Ouns El Harzli
Interests
In my research, I explore a wide range of mathematical formalisms and tools to try and answer two intertwined (and unsolved) questions in deep learning theory: 1) why do deep neural networks generalise well ? and 2) how to make their predictions explainable ?
Mathematical frameworks that I have used in my research/of interest: formal logic, computational complexity theory, algorithmic information theory, statistical physics, random matrix theory, probability theory, Bayesian statistics, statistical learning theory, continuous optimisation, discrete optimisation.
Deep learning research topics that I have covered in my research/of interest: explainability, generalization, emergent phenomena (e.g. double-descent, grokking, neural collapse), training dynamics (e.g. unconstrained feature models), algorithmic stability, model compression, adversarial attacks, graph neural networks, transformers, diffusion models, large language models, mechanistic interpretability (e.g. sparse auto-encoders), neurosymbolic architectures (e.g. fibring, logic tensor networks), formal verification, Bayesian neural networks.
In terms of applications, I have worked on knowledge graphs, bioinformatics, image generation and more recently, integer programming.
Biography
I am a final-year DPhil student in Computer Science working on mathematical theories of deep learning, in particular interpretability and generalisation. My goal is to develop a principled understanding of the generalisation capabilities of neural-network-based AI systems and use the latter to design interpretability methods with formal guarantees.
Prior to this, I obtained a MSc in Mathematics from University of Oxford and a Diplôme d'Ingénieur from Ecole des Mines de Paris, after attending Classe Préparatoire in Mathematics and Physics. Through various internships, I gained experience in industrial engineering, applied machine learning, operations research and finance. Prior to my DPhil, I also did graduate-level research on computer vision and astrophysics.