Projects

In this work we introduce a custom distance metric for non-parametric supervised learning algorithms when each data sample comes as a collection of pairs of means and standard deviations, rather than a single fixed vector. Unlike traditional metrics such Euclidean distance, this approach accounts for uncertainty and variability in each feature, making it more robust to fluctuations. By leveraging variance as informative data, the metric can better discriminate between stable and volatile samples, improving on classic algorithms performances in contexts where distributional information matters.

The Artificial Reasoning Corpus (ARC) is a dataset designed to evaluate machine reasoning capabilities. The ARC Prize 2024 is an award for advancements in AI reasoning, recognizing innovative research that pushes the boundaries of machine reasoning performance using the ARC dataset. The limited number of samples per puzzle prevents statistical learning of puzzle peculiarities, calling for neuro-symbolic approaches instead. Current methods heavily rely on human-crafted elementary transformations, which are combined in search of the puzzle's transformation, a process that falls short of the challenge's goals.

Investigation of Graph Signal Processing (GSP) applied to analysing brain fMRI data. In collaboration with Ines Vati, we leveraged GSP along with other time series analysis tools to predict how much subjects liked pictures shown to them during fMRI recordings. For a 3-class classification task, our best approach achieved a 40% accuracy, which is barely above random chance. This approach involed adapting Dynamic Time Warping metric to a series of features obtained through the graph Fourier transform. This project highlighted the critical need to collaborate with experts to successfully conduct such subtle analysis.

Study of the limitations of a novel Riemannian method for inferring brain connectomes from Diffusion Weighted Imaging (DWI) data, aiming to enhance method's ability to handle fiber crossings effectively. This framework is intended to improve individual tractography methods but still yield doubtable results due to noise in DTI data, when it comes to build a robust average connectome atlas.

I spent the free-time I had during my last 2 years of highscool studying autonomously Olympiad mathematics, on Mathraining. I've took additional courses and attended a summer school to deepen my understanding. In addition, I founded a mathematics club at my high school where I taught Olympiad-level math to other students. I’ve also competed in several contests, and I am proud to have been selected by the POFM, the French organization responsible for preparing students for the International Mathematical Olympiad (IMO)