Dr. Kellian Cottart

Kellian Cottart

Machine Learning, Hardware, and Neuromorphic Researcher

◉ Palaiseau, France⌂ C2N, Université Paris-Saclay✉ kellian.cottart@universite-paris-saclay.fr

I bridge algorithmic theory and extreme execution efficiency: deriving discrete Bayesian learning rules, writing custom autograd GPU kernels, and co-designing with physical neuromorphic devices.

Research
Hardware-efficient ML, low-precision and binary architectures
Engineering
Custom GPU kernels (Triton, Pallas); scaled training and data pipelines on HPC (Jean Zay, Slurm)
Published
First author at ICML 2026, second author in Nature Communications

Download the PDFAvailable from January 2027

Education

  1. 2023–2026

    Ph.D. in Machine Learning and Neuromorphic Computing

    Université Paris-Saclay, CNRS / C2N, Palaiseau

    Bayesian Metaplasticity: Uncertainty as a Computational Principle for Resource-Constrained Learning. Advisor: Prof. Damien Querlioz. Defended 5 Oct 2026. Video summary.

  2. 2022–2023

    M.Sc. in Data Sciences (M2DS)

    École Polytechnique, Institut Polytechnique de Paris, Palaiseau

    Reinforcement learning, optimisation for data science, Monte Carlo methods, advanced deep learning, computer vision, data stream processing, big data.

  3. 2018–2023

    Engineering Degree (M.Sc.) in Computer Science

    École Centrale d'Électronique (ECE Paris)

    Software and data engineering major; Erasmus semester in London. Best research project of the minor: a low-cost scanning tunnelling microscope built from scratch.

Experience

  1. 2023–2026

    Doctoral Researcher in Machine Learning

    CNRS, Centre de Nanosciences et de Nanotechnologies (C2N), Palaiseau — advisor: Prof. Damien Querlioz

    • BiMU & hardware co-design (ICML 2026, first author): discrete Bayesian rules for 1-bit weights.
    • Custom fused GPU kernels (Triton/Pallas/JAX) with hand-derived VJPs; multi-GPU HPC scaling on Jean Zay.
    • Binary Bayesian Transformers (JAX/Equinox): 1-bit QK=V cache via variational inference, Gumbel-softmax.
    • MESU (Nature Communications 2025, second author): variance-scaled learning/forgetting, 200+ tasks in JAX.

  2. 2023

    Research Intern, Physics-Informed Deep Learning

    Dassault Systèmes (CATIA Composites R&D), Vélizy-Villacoublay

    • PINN surrogate for micromechanics: refined training-geometry generation (co-author, J. Composite Materials) and predicted nanoscale shear stress.

  3. 2022

    Infrastructure and Data Platform Intern

    Adaltas, Paris

    • Deployed bare-metal Kubernetes clusters and a proof-of-concept distributed data platform with Onyxia (INSEE open-source datalab) and S3 storage.

Publications all publications

  1. ICML 2026

    Active Continual Learning with Metaplastic Binary Bayesian Neural Networks

  2. Nature Communications 2025

    Bayesian Continual Learning and Forgetting in Neural Networks

  3. Nano Letters 2025

    Dynamic Control of Weight-Update Linearity in Magneto-Ionic Synapses

  4. Appl. Phys. Lett. 2026

    Differential Magneto-Ionic Artificial Synapse

  5. J. Composite Materials 2024

    A Unified Method to Generate Representative Volume Elements with Tailored Random Fibre Arrangements to Estimate the Shear and Transverse Behaviours of Unidirectional Continuous Fibres Composite Plies

Talks & Posters all talks

A poster at ICML 2026 (Seoul) and two talks in September 2025 — the Neuromorphic Symposium and the UQ4ML / COMETA workshop at CEA Paris-Saclay — both presenting the same Bayesian continual learning and forgetting result.