Grab any physics textbook and you’ll find formula after formula describing how things wobble, fly, dodge, and stop. The formulas describe actions that we can observe, but behind each there may be factors that are not immediately obvious.
Now a new AI program developed by researchers at Columbia University has apparently discovered its own alternative physics.
After being shown videos of physical phenomena on Earth, the AI has not rediscovered the actual variables we are using; Instead, it actually evolved new variables to explain what it saw.
To be clear, this does not mean that our current physics are flawed or that there is a better fitting model to explain the world around us. (Einstein’s laws have proven incredibly robust.) But these laws could only exist because they were built on the basis of a pre-existing “language” of theories and principles established through centuries of tradition.
Given an alternate timeline, where other minds approached the same issues from a slightly different perspective, would we still design the mechanisms that explain our universe the same way?
Even with new technology imaging black holes and discovering alien, distant worlds, these laws have endured time and time again (side note: quantum mechanics is a whole different story, but let’s stick to the visible world here).
This new AI has only watched videos of a handful of physical phenomena, so it’s in no way suited to coming up with new physics to explain the universe or try to defeat Einstein. That wasn’t the goal here.
“I’ve always wondered, if we’d ever encountered an intelligent extraterrestrial race, would they have discovered the same physical laws we do, or could they describe the universe in a different way?” says roboticist Hod Lipson of the Creative Machines Lab at the Columbia.
“In the experiments, the number of variables was the same each time the AI was restarted, but the specific variables were different each time. So yes, there are alternative ways to describe the universe, and it’s entirely possible that our choices aren’t perfect.”
In addition, the team wanted to know if AI can actually find new variables – and thus help us explain complex new phenomena emerging in our current data deluge that we currently don’t have the theoretical understanding to keep up with.
For example, the new data emerging from giant experiments like the Large Hadron Collider pointing to new physics.
“What other laws are we missing just because we don’t have the variables?” says mathematician Qiang Du of Columbia University.
So how does an AI find new physics? First, the team fed the system raw video footage of phenomena it already understood and asked the program a simple question: What are the minimum basic variables required to describe what is going on?
The first video showed an oscillating double pendulum known to have four state variables at play: the angle and angular velocity of each of the two pendulums.
The AI pored over the footage and the question for a few hours, then spat out an answer: This phenomenon would require 4.7 variables to explain, it said.
That’s close enough to the four we know… but it still didn’t explain what the AI thought the variables were.
So the team then attempted to match the known variables to the variables the AI had chosen. Two of them loosely matched arm angles, but the other two variables remained a mystery. Still, the AI could make accurate predictions about what the system would do next, so the team figured the AI must have encountered something they couldn’t fully understand.
“We tried to correlate the other variables with anything and everything we could think of: angular and linear velocities, kinetic and potential energies, and various combinations of known quantities,” says software researcher Boyuan Chen, now an assistant professor at Duke University, who developed the line had the job.
“But nothing seemed to fit together perfectly…we don’t yet understand the mathematical language it speaks.”
The team then showed the AI more videos. The first showed an ‘air dancer’ with a wavy arm waving in the wind (the AI said this had eight variables). Lava lamp recordings also produced eight variables. A video clip of flames came back with 24 variables.
The variables were unique each time.
“Without prior knowledge of the underlying physics, our algorithm uncovers the intrinsic dimension of the observed dynamics and identifies candidate sets of state variables,” the researchers write in their paper.
This suggests that, in the future, AI could potentially help us identify variables underpinning new concepts that we are currently unaware of. Watch this area.
The research was published in natural informatics.