Artificial intelligence and science, cross-pollination

toOn 15 April, the European Commission's Scientific Advisory Mechanism (SAM) published a Report on the role of artificial intelligence in science. This draws an observation that applies to any social and economic sector and reaches the usual conclusions: Europe's underdevelopment, the dominance of US corporations, the need for computing power, data, and talent, a “European” ethical approach, the impact of energy, and so on. Although climate, global Earth modelling, materials science, and the humanities and social sciences are briefly mentioned, the report focuses more on devices aimed at facilitating the use of AI in research and innovation than on the place it is set to take globally. The heart of science itself

To be sure, researchers need computing power and sufficient amounts of reliable data to do their work and must be trained in the responsible use of artificial intelligence. However, should we go beyond sharing infrastructures, when scientific communities have their own practices and are better placed to know when and for what purposes to use AI?

Beyond these primarily administrative considerations, AI has already had a tangible impact on science. In 2015, a team of astronomers and machine learning specialists identified… 36 new exoplanets In data collected by NASA to extend the Kepler mission. AlphaFold, developed by DeepMind in 2021, exists so far It predicts the 3D structure of more than 200 million proteins of the corresponding amino acid sequences. Applications of artificial intelligence to the “hard” sciences go much further biology And the Cosmology To diagnose Alzheimer's diseaseAI's contributions also relate to the humanities and social sciences, in particular new methods for analyzing quantitative data in fields ranging from archeology And Ethics (study of animal behavior) History of science And the Paleontology (Study of ancient writings).

Measuring trust

Science is also an enormous field of applications for AI researchers, offering both challenges and opportunities. Challenges, since the millions of annotated data typically needed for supervised training of an exoplanet detector, for example, are simply not available: only a few thousand exoplanets have been discovered so far. We must therefore resolve to use partially synthetic training data, with the advantage of this type of field being that there are accurate models of the physical process of generating the data, unlike most commercial applications.

You have 23.81% of this article left to read. The rest is reserved for subscribers.