Mohammad Sadegh Salehi

Mohammad Sadegh Salehi

PhD student in Applied Mathematics (SAMBa CDT)

Department of Mathematical Sciences University of Bath

Biography

I am a PhD student at the University of Bath and a part of the SAMBa CDT, supervised by Dr. Matthias Ehrhardt and Dr. Subhadip Mukherjee. I am also a member of the Numerical Analysis group and the Bath Imaging Group. My interests span Optimization, Machine Learning, Imaging Inverse Problems, and Data Science. My research revolves around algorithms for learning hyperparameters, particularly when the task is modeled using variational regularization approach. This approach is standard in fields like image reconstruction, image processing, machine vision, and data science.

Interests
  • Scalable Methods for Bilevel Learning
  • Variational Methods & Inverse Problems
  • Deep Image Priors
  • Neural Networks as the Regularizer
  • Convex Optimization
Education

  • PhD in Statistical Applied Mathematics, October 2021– present

    University of Bath, UK

  • Master of Applied and Theoretical Mathematics, 2020– 2021

    PSL Research University- Dauphine, Paris, France

  • Bachelor of Mathematics and Applications, 2016– 2020

    University of Tehran, Iran

Recent News

  • (Dec-24) New Research Preprint

    New preprint "Bilevel Learning with Inexact Stochastic Gradients" Subhadip Mukherjee (IIT Kharagpur), Lindon Roberts (University of Sydney), and Matthias Ehrhardt (University of Bath).

  • (Dec-24) New Research Preprint

    New preprint "An Adaptively Inexact Method for Bilevel Learning Using Primal-Dual Style Differentiation" Lea Bogensperger (University of Zurich), Matthias Ehrhardt (University of Bath), Thomas Pock (Graz University of Technology), and Hok Shing Wong (University of Bath). We proposed a posteriori error bounds for the hypergradients calculated through primal-dual style differentiation, combined it with adaptive line search algorithms and showed the application on learning data-adaptive regularisers like input-convex neural network (ICNN).

  • (Sep-24) Conference Presentation in 4th IMA Conference on Inverse Problems

    I presented our research findings 4th IMA Conference on Inverse Problems from Theory to Application, Bath, UK.

  • (Jun-24) Conference Presentation in EUROPT 2024

    I presented our research findings in 21ST Conferenc on advances in continuous optimization, Lund, Sweden.

  • (May-24) Conference Presentation in ICMS Big Data Inverse Problems

    I presented our Algorithm Method of Adaptive Inexact Descent (MAID) in ICMS Workshop on Big Data Inverse Problems as an invited speaker, Edinburgh, UK.

  • (Sep-23) Conference Presentation in EUCCO 2023

    I will be presenting our research findings in European Conference on Computational Optimization (EUCCO), Heidelberg, Germany.

Research Projects

*
  • Scalable Algorithms for Bilevel Learning
    Supervisors: Dr. Matthias Ehrhardt, Dr. Subhadip Mukherjee
  • Modified Analytical Deep Image Priors
    Supervisors: Dr. Tatiana Bubba, Dr. Yuri Korolev
  • Gradient-Free Training of Neural Networks (ADMM Approach)
    Supervisor: Dr. Matthias Ehrhardt
  • Impact of Neural Network Architecture and Initialization on Gradient Confusion and Stochastic Gradient Descent
    Supervisor: Dr. Sandipan Roy
  • Interdisciplinary Research Project on Portfolios Design for Clinical Trials
    Supervisor: Dr. Alex Cox
    Partnership with: Roche Company
  • Master Thesis: Linear and Non-linear Acceleration Methods for Alternating Projections and Dykstra's Algorithm
    Supervisor: Prof. Antonin Chambolle

Publications

Quickly discover relevant content by filtering publications.
M. S. Salehi, S. Mukherjee, L. Roberts, M. J. Ehrhardt (2023). An adaptively inexact first-order method for bilevel optimization with application to hyperparameter learning.

Preprint

L. Bogensperger, M. J. Ehrhardt, T. Pock, M. S. Salehi, H. S. Wong (2024). An Adaptively Inexact Method for Bilevel Learning Using Primal-Dual Style Differentiation.

Preprint

M. S. Salehi, S. Mukherjee, L. Roberts, M. J. Ehrhardt (2024). Bilevel Learning with Inexact Stochastic Gradients.

Preprint

Talks

  • Inexact Algorithms for Bilevel Learning

    International Congress on Industrial and Applied Mathematics (ICIAM) - Waseda University, Tokyo, Japan

  • Inexact Algorithms for Bilevel Learning

    International Conference on Bilevel Optimization - University of Southampton, UK

  • Deep Data-Adaptive Regularisers for Inverse Problems

    Online, Math4DL Seminars (Jointly by the University of Bath, University of Cambridge, and UCL)

Teaching

    University of Bath

  • MA20278: Machine Learning 1 Tutor: Semester 2 2022-2023
  • MA10276: Programming Lab (Python) Tutor: Semester 2 2022-2023
  • MA10207 Analysis 1A Tutor: Semester 1 2021-2022
  • MA10207 Analysis 1B Tutor: Semester 2 2021-2022

Contact