Tyler Farghly

computation · mathematics · music

prof_pic.jpg

PhD student @ Oxford Statistics; working on theoretical foundations of machine learning. Supervised by Patrick Rebeschini and Arnaud Doucet in the OxCSML research group. Currently interested in diffusion models, stochastic optimisation and algorithm-dependent theories of generalisation.

Also a musician, primarily focussed on Jazz drums. Regularly perform around Oxford and London. Occasionally produce electronic music.

news

Sep 2025 Visiting University of Copenhagen Mathematics. Giving a talk on Sept 4th
Jul 2025 ♫ Showing a soundscape as part of a collaborative exhibition, ‘This Place is a Message’ at Mezzanine Studios
Jul 2025 ✎ New preprint! Implicit Regularisation in Diffusion Models: An Algorithm-Dependent Generalisation Analysis
May 2025 ✎ Article featured in Journal for the Philosophy of Planetary Computation: Cognitive Infrastructures: Conjectural Explorations of AI as a Physical Actor in the Wild
May 2025 ♫ Released an album with the Small Claims Trio: Listen to ‘Small Claims’
Mar 2025 Presenting our work at ALT 2025 in Milan: Generalisation under gradient descent via deterministic PAC-Bayes
Jun 2024 Working at Antikythera’s Cognitive Infrastructures Studio as a Studio Researcher over the summer
Apr 2024 ✎ Article featured in JMLR: Mean-Square Analysis of Discretized Itî Diffusions for Heavy-tailed Sampling

selected papers

  1. Diffusion Models and the Manifold Hypothesis: Log-Domain Smoothing is Geometry Adaptive
    T Farghly, P Potaptchik, S Howard, G Deligiannidis, and J Pidstrigach
    Draft available on request 2025
  2. Implicit Regularisation in Diffusion Models: An Algorithm-Dependent Generalisation Analysis
    T Farghly, P Rebeschini, G Deligiannidis, and A Doucet
    arXiv preprint 2025
  3. Towards a Complete Analysis of Langevin Monte Carlo: Beyond Poincaré Inequality
    A Mousavi-Hosseini, T Farghly, Y He, K Balasubramanian, and M Erdogdu
    In COLT 2023
  4. Time-independent Generalization Bounds for SGLD in Non-convex Settings
    T Farghly, and P Rebeschini
    In NeurIPS 2021