🇫🇷 Version française

I'm an M1 student in Artificial Intelligence at Université Grenoble Alpes (UFR IM²AG and Ensimag), holder of the EFELIA-MIAI excellence scholarship and of the French student-entrepreneur status (SNEE).

Everything below, at more length and with the figures: tewf.github.io

bilinear-tensor-optimization · browse it ↗ · LJK, Université Grenoble Alpes · supervised by Jean-Guillaume Dumas · May to July 2024

The rank of a bilinear map is the number of multiplications it needs; Strassen's seven-instead-of-eight is the known case. A heuristic search over alternative bases takes 5×5 polynomial multiplication over GF(2) from 25 multiplications to 14, and four benchmark maps down by 37% to 44%. It does not prove those decompositions optimal.

tensor-rank-toolkit · browse it ↗ · the same question, continued independently in C++20 on Givaro · v0.4.1

Where the internship searched heuristically, this decides. It runs from a cheap descent up to a SAT encoding whose refutations are checkable as DRAT. On the same 5×5 map over GF(2) that the internship took to 14 products, an exhaustive search settles the rank at exactly 13, refuting 12 over 146 million nodes. Nothing here is ever a float, so a reported rank is a fact about the map and not an artefact of rounding.

Multiplications are half the cost; the additions are set by how many nonzero entries the operators carry. On the operators of a published rank-23 ⟨3,3,3⟩ scheme it takes 221 nonzeros to 128, and 128 is the least any change of basis can leave rather than the best found. main holds what won; rejected-experiments holds what lost, with evidence.

IA-Economie-Strategique · browse it ↗ · the article ↗ · GAEL, UGA and INRAE · supervised by Alexis Garapin and Olivier Bonroy · January to April 2025

Do pricing algorithms sustain tacit collusion or break it? The internship proposed two mechanisms without running either, and I have since run both. The imitation agent finishes eighth of eight, behind a coin flip. Five local language models then played 220 matches: handed a mutually defecting opening, three of the four readable ones never leave it, and a non-binding message frees exactly one. Escaping an imposed regime is a property of the model rather than of the channel.

after-hours · browse it ↗

A CNN learning Flappy Bird from raw pixels, where the largest gain came from the observation rather than a hyperparameter: taking the blue channel instead of a luminance greyscale moved the agent from 0.4 pipes to 12.65 at the same 250k steps. Alongside it, a 3-SAT solver over GF(2) and the French income tax modelled with the Lambert W function.

University-Coursework · browse it ↗

The Licence MIASHS and the M1 AI at UGA, end to end: hedonic pricing where the log-linear specification is what makes the model work, five classifiers evaluated over 24 000 fragrances and the four retained compared on ROC, and a Battleship bot winning 74.3% of 300 round-robin games.