Victor Mercklé

PhD Student in Machine Learning @ Grenoble
GitHub - LinkedIn - Resume(CV)


Illustration

Hidden Convexity of ReLU Neural Networks

Peer-reviewed article published in ICLR 2024 and presented as a poster in Vienna. It explains the link betwwen ReLU Neural Networks and Convex Optimisation that has been established but was not well understood. Graphics and training a Neural Network with a Convex formulation are reproducible using a Jupyter Notebook (html)

Illustration

Neural Network Analysis Framework for Research

Gradient Work(GitHub - Python Project) is a research framework for short to long (+20 hours) experiments, targeted at iterative algorithms like gradient descent. It also implements state of the art algorithms for Wasserstein Gradient Flow discretization.
Used and developped during my PhD Thesis to understand the learning dynamics.

Illustration

Monte Carlo Tree Search (MCTS) PSPACE Universal Solver

Conversion of a PSPACE (harder than NP) program to a game tree, which is then solved by my implementation of Monte Carlo Tree Search in OCaml.
This M2 Internship was Supervised by Abdallah Saffidine, Frédéric Maris et Andreas Herzig.
The report can be found here. Code is shared on GitHub(OCaml)

Illustration

Deep Learning GPU Memory Optimization by SAT solving

Describing the problem of optimal GPU memory usage as a NP problem (SAT), developping strategies to find an approximate solution using C++ SAT-solver Glucose.
Supervised by Jean Marie-Lagniez.
Code is shared on GitHub(C++)