The theories that allow us to understand physics, and everything around us, are based on variables. One can for example cite Einstein’s famous formula E=MC^2 which describes the relationship between energy, speed and mass.

But finding these variables is in itself one of the most time-consuming tasks. And at the forefront of science, we increasingly find ourselves in situations in which this task is rendered practically impossible by the deluge of data.

This AI could well give a boost to science

This is the case for example in genetics, biology or even cosmology. Researchers therefore wondered if it was possible to entrust the search for variables to a specially trained AI. And to read their study published on July 25 in Nature Computational Science they got off with a few surprises.

Initially, they fed their AI with a video of a phenomenon of which they know the number of variables: the images show a double pendulum oscillating movement which they know depends on four state variables. Yet the AI ​​did not fall on the same number as the researchers.

After several hours of processing, the AI ​​came up with the astonishing figure of 4.7. At first, the researchers figured this was close enough to the 4 expected state variables. Then they gave the AI ​​more images, and tried to pinpoint the exact characteristics of each of the variables the AI ​​found.

This in itself is no small feat: the structure of machine learning algorithms is very difficult to analyze, and the characteristics and deeper meanings of the output responses can be even more so. But with time and determination, they realized that two of the variables correspond to angles, while the others remained impossible to characterize.

Despite this, the AI ​​was able to go the other way, if necessary, with the variables it determined itself – as if it were able to understand the world differently and determine new physical laws.

“We tried to correlate the other variables with anything and everything: angular and linear velocity, kinetic and potential energy, and various combinations of known quantities”explains Boyuan Chen, one of the researchers.

And to add: “but nothing matched perfectly. We do not yet understand the mathematical language that [notre IA] used”. After validating the same observation with other phenomena whose variables are known, the researchers became interested in other phenomena for which they do not know the number of state variables.

A video of a dancer resulted in 8 variables, as well as a video showing a lavalamp in operation. A chimney fire video resulted in 24 variables. For researchers, this improved AI could greatly speed up research – helping researchers tackle the problem of variables with a much faster-than-human tool.

Read also – Ex-Google boss compares artificial intelligence to a nuclear bomb

“Perhaps some phenomena still seem too complex simply because we are trying to understand them with the wrong variables. Although we have used video data in this work, any data source can be employed – radar or DNA for example”concludes the researcher.