Artificial intelligence echoed Copernicus' discovery

Special artificial intelligence for the analysis of scientific data was tested on astronomical objects - the Sun and Mars. Only based on information about their motion in the Earth's celestial sphere. the neural network has built a heliocentric model of the solar system. Thus, in a few minutes, the computer algorithm repeated the achievements of the great scientists of the past. who formulated the laws of astronomy throughout their lives.

Modern science produces so much data that it takes an incredibly long time to process it. Furthermore, a significant amount of information is not studied as well and carefully as scientists would like it to be. So there are several artificial intelligence projects that will do all that. However, there is one problem with modern neural network algorithms - their "thinking" process is a "black box". You cannot just take in and understand how the program came to its conclusions.

That is, even if the AI ​​finds some kind of law of nature, it will be just a finished answer or just a formula, to see which ones you will need to re-do all the calculations. To circumvent this limitation, artificial intelligence was created on the basis of two consecutive neural networks. The first analyzes the data set and finds patterns that are mathematically formulated as concisely as possible. These rules are transferred to another program after a very limited number of links. This, in turn, re-applies these formulas to the data set and is already building a specific model.

Artificial intelligence echoed Copernicus' discovery
Schematic of the classical neural network. Purple circles are entered, orange circles are, and green circles are nodes. There can be a large number of the latter, and the algorithm itself decides what connections will be made between them. Thanks to these connections, a process of information processing, that is, of a kind of thinking, takes place. By analogy with the human brain, which creates new neural connections for each task. It is usually impossible to know exactly how the data handles the nodes. The entire neural network is trained to perform certain tasks (for example, pattern recognition) and gives a definitive result at the output.

how writes Nature, during one of the AI ​​tests, was "fed" with information about how some celestial bodies move through the sky. It was similar to the past of astronomers and mathematicians, using the primitive tools at their disposal to try to understand the universe. The program was able to build a heliocentric model of the solar system without any trace or modification of human data.

Physicist Renato Renner of the Swiss Technical High School in Zurich is pleased with the first results of the project. By creating such a "lobotomized" neural network, whose course of "reasoning" can be traced, the scientist and his colleagues hope for a serious advancement of science. Thanks to such artificial intelligence, it is finally possible to resolve the contradictions between quantum mechanics and other branches of physics. According to him, we may be on the verge of discovering completely new laws of nature right now.

Source link