Artificial intelligence, engaged in physics, can deduce the laws of imaginary universes
After learning AI tricks that physicists use to understand the real world, an extremely powerful machine turns out
There is a famous story about how Galileo watched the lamp swing in the Cathedral of Pisa, and measured it in relation to his pulse. He came to the conclusion that the period is constant and does not depend on amplitude.
Galileo suggested that the pendulum could control the clock, and later developed a similar device, although the first clock of this type was built by Huygens 15 years after the death of Galileo.
Making a discovery, the genius of Galileo ignored all the unpleasant details that could be taken into account - air resistance, temperature, flickering of light, noise, other people, etc. He considered the simplest model of a swinging lamp, using only its period, concentrating on the most noticeable feature.
Many historians believe that the Galilean approach represents the earliest stage in the evolution of the scientific method - the process that gave us flights, quantum theory, electronic computers, General Relativity and artificial intelligence.
In recent years, AI systems have begun to find interesting patterns in the data and even independently derive certain laws of physics. But in these cases, the AI has always studied a particular set of data, isolated from the distractions of the real world. The abilities of these AI systems do not greatly reach the capabilities of people like Galileo.
This raises an interesting question: is it possible to develop an AI system that generates theories, as Galileo did, concentrating on the information needed to explain the various aspects of the world that it observes?
Today, thanksThe work of Taylin Woo and Max Tegmark from MIT, we know the answer. They developed an AI, a replicating approach of Galileo and some other tricks that physicists have learned over several centuries. Their AI Physicist system is able to deduce several laws of physics in mysterious worlds, specially created to simulate the complexity of our Universe.
Wu and Tegmark began by identifying a significant weakness in modern AI. On a large data set, they usually look for a single theory that governs the entire set. But the larger the data set and the more fragmented the data set becomes, the more difficult it is to do. For the current AI it would be impossible to search for the laws of physics in the cathedral.
To cope with this problem, physicists use various thinking methods that simplify the problem. The first is to develop theories that describe a small portion of the data. The result is several theories describing various aspects of data — for example, quantum mechanics, or the theory of relativity. Wu and Tegmark developed AI Physicist to resemble large data sets using the same method.
Another of the basic rules of physicists is Occam's razor, or the idea of the superiority of simple ideas. Therefore, physicists reject theories that require the creator who created the Universe or the Earth: the existence of a creator raises his own set of questions about his nature or origin.
It is known that AIs are prone to issuing overly complex models describing the data on which they are trained. Therefore, Wu and Tegmark also taught the system to prefer simpler theories to complex ones. They used a simple measure of complexity, based on the amount of information that theory covers.
Another of the famous tricks of physicists is the search for ways to unite theories. If one theory is able to cope with the tasks of two, it is most likely better. This prompted physicists to search for a single law governing everything (although there is practically no real evidence of the existence of such a theory).
The last principle that helped physicists in their research: if something worked before, it could work with future tasks. Therefore, AI Physicist from Wu and Tegmark remembers the obtained problem solutions and tries to apply them to future tasks.
Armed with these techniques, Wu and Tegmark sent AI Physicist to work. They developed 40 mysterious worlds, governed by the laws of physics, changing from place to place. In one of these worlds, an abandoned ball can fall under the influence of gravity into an area controlled by an electromagnetic potential, and then fall into an area controlled by a harmonic potential, and so on.
Wu and Tegmark wondered if their AI Physicist could deduce the relevant laws of physics, simply by studying the motion of the ball. They compared the behavior of AI Physicist with the behavior of the “newborn physicist”, using a similar approach, but without learning opportunities, as well as with the work of the classical neural network.
It turns out that both AI Physicist and the “new-born physicist” can deduce the correct laws. “Both subjects are able to sort out more than 90% of all 40 mysterious worlds,” they say.
The main advantage of AI Physicist over the "newborn" is an accelerated learning process and the need for a smaller data set. “It seems that experienced scientists are able to solve new problems faster than a beginner, based on existing knowledge of similar problems,” said Wu and Tegmark.
Their system works much better than a conventional neural network. “Our AI Physicist usually learns faster and produces a standard error of prediction that is billions of times smaller than a standard direct-acting neural network of similar complexity,” they say.
This is an impressive work, suggesting that AI can significantly affect scientific progress. Of course, the real test would be to let AI Physicist into reality, for example, to place it in the Cathedral of Pisa, and see if it would deduce the principle of the mechanical clock. Or push it on other complex data, for example, on data that baffled economists, biologists, and climatologists. This is clearly an easy task for such a system.
And if the work of AI Physicist is successful, the historians of science will be able to consider it the first step of the new era of the evolution of the scientific method from the time of Galileo and his fellow humans. No one knows where she can lead us.