Lenka Zdeborová worked at the Atomic Energy Center for the CNRS before becoming, in 2020, professor of statistical physics and computer science at the Federal Polytechnic School of Lausanne (EPFL). She is a member of the editorial board of the journal Machine Learning: Science and Technology, devoted to the contributions of artificial intelligence (AI) for research. In particular, she teaches machine learning courses for physicists and tries to understand how AI works.
Why, as a physicist, do you study AI systems?
In general, an algorithm is a succession of operations that we can understand and follow one by one. A neural network with millions or even billions of parameters is not the same thing. When will it work? Under what conditions? There, one cannot often answer in advance. We therefore lack a theoretical understanding of these objects. I study them like other complex systems, such as the spread of an epidemic or a black hole, with statistical physics tools developed in particular by Giorgio Parisi, the 2021 Nobel Prize in Physics.
What purpose ?
For certain uses, for critical applications or to study systems in science, we have more rigorous criteria than for a simple film recommendation application. We need to have great confidence in the outcome. We need to be able to tell if the program’s response is solid, under what conditions a correct result is obtained, which architecture is better, faster or more accurate than another. We wonder if the same result could be obtained with less training data. But my motivation is also a bit more philosophical.
What do you mean ?
For example, one might wonder if AI is changing the scientific method. Until today, we were a little stuck, in our reasoning, by equations and models. But, now, there are situations where predictions are made without a model, and where hypotheses come out of the data and no longer out of the head of the researcher. We note that, in some cases, predicting the behavior of a system or controlling it can be done without having simple models, unlike in the past. Basically, this situation looks a bit like the ones we experienced by performing very complicated simulations. We might even be a little frustrated that we couldn’t come up with a simple understanding. You will have to get used to the fact that the AI multiplies this type of situation.
You have 36.41% of this article left to read. The following is for subscribers only.
We would love to thank the author of this write-up for this outstanding material
“Behind artificial intelligence, there are always humans”
You can find our social media profiles as well as other related pageshttps://www.ai-magazine.com/related-pages/