DeepMind AI is able to make algorithms better than humans have imagined, creators claim

Andrew Griffin
Wednesday 05 October 2022 16:46 BST
<p>Scientists reported on 23 September, 2022, that vast quantum physics equations have been simplified by a new AI algorithm</p>

Scientists reported on 23 September, 2022, that vast quantum physics equations have been simplified by a new AI algorithm

A new artificical intelligence system is able to create algorithms better and more powerful than humans have imagined, its creators say.

DeepMind, the Google sister company that has brought a number of machine learning breakthroughs, says the latest discovery is one step on its mission to use artificial intelligence to advance science.

Despite the fact that algorithms are used everywhere today, finding them is still incredibly difficult. For instance, mathematicians relied on a specific way to do matrix multiplication – and it wasn’t until 1969 that a researcher called Volken Strassen showed that there was actually a better way, which has stayed largely unchanged since.

That kind of algorithm is still used in computers to this day, for everything from sharing videos on the internet to predicting the weather. And so even the most minor change to its efficiency could have widespread effects for the way our world works.

From the beginning, the system was able to discover previously known algorithms, verifying that the system can work effectively.

But DeepMind’s system, called AlphaTensor, has also already found algorithms that are more efficient than the existing state of the art systems for doing those calculations, its creators say. The algorithms that the AI is able to generate perform better than those that scientists have been considering and using for hundreds of years.

If the system is able to do that more and for other purposes, it could lead to considerable changes in the way that computers work, scientists say.

The system works by effectively playing a game with itself: one in which the aim is to find the best way of multiplying two different matrices, or arrays of numbers. DeepMind has already used its AI systems to play games, such as chess or Go, but this is both more productive and far harder.

The game that AlphaTensor is playing sometimes requires around a trillion more actions than those already complex board games. And it is vast as well as complex, with the number of possible algorithms much greater than the number of atoms in the universe.

But by playing that game, the system can very quickly work through different ways of multiplying those matrices. As it did, it found the existing algorithms and went on to find completely new ones, unimagined before, which work even more effectively.

The system does still need some work, the authors say in a new paper published in Nature today. It needs to be given some information to start with, and so might miss some particular algorithms.

But the researchers say that it is another example of how artificial intelligence could rapidly change the state of mathematics, science and more, by finding solutions that people have never been able to think of before.

Join our commenting forum

Join thought-provoking conversations, follow other Independent readers and see their replies


Thank you for registering

Please refresh the page or navigate to another page on the site to be automatically logged inPlease refresh your browser to be logged in