This exploratory work could one day lead to revolutionary advances in many disciplines.
Modern science is largely based on the principle of iteration. We start from certain simple and verifiable statements to build even more convoluted theories that, once validated in turn, will be used to establish new models, and so on.
This approach has proven to be sound, and today we owe it an immense amount of progress, which has unquestionably advanced our civilization… but that doesn’t necessarily mean it was the only possible clue. Had the circumstances been different, our scientific method might very well have evolved very differently.
This is a question that most science fiction fans have already asked themselves; for example, many, many observers have wondered how an extraterrestrial species could have conceptualized what we call physics or mathematics.
Until very recently, all of this reasoning was more a matter of thought experiment; but the game has started to change with the explosion of artificial intelligence. This technology is incredibly powerful when it comes to juggling different elements that can be very numerous and above all quite abstract. It is for this reason that AI is doing wonders in areas like computer vision.
Reinvent physics from scratch
Therefore, researchers at Columbia University decided to carry out a very original experiment: asked an AI to rediscover the laws of physics on its own that govern the behavior of matter. But, above all, he had to do it only from concrete examples. She did not have access to no theoretical basis such as Newton’s theorems, or any information on geometry.
His work is based on a camera that observes the evolution of a physical system, like a pendulum. And this is the only resource at your disposal. From these simplistic visual examples, the AI is in charge of determining the number of parameters necessary to describe the behavior of the system in question. In a very colorful way, he’s a bit like a scientific genius rediscovering physics in real time in a parallel dimension.
Take the well-known example of the double pendulum: a pendulum hanging from the end of another pendulum. To describe it within the framework of physics as formalized by Newton, it is necessary four parameters — we are talking about state variables — namely, the angle and angular velocity for each of the two arms.
The researchers were therefore curious to see if the AI would also find four parameters, which could possibly indicate that it would have followed the same reasoning as humans. But the proposed answer was very surprising: to describe the double pendulum, the system estimated that it would be necessary… 4.7 settings.
AI has its reasons that reason ignores
At this point, the problem becomes complicated. Because the “reasoning” process of these neural networks is by nature very difficult for humans to decipher; one can understand the meaning of the proposed result, but it is often iIt is impossible to determine precisely what algorithmic tricks allowed the system to reach this conclusion.
Therefore, the researchers could not know what this figure corresponded to, to say the least curious. How the heck can a parameter number be anything other than an integer? What can this 0.7 mean in practice? For humans, does it make any sense to reason with fractional parameters?
In an attempt to answer these questions, researchers have released a bunch of new computer simulations. The objective: to compare these virtual parameters with those of real life. They were able to determine that two of the parameters proposed by the AI corresponded more or less to the angle of the arms… but about the others, they have no idea. And it’s not for lack of looking.
” We try to correlate the other variables with absolutely everything and anything. “, explains Boyuan Chen, lead author of the study. “ Angular and linear velocities, kinetic and potential energies, various combinations of other known parameters… he quotes ” But nothing fit perfectly. “, the Mint. “We still don’t understand the mathematical language that AI speaks”summarizes.
And this is where the problem becomes fascinating. Because even if the researchers did not understand the path of the algorithm from him, he managed to predict the behavior of the studied systems with great accuracy. Conclusion : whatever the reasoning behind this, it works fine. The alternative physics model built by the AI is as efficient as ours, even if it is incomprehensible.
A true generator of “Eureka moments”?
Therefore, the researchers repeated the experiment with other well-documented mechanical systems. And each time, the result was the same: the algorithm always managed to predict the evolution of the mechanical system based on completely new variables that did not correspond to any parameter of Newtonian physics.
“ Without any prior knowledge of the physical mechanisms involved, our algorithm discovered the intrinsic dimensions of the observed dynamics and identified sets of state variables. “, explain the researchers. In short, this AI doesn’t just think outside the box; she frankly imagines new ways to move.
This highly exploratory work may seem as pointless as it is anecdotal, but its implications could be extremely profound. They reinforce the idea thatthere are potentially many other ways to describe observable reality. And some of these approaches could be even more efficient than what we know today.
Therefore, the challenge will be to explore these new approaches in the hope ofidentify those that would be exploitable by humans. This could generate great conceptual revolutions in already very advanced disciplines, where the slightest advance requires an enormous effort of imagination and experimentation on the part of the human being.
Concrete potential in certain areas
In all honesty, there’s little chance that humanity will end up becoming a ” new physics » formalized by an AI; ruining the current foundations of science would probably be counterproductive, at least in the short term. On the other hand, this approach could work wonders in certain disciplines that work on rather obscure phenomena.
The most obvious example is undoubtedly that of quantum computing. Everyone agrees that this technology has enormous potential, but it is still moving rather slowly; some of the underlying mechanisms remain poorly understood, often forcing researchers to feel their way, very empirically.
In such a context, one can imagine that an AI could offer very interesting clues that would then allow humans totackle these problems in a radically different way —enough to pave the way for revolutionary breakthroughs.
Starting from scratch each time, it would be possible to reinvent certain concepts from radically different and potentially more relevant bases. In the case of this study, the parameters formalized by the AI referred to the movement of physical systems, but the concept as a whole goes far beyond this area.
This approach could also be used in much more specific areas as the logistics, urban planning, climatology or public health, for instance. These are activities where AI has already caused great disruption. But until then, only complementary elements allowed to optimize the concepts imagined by humans.
Such a system, on the other hand, could make it possible to highlight phenomena and approaches that would have completely eluded researchers until now… including the operation of the AIs themselves!
Certainly, among the AIs that are already revolutionizing scientific research, those that are writing scientific papers about themselves and works of this type, there is much to be excited about the future of AI in research.
The technical documentation related to this work is available here.
#Une #intelligence #artificielle #découvert #des #lois #alternatives #physique