Compared with AI such as Siri, what is the difference between AlphaGo and others?

On March 9, 2016, Google's Deepmind's Go program "AlphaGo" will face off against professional ninth-dan Lee Sedol. In October last year, this program defeated Fan Hui, a second-dan professional player from China; that was the first time that a Go AI defeated a professional player in a fair game. This result was published in the journal "Nature" in January this year, and also triggered an extremely heated discussion - and the most commonly asked question is, will AI finally take over the world?

Siri can sing and tell cold jokes. Image source: Apple

This problem is not unfounded. In a sense, AI has already taken over: from Apple’s Siri, From the search engines we browse every day to the article recommendations and product recommendation systems on the Internet, these are all artificial intelligences—even if they are not the ones in science fiction, it is already difficult to separate them from them in our daily lives.

But AlphaGo is different from these common AIs. Their difference lies in the versatility of learning methods and techniques.

Siri: A scripted assistant

Siri is an "intelligent assistant" that can understand our verbal commands, help us search the Internet, and help us find contacts in lists . But its principle is simple: through sound recognition technology, sounds are converted into basic elements of language, such as vowels, consonants, and words, and then compared with special commands built into the system. If the comparison is a practical question, then execute the corresponding instruction; if the comparison is a vague question, select a joke from the corresponding joke library.

So its problem is clear at a glance: if you order it to do a command that is not in the system, it will crash. Although Siri is AI, it is a very limited AI: it can only solve pre-written problems.

Faced with the provocation from the Northeastern eldest brother, Siri was confused (or maybe just scared. Image source: Apple

Deep Blue: No one can beat chess, but only in chess)

In 1997, the chess machine "Deep Blue" manufactured by IBM defeated the then world chess champion Kasparov. This was a landmark event in the history of artificial intelligence. However, although Deep Blue defeated. It has the same shortcomings as Siri: it is too specialized.

As a program, Deep Blue's software is specially designed for chess. It evaluates the board's four criteria including piece strength and chess position. , king security and layout rhythm - obviously, these indicators are completely dependent on the rules of chess itself and have no scalability.

Scene of Kasparov versus "Deep Blue". ：muse.jhu.edu

Even so, it still relies heavily on "brute force". Deep Blue's hardware was the fastest chess-playing machine at the time. Although there was a system to help filter, it still failed every second. In response to this need, IBM developed custom hardware for it.

The result is that it is more like an international computer than a chess program. Chess machine. Deep Blue can only play chess, but cannot learn Go, and cannot even learn simple backgammon. In comparison, Kasparov, as a human, can learn Go, backgammon, and painting. . Deep Blue’s technology is like a key designed for chess, with great limitations

Self-driving cars: taking a new direction

The principles of self-driving cars can be simplified. The following steps:

First, it understands the surrounding environment through sensors, just like a driver uses his eyes to observe the surrounding situation;

Then it obtains the route of the road through the Internet, just like when we drive a car When using navigation software;

Then the computer program determines nearby pedestrians and how the car will move;

Finally, it calculates its best route and controls the speed and speed of the car according to this route. Direction.

Google's self-driving car. Image source: Google

It is specific to the field of self-driving, but the basic idea is somewhat close to AlphaGo.

IBM Watson: Universal Intelligence

In 2011, IBM Watson defeated human contestants on the American live-action quiz show Jeopardy!. Its technical concept is more like AlphaGo. Watson's decision-making consists of four steps: first observation, collecting data from the environment, then making assumptions about the data, then evaluating those assumptions, and finally making a decision. However, there are some differences from AlphaGo. First of all, it is designed as a question and answer machine. Secondly, training Watson requires the participation of human experts-for example, on cancer issues, scientists need to eliminate outdated information from massive books and papers. Wrong information, feed the compiled data to the machine. But at least its ability to handle many fields makes it much more scalable than its peers: Watson is already being used in the medical field.

IBM Watson's logo. Image source: IBM

So, what is the technical idea of ??AlphaGo?

Deepmind created AlphaGo in an attempt to build general artificial intelligence through reinforcement learning technology (Reinforcement learning). Its concept contains two entities, one is the artificial intelligence itself, and the other is the environment in which it exists. There are two types of relationships between artificial intelligence and the environment. One is sensing data through sensors, and the other is affecting the environment through specific actions. Because of the complexity of the environment, it cannot obtain all information, so it needs to continuously repeat the perception-reaction cycle in order to achieve maximum benefit in the environment. Most mammals, including humans, conform to this set of rules.

Reinforcement learning technology continuously senses and feeds back information in the environment. Image source: Google

Before AlphaGo, they had already used this idea to let AI play games. A 2015 paper published in the journal Nature described how to get an algorithm to play different Atari programs, including games like Space Invaders and Arkanoid. AI watches game videos like people, operates games like people, and slowly learns from a game novice to a game expert. AlphaGo is also based on the same principle and simulates the way people learn to play Go. It plays chess like people and slowly learns how to think like an expert.

This technical concept requires original data, so it is more versatile than methods that require the input of manually organized data. In principle, if you want to learn Go and Go backgammon with AlphaGo, it won't be a problem.

The reason AlphaGo's technology was first used in games is because games are much simpler than real-life problems, whether they are board games or computer games. Games are also likely to be the first areas where similar technologies have been put into practical use: after all, with the development of game technology, game developers have gradually realized that good AI and realistic graphics are equally important, whether it is real-time strategy games, such as "Star Wars" "Conflict for Hegemony" is also an NPC in the role-playing game. Advanced artificial intelligence can not only become a powerful opponent, but also an excellent team partner.

However, its strongest point is of course its adaptability and learning ability. Deepmind claims that this technical concept will soon be applied to the medical field to try to solve the problem of personalized medicine. And this is definitely just the first step.