The Birth of AI Copernicus! Not Only Discovering Heliocentrism, but Also Offering Hope to Resolve Quantum Mechanics Contradictions

2019-11-09

The sun is the center of the solar system, a fact that took astronomers centuries to figure out. But now, a neural network that learns physics principles on its own has been able to come to the same conclusion based on the movements of the sun and Mars and their positions in the sky as seen from Earth. This feat is the first test of this technology, and researchers hope that they can discover new laws of physics and redefine quantum mechanics through patterns found in large datasets.


Physicists Renato Renner and his colleagues at the Swiss Federal Institute of Technology (ETH) in Zurich wanted to design an algorithm that could distill large datasets into a few fundamental equations, mimicking the concise equations proposed by physicists such as E = MC². To do this, the researchers had to create a new type of neural network, one inspired by the structure of the brain.Traditional neural networks learn to recognize objects, such as images or sounds, by training on massive datasets. They can discover general features, such as "four legs" and "pointy ears," that might be used to identify a cat. They then encode these features in mathematical "nodes" (artificial equivalents of neurons). However, neural networks do not distill information into a few easily interpretable rules like physicists do. Instead, they are black boxes that distribute their acquired knowledge across thousands of nodes in unpredictable and hard-to-explain ways.


Therefore, Renner's team designed an "integrated" neural network: two subnetworks that can communicate with each other through only a few links. The first subnetwork learns from the data like a typical neural network, and the second subnetwork uses this knowledge to make and test new predictions. With very few connections between them, the first network is forced to compress the information it passes to the other side. Renner compares this to how a teacher imparts knowledge to a student.One of the initial tests was to provide the network with simulated data about the movements of Mars and the sun in the sky as seen from Earth. From this perspective, the orbit of Mars around the sun appeared to be unstable, with periodic "retrograde" motion where it reverses its direction. For centuries, astronomers believed that Earth was at the center of the universe and explained Mars' motion through small circles called epicycles. But Nicolaus Copernicus discovered that if both Earth and the planets revolve around the sun, the motion can be predicted with simpler formulas.


Physicist Mario Krenn from the University of Toronto in Canada said that the team's neural network came up with Copernican-like equations for the orbit of Mars, rediscovering "one of the most important shifts in the history of science."Renner emphasizes that while the algorithm derived the equations, humans are still needed to interpret them and understand their relationship to the motion of planets around the sun.


Hod Lipson, a roboticist at Columbia University in New York City, says that this work is important because it can identify key parameters that describe physical systems. He says, "I think this kind of technology is our only hope to understand and keep up with physics and even more complex phenomena."


Renner and his team hope to develop machine learning techniques to help physicists resolve apparent contradictions in quantum mechanics.


Renner says, "The current formulation of quantum mechanics may be to some extent just a human artifact, and a computer could come up with a formulation that doesn't have these contradictions, but our current technology is not yet good enough. To achieve this goal, we are trying to develop a version of our neural network that can not only learn from experimental data but also propose entirely new experiments to test its hypotheses."

RECENT AI NEWS

RECENT AI TOOLS