How creative is DeepMind

AlphaZero: The computer is getting creative

A computer scientist has invented something that could be more inventive than himself. David Silver of the Google subsidiary DeepMind had already shown with AlphaGo - a software that mastered the extremely difficult board game Go at grandmaster level - where the journey of artificial intelligence is headed. Computers analyze games by human masters and learn from them.

Silver's latest creation, AlphaZero software, goes even further. She learns board games like Go, Chess or Shogi by practicing - while playing against herself. With millions of auditions, AlphaZero is supposed to be able to learn strategies that humanity took millennia to develop.

Could such an AI one day solve problems that were previously hidden from the human mind? In an interview with Technology Review, Silver explains what's in store for us.

Technology Review: In a famous game against possibly the best Go games of all time, your AlphaGo software used a brilliant move that human observers initially dismissed as a mistake. Was the machine creative at that moment?

Silver: "Zug 37", as it later became known, surprised everyone, including the Go community and us, the creators of AlphaGo. It was so far outside of the usual way of playing Go as humans developed it over thousands of years. For me this is an example of creativity, yes.

Since AlphaZero no longer has to learn from people: is the software even more creative?

If you let something learn for yourself, then it builds up your own knowledge from scratch. That in turn is almost the essence of creativity. AlphaZero has to find out everything for itself. Every single step is a creative leap. These insights are creative because they were not given to him by humans. And such leaps continue until you arrive at something that humans couldn't and that has the potential to astonish us.

They pitted AlphaZero against the best conventional chess software called Stockfish. What did you learn from it?

Stockfish is a very advanced search engine, but the heart is in a module that says, "According to people, this is a good or a bad position." Accordingly, people play a very profound role here. It is difficult to part with it and take a position that is so fundamentally different.

AlphaZero learns by understanding the position for itself. There was a very nice game where we just watched four pawns in a row give up and even wanted to sacrifice a fifth. Stockfish thought it would win big, but AlphaZero was very pleased with itself. It found a way to understand the position that was actually unthinkable from the norms of chess. It understood that position is more important than keeping four pawns.

Does AlphaZero suggest that AI could play a role in scientific innovation in the future?

Machine learning has so far been dominated by an approach called supervised learning, in which you start with everything a person knows and then distill a computer program from it that does the same. The nice thing about our new approach, reinforcement learning, is that it is a system that learns by itself how to achieve the goals set, from the first principle onwards. It's about millions of small insights, one at a time, that lead to a creative mindset. And when you can do that, it leads to something with immense power, immense ability to solve problems. Hopefully, this will result in major breakthroughs.

Are there aspects of human creativity that cannot be automated?

If we think about the capabilities of the human mind, we are still a long way from this point. We can achieve results in specialized areas such as chess or go if we dedicate massive computing power to a task. Yet the human mind is capable of radically generalizing to something else. You can change the rules of the game and it won't take another 2,000 years to figure out how to play.

I would say that the big limit of AI today is to increase the bandwidth and flexibility of our algorithms so that we can cover the full range that the human mind dominates. We want to go there, but that will take a long time.

How could you get that far?

I want to keep the idea of ​​a system that can do something without being limited by human knowledge. A baby is not interested in what career it will have or how many children it will one day have. It plays with toys and learns how to use them. There is a tremendous amount that can be learned about the world if there is no definitive learning goal. Our systems could and should work in the same way.

(bsc)

Read comments (4) Go to homepage

Newsletter

Research & technology trends from the upcoming MIT Technology Review at a glance. Appears monthly.

E-mail address

You can find detailed information on the dispatch procedure and your cancellation options in our data protection declaration.

Ad ad