winning. It also improves itself through playing against the older versions of itself.

Source: Nature: Mastering the Game of Go with Deep Neural Networks and Tree Search

To make truly intelligent AI, there are more tasks that are needed to be done. For example, to do a simple task such as driving through a city, the AI needs to have the concepts of cars, houses and so on. The AI needs to store and retrieve both long term and short-term memory and also associate old knowledge to new facts. Unfortunately, there is no computer capable of solving this task. However, recent advancement suggests that we might not be far from it. The new invention is called Neural Turing Machine, or more recently, the Differentiable Neural Computer. We believe that this could be the beginning of a really profound revolution of the computer and artificial intelligence industry. We will devote our next blog on it.

At Finaius, we intend to utilize the latest research and discoveries of the human brain and rapidly apply them to our software systems. The idea behind AplhaGo is straightforward. The key innovation is to combine the learning capability of neural networks and the rigor of mathematical decision-making process. Here are three applications that we believe this technology can have:

1. Scientific Experimentation. We can have such a system to conduct scientific experiment. The network can learn to evaluate results and make highly efficient decisions, in order to achieve a goal. We can imagine giving the AI all sorts of chemicals and desired effects of a new drug. Through the same learning and optimizing process, the AI can try millions of different combinations, building up knowledge and eventually come up with a solution.
2. Business Planning. The network can learn to evaluate business factors at each stage and propose actions that result the highest value.
3. Investing. Our financial market is highly complex. It involves with millions of parties, each making decisions hoping to generate rewards. So, in a way, it is like a game. The neural network can learn to evaluate all those market moving factors and make investing decisions based on that. What’s more? Through making those decisions, AI can figure out best winning strategies that humans might have never even considered.