Despite the success of neural networks in solving concrete problems, their use as a general-purpose scientific tool is still in its infancy.
Among other efforts, scientists from the Federal Polytechnic School of Zurich (ETH-Z) have developed a new neural network architecture called SciNet, which is a little closer to the way of reasoning of the human mind. This has facilitated progress in the area of computer-aided scientific discovery. As proof of concept for these advances, these researchers have applied this new network to several simple settings in Physics. Their results show that it is possible to re-obtain some of the basic laws of nature such as those of conservation (linear momentum, energy, etc.). It is also possible to derive the value of fundamental parameters and constants or even infer the heliocentric view of the sky (where the Sun is at the center of the Solar System), as postulated by Nicolás Copernicus in his work On the gyrations of the celestial orbs of 1543.
Indeed, a neural network that tries to emulate the same reasoning process about the world around us, can be used to extract basic concepts from Physics. In addition, this can be achieved only on the basis of previously measured experimental data. In this context, if you ask such a computer architecture to make predictions about the positions of the Sun and Mars in the sky, it will be able to do so based only on known previous positions. What’s more, you will achieve this milestone without having any idea of the underlying physics (no Newton’s gravitation, no Kepler’s laws, nothing!). This is what SciNet does. This program is capable of, among other wonders, autonomously veering (rediscovering) towards the heliocentric model that we have discussed above.
Internally, SciNet is separated into two parts. The first one compresses the information supplied by the experimenters (the observations) into a simpler representation. The second uses this data to answer questions (which must also be coded appropriately) about the physical system based on this representation. If the reader is aware, this is how the brain of a researcher and the scientific method in general works, very roughly. Interestingly, SciNet makes it possible to retrieve the same results that any student can consult in a science textbook, but without the need to previously supply information that we know about the laws of nature that describe a specific phenomenon. SciNet only needs experimental data and she will make all the necessary deductions herself.
To date, physicists used prior knowledge about the system they were studying. However, neural networks, and specifically SciNet, represent a first step in the use of these new inference techniques to extract new knowledge (or reproduce the existing one, but in another way). But make no mistake, people are still necessary to interpret the results found by the network. The researcher is still essential. The ETH-Z scientists believe that these new methods will eventually complement the traditional scientific method. This is based on the metaphor of the scaffold, where each new discovery rests on more or less solid foundations related to previous knowledge. And while drawing on this preceding understanding is certainly helpful in advancing science, it can also lead to biases, unintentional biases, wrong biases, and preconceptions. Neural networks can provide a new look, perhaps something naive (and even childish, as we mentioned just at the beginning), but simultaneously clean, efficient and ally of the researcher.
Alberto Corbi is a Doctor in Physics and Director of the Degree in Physics at the International University of La Rioja (UNIR).
