Watched only short clips, an AI can discover its own physics theories

Open any physics textbook and you will find there are countless different formulas to explain how objects move, such as flying, changing direction or stopping. Although these formulas describe the actions we observe, behind them are physical theories with individual variables to explain them.

Now, a new AI program developed by scientists at Columbia University appears to be uncovering physical theories of its own. After watching a series of videos about physical phenomena on Earth, instead of re-detecting the variables we are currently using, it introduced a new set of variables to account for what it saw.

This does not mean that our current theory of physics is flawed or that there is a better physical model to explain the world around us – for example Einstein’s laws have been proven firmly. . But these laws can only exist by being built on a series of theories and principles established by centuries-old tradition.

But if we were in a different timeline, with other genius minds still solving those problems but from a different perspective, would we be framed in the interpretation of the Universe as it is? are not?

One thing to note is that this AI can only watch videos of a handful of physical phenomena, so don’t expect it to come up with new physics theories to beat Einstein in terms of ability. explain the universe. Anyway, this is not the goal of this study.

” I’ve always wondered, if we were to meet a race of intelligent aliens, would they discover the same laws, ” said roboticist Hod Lipson from Columbia University’s Innovative Machines Lab. physics like us, or could they describe the Universe differently? ”

” In these tests, the number of variables was the same each time the AI ​​was restarted, but the specific variables were very different. It’s true, there are different ways to describe the Universe and very Maybe our choices aren’t perfect .”

What’s more, the team wanted to know if this AI could find new variables on its own – thus helping us to explain complex new phenomena emerging in the current sea of ​​data that we currently cannot. enough theory to understand them.

For example, new data emerging from experiments in the Large Hadron Collider is hinting at new theories of physics.

How does this AI find new physics?

Initially, the researchers would feed it raw videos of known physical phenomena and then ask the program a simple question: What are the minimum fundamental variables to describe things? What is happening?

The first video shows a swinging double pendulum with four state variables in motion: the angle and angular velocity of each pendulum. The AI, after watching the video and the question for a few hours, was able to come up with an answer: This phenomenon requires 4.7 variables to explain – but this doesn’t explain what variables the AI ​​thinks it is. .

So the team tries to compare the known variables with the variables the AI ​​has chosen. The two variables match the pendulum’s inclination relative, but the other two remain a mystery. Even so, the AI ​​can still predict exactly what this pendulum system will do next, so the team thinks that the AI ​​seems to have discovered something that they did not grasp.

The team then showed the IA other videos. The first is an arm that moves in waves like an “air dancer” blowing in the wind – the AI ​​says the system has eight different variables. The scene of the Lava Lamp also shows 8 variables. A clip of the fire shows 24 variables. The variables in each of these videos are unique.

” Without any basic knowledge of physics, our AI algorithm still uncovers the intrinsic aspects of the observed motions and identifies the motions, ” the researchers said in their paper. define sets of potential state variables .”

This suggests that in the future, AI has the potential to help us identify the underlying variables for new physical concepts that are still unknown to us. A potential that cannot be taken lightly.

Refer to ScienceAlert