Machine learning principles and applications in physics

In the first part (3 lectures), we will expose the bases of machine learning (supervised/unsupervised ML, optimization, regularization, neural networks, deep neural networks etc.) as well as the Python tools required to initiate a project (PyTorch).

After this short introduction we will present a set of projects (some are listed below) on which the students can work in groups of two during the whole semester. All the topics are linked with machine learning applied to a physics problem.
After some bibliographical work, students will have to find appropriate datasets and pre-process them to build and implement an algorithm that answers the problem. The work is done in partial autonomy with weekly interactions planned for each group with the supervisors.

A non-exhaustive (and not definitive) list : 
Estimating cosmological parameters from halo distribution in cosmological simulations, 
Augmenting dark matter simulations’ resolution by deep generative models, 
Deep symbolic regression for new physics, 
Contrastive divergence learning in energy-based models, 
Applications of autoregressive networks in statistical mechanics problems, 
Accelerated sampling of energy landscapes with neural networks.

Python basis for numerical analysis (ML-specific libraries are not required), knowledge in statistics (estimation, inference).

The exam consists of a 20-minute oral presentation of the proposed solution but also of the scientific approach followed throughout the project.

Carleo, Giuseppe, Ignacio Cirac, Kyle Cranmer, Laurent Daudet, Maria Schuld, Naftali Tishby, Leslie Vogt-Maranto, and Lenka Zdeborová. "Machine learning and the physical sciences." Reviews of Modern Physics 91, no. 4 (2019) : 045002. 
Goodfellow, Ian, Yoshua Bengio, and Aaron Courville. Deep learning. MIT press, 2016. Freely available here : https://www.deeplearningbook.org/ 
PyTorch : https://pytorch.org/.