I graduated from Mines Paristech this year, here is the official nomination. “Théorie et pratique” is the motto of the school. One of the best engineering schools in France, it proposes a fast track for selected people from ENS, Polytechnique and ESPCI, that keeps all the specificities of the school.

The formation includes a project on a industry-based problem. I worked on designing an autonomous robot for children in hospital in one month. The initial aim of the École des mines de Paris, namely to train high-level mining engineers is still present: we analyzed the formation of Paris region, and I follow a lecture on Geostatistics, especially Kriging was historically formalized at the school.

Aside from classical engineering lectures, having a critical mind about the role played by engineers is developped. Students organize a mandatory seminar on environmental challenges and climate change. I am in an option founded by Bruno Latour, focused on Sociology of Sciences.

The MVA master stands for “Mathématiques, Vision, Apprentissage” (Mathematics, Computer Vision, Machine Learning) and was one of the best place to learn a lot from the excellent researchers both theory and applications in machine learning.

One of the difficulty is to chose lectures when you would like to attend nearly every of them. I finally pick 11 of them for validating (minimum is 8). Here the list:

• Graphs in ML Very interesting lectures Tutorials allow to test various algorithms, with application on photo, real time face recognition and image segmentation. Spectral clustering, graph net. I study a paper that propose GenDice, a general stationary distribution correction estimation algorithm.
• Reinforcement Learning One of the very popular lecture of the master, it introduces basics, but also quite advanced topics in RL. Little coding in complement.
• Topological Data Analysis A good introduction to the domain, with not so easy tutorial. I studied the paper Progressive Wasserstein Barycenters of Persistence Diagrams
• Clouds and models An overview of reconstruction methods from clouds of points. Implementation in Python, so it was often to slow for real applications, but it was very clear. I did some experiments on the paper On Fast Surface Reconstruction Methods for Large and Noisy Point Clouds
• Optimal transport One of the most crowded lectures, as everyone wants to hear Gabriel Peyré ! I worked on the paper Optimal Transport to a Variety
• Kernel methods One of the famous lectures of the MVA, given by Julien Mairal and Jean-Philippe Vert. Really focus on the mathematical point of view.
• Convex Optimization Alexandre d’Aspremont really focuses on the mathematical understanding of the methods, and how convexity is often the limit between easy and hard problems. Even if the content seems more basic (I didn’t wait until my last year of Master to use Lagrangien and dual of linear problems.), it was a rich lecture.
• Large Scale Optimization Methods to optimize in parallel fashion, with implementations and derivation of the equations.
• Random Matrix Theory One of the most interesting and difficult lectures. How does spectral clustering works ? Or not ? Complex analysis provides answers and it is beautiful.
• Prediction for Individual Sequences Classic bandits and experts algorithms, with proofs on the blackboard.
• Convolutional Neural networks Lectures of Stéphane Mallat, he puts my lecture notes on the Course’ page

In addition to my scientific curriculum, I wanted to have a better overview of the social impacts and how to place Machine Learning in the philosophical analysis of science. This formation is perfect for acquiring a solid background in philosophy while keeping freedom to chose essays and research work topics.

The teachers were inspiring: always motivated to discover new things and happy to discuss with their students, making connections with other works and authors. I especially thank Alberto Naibo and Maël Pégny for supervising my first year thesis, and Marco Panza for joining my supervision for the final Master Thesis.

My Master Thesis on differential privacy was published on Dumas.

I did the first semester in ENS de Lyon : information theory, parallel programming, optimization and approximation algorithms, compilation and probabilities. I did the second semester at ETH as a SWEP student. I did a research internship with Hoda Heidari, in the LAS group .

ENS Lyon

With a mainly theoretical approach and projects. The lectures include both proofs on halting problems and writing assembly code (I did a Tetris), NP-completeness reduction and preparation to coding competitions.

I realized a research internship of six weeks to finish my Bachelor. I worked in the team Mnemosyne at INRIA Bordeaux under the supervision of Nicolas Rougier. Here is my report.