Aniss Medbouhi
Doctoral student
Details
Researcher
About me
I am completing a PhD in Computer Science specialized in Machine Learning, under the supervision of Professor Danica Kragic Jensfelt at the Department of Robotics, Perception and Learning within the EECS School.
My PhD research focuses mostly on hyperbolic geometric inference for hierarchical data analysis (or we could say "hyperbolic data analysis"), with some applications to biology and cognitive neuroscience (see publications below).
Topics: geometric data analysis, non-Euclidean computational geometry, hyperbolic machine learning, applications to biology and neuroscience, topological data analysis, brain-computer interface.
My goal is to develop new tools for machine learning and data analysis using mathematics, and then, hopefully, apply them to solve concrete biomedical problems. I am particularly interested in neurorehabilitation. Please feel free to contact me if any interest!
Context and summary of my research:
Hyperbolic geometry is a non-Euclidean geometry characterized by constant negative curvature, which is particularly suitable for low-dimensional tree embedding. In contrast to the Euclidean case, a hyperbolic space requires only two dimensions to embed any tree with arbitrary low distortion due to the exponential volume growth away from the origin (Sarkar 2012). This has motivated the use of hyperbolic geometry in machine learning for representing data that exhibit a hierarchical structure. Examples of such data include geographic communication networks (Kleinberg 2007), internet networks (Boguñá et al. 2010), words in a natural language (Nickel & Kiela 2017, Tifrea et al. 2019), or single-cell data in biology (Klimovskaia et al. 2020). Recent work in dimensionality reduction (Sala et al. 2018, Chami et al. 2021, Guo et al. 2022) and generative modeling (Mathieu et al. 2019, Nagano et al. 2019) has shown that performance in downstream tasks is improved when hierarchical data is represented in a hyperbolic space.
Despite the abundance of hyperbolic machine learning models, methods for geometric inference and data analysis in hyperbolic spaces remain limited.Our research work contributes to the emerging toolbox for hyperbolic data analysis, that is, methods designed to analyze, compare, and interpret data once embedded in hyperbolic space. To this end, we propose HyperDGA to measure the geometric alignment between two sets represented in a hyperbolic space, HyperSteiner to compute a heuristic hyperbolic Steiner minimal tree with applications to hierarchy discovery, and Randomized HyperSteiner to improve the Steiner minimal tree inference when deterministic search is too limited. All three methods are based on a hyperbolic Delaunay triangulation that takes into account the topology and geometry of the data representation. We validate our methods on synthetic data and explore applications to real-life biological data, in particular to hyperbolic embeddings of single-cell RNA sequencing which exhibit a hierarchical structure due to the cellular differentiation process where stem cells specialize into, for example, muscle or neuron cells.
Publications:
Medbouhi*, García-Castellanos*, Marchetti, Pelt, Bekkers, Kragic
International Conference on Artificial Intelligence and Statistics (AISTATS), 2026.
García-Castellanos*, Medbouhi*, Marchetti, Bekkers, Kragic
Proceedings of the Symposium on Algorithm Engineering and Experiments (ALENEX), Society for Industrial and Applied Mathematics (SIAM), 2025.
Medbouhi, Marchetti, Polianskii, Kravberg, Varava, Kragic
Joint European Conference on Machine Learning and Knowledge Discovery in Databases (ECML PKDD), 2024.
Medbouhi, Polianskii, Varava, Kragic
Machine Learning and Knowledge Extraction (MAKE), MDPI journal, 2023.
On-going work:
Medbouhi*, Taleb*, Marchetti, Kragic
Published as an extended abstract at Computational and Systems Neuroscience (COSYNE), 2025.
Accepted as a short paper at Cognitive Computational Neuroscience (CCN), 2025.
Prior to start my PhD:
Previously, I was a research engineer in Machine Learning and Brain-Computer Interface in the same lab supervised by Prof. Danica Kragic Jensfelt and Dr. Ali Ghadirzadeh (Stanford University). Before that, I realized my master thesis, also in the same lab, in Machine Learning and Topological Data Analysis (or "Topological Machine Learning") supervised by Dr. Anastasiia Varava and Vladislav Polianskii.
I studied (mostly) mathematics and physics in scientific Classes Préparatoires aux Grandes Ecoles at the Lycée du Parc in Lyon (France), and then I integrated the Ecole Centrale Marseille (now called Ecole Centrale Méditerranée) where I studied General Engineering with a major in Biomedical Engineering. At that time I realized two research internships: one at the Fresnel Institute (CNRS/Ecole Centrale Marseille/Aix-Marseille University) supervised by Prof. Salah Bourennane where we developped signal processing and machine learning methods to classify if patients are epileptic based on their electroencephalograms (EEG), and a second one in predictive modelling for pharmacokinetics at the Center for Research on Cancer of Marseille (INSERM/CNRS/Aix-Marseille University) in the SMARTc team supervised by Ing. MD. Raphaël Serre, Dr. Elissa Cousin and Prof. Joseph Ciccolini, where we developed an algorithm of toxicity risk prediction for anti-cancer chemotherapy. During my 1st year at Ecole Centrale Marseille, in parallel with my engineering studies I completed a B.Sc. in (pure) Mathematics at Aix-Marseille University. I then continued my studies with a double degree at KTH Royal Institute of Technology through TIME (Top International Managers in Engineering) mobility, where I followed the courses of the master's degree in Machine Learning. Then, you already know... I stayed in Stockholm for my master thesis, research engineering position, and PhD..."la boucle est bouclée".