Sep-24
Conference Presentation in 4th IMA Conference on Inverse Problems
Presented research findings at the 4th IMA Conference on Inverse Problems from Theory to Application, Bath, UK.
Mohammad Sadegh Salehi
Machine Learning Research Engineer at Zero One Creative
Generative Models - Computer Vision - Optimisation - Inverse Problems
Biography
I am a machine learning researcher and engineer focused on generative 3D, computer vision, optimisation, inverse problems, and spatial AI. At Zero One Creative, I work on Amara, an AI platform for generating and iterating 3D environments from simple prompts.
I built EviBind, an evidence-binding layer for fail-closed LLM tool execution: the model can select policy-admissible evidence, but it cannot invent protected values at the release boundary.
My research spans hyperparameter learning, vision-language and diffusion models, and test-time adaptation across academic and industrial R&D. Academically, I undertook PhD studies and an MRes in Statistical Applied Mathematics under Dr. Matthias Ehrhardt and Dr. Subhadip Mukherjee, after earning an MSc in Applied and Theoretical Mathematics from PSL Research University in Paris.
Interests
Education
PhD studies in Statistical Applied Mathematics
University of Bath, UK
October 2021 - February 2025
Master of Applied and Theoretical Mathematics
PSL Research University - Dauphine, Paris, France
2020 - 2021
Bachelor of Mathematics and Applications
University of Tehran, Iran
2016 - 2020
Contact
- London, UK
- X
- Google Scholar
- GitHub
News
May-26
Bilevel Learning with Inexact Hypergradients talk now available
My talk from 26 May 2026 on Bilevel Learning with Inexact Hypergradients is now available on YouTube. The talk focuses on bilevel optimisation, inexact hypergradients, and scalable learning algorithms.
May-26
AmaraSpatial-10K dataset & preprint released
We released AmaraSpatial-10K, a spatially and semantically aligned 3D dataset of 10,100 synthetic assets for spatial computing and embodied AI. Available on Hugging Face with the accompanying preprint on arXiv.
May-25
SSVM 2025: oral, poster & support grant
Attended SSVM 2025 (Scale Space and Variational Methods in Computer Vision), where I gave an oral presentation of "Fast Inexact Bilevel Optimization for Analytical Deep Image Priors" and presented a poster on "Bilevel Learning with Inexact Stochastic Gradients". Honoured to have received an SSVM support grant to attend the conference.
May-24
Conference Presentation in ICMS Big Data Inverse Problems
Invited speaker at the ICMS Workshop on Big Data Inverse Problems, Edinburgh, UK, presenting MAID.
Jun-26
Presented Amara at Plug and Play UK's Mobility & Physical AI Innovation Day
Shared an update from Plug and Play UK's Mobility & Physical AI Innovation Day, where we presented Amara, Zero One Creative's agentic 3D creator for artist-led 3D prototyping and storytelling.
Jun-24
Conference Presentation in EUROPT 2024
Presented research findings at the 21st Conference on Advances in Continuous Optimization, Lund, Sweden.
Feb-26
We Launched Amara
01C launched Amara, an AI platform for generating and iterating full 3D worlds from simple prompts for creative and production workflows.
Feb-25
Papers accepted to SSVM 2025
Our papers "Bilevel Learning with Inexact Stochastic Gradients" and "Fast Inexact Bilevel Optimization for Analytical Deep Image Priors" accepted to 10th International Conference on Scale Space and Variational Methods in Computer Vision (SSVM 2025).
Dec-24
New Research Preprint (ISGD)
New preprint "Bilevel Learning with Inexact Stochastic Gradients" with Subhadip Mukherjee, Lindon Roberts, and Matthias Ehrhardt.
Dec-24
New Research Preprint (Piggy)
New preprint "An Adaptively Inexact Method for Bilevel Learning Using Primal-Dual Style Differentiation" with Lea Bogensperger, Matthias Ehrhardt, Thomas Pock, and Hok Shing Wong.
Aug-23
Conference Presentation in ICIAM 2023
Presented research findings at the minisymposium "From model-blind to model-aware learning of inverse problems in imaging", in ICIAM 2023, Tokyo, Japan.
Apr-26
Book chapter published in Handbook of Numerical Analysis
Our chapter "Learning regularization functionals for inverse problems: a comparative study" was published in the Handbook of Numerical Analysis (Elsevier). Available here (DOI: 10.1016/bs.hna.2026.04.001).
Apr-25
Papers accepted to SIMODS
Our paper "An Adaptively Inexact First-Order Method for Bilevel Optimization with Application to Hyperparameter Learning" has been accepted to the SIAM Journal on Mathematics of Data Science (SIMODS).
Projects
Evidence Bound Tool Calls
EviBind
Authenticated evidence binding for fail-closed LLM tool execution.
01C-Amara
Authenticated evidence binding for fail-closed LLM tool execution. Trusted code authenticates model-selected evidence, replays derivations, validates the tool contract, and materializes the executable call atomically. The model can select policy-admissible evidence; it cannot invent or rewrite protected values at the release boundary.
0/144AgentDojo banking attacks after binding
Fail-closedrelease boundary
3.11–3.13Python
Open source
diffusion-inverse-problems
Diffusion-based image restoration and inverse problems, with implementations of DiffPIR, DPS, RePaint, and DDRM for inpainting, CT, and deblurring.
NeuralUVAtlas
Official implementation of continuous neural reparameterization as a deep geometric prior for robust fixed-chart UV repair.
GNN-PDE
Physics-informed graph neural networks for learning PDEs on 3D meshes.
MAID
Method of Adaptive Inexact Descent for bilevel learning and hyperparameter optimisation.
ISGD
Inexact stochastic bilevel learning, with code for the SSVM paper on inexact stochastic gradients.
Analytical-Deep-Priors
Fast inexact bilevel optimisation for analytical deep image priors.
Deep_Learning_ADMM
ADMM approach to train neural networks without gradients.
Gradient_confusion
Impact of neural network architecture and initialisation on gradient confusion and stochastic gradient descent.
Amara-spatial-dataset
Scripts for reproducing AmaraSpatial-10K evaluations and benchmarks.
LearnedRegularizers
Code for the comparative study of learning regularisation functionals for inverse problems.
text-to-3D
From prompt and LLM parsing to semantic understanding and 3D generation using diffusion models.
Publications
arXiv preprint2026
Continuous Neural Reparameterization as a Deep Geometric Prior for Robust Fixed-Chart UV Repair
M. S. Salehi
Handbook of Numerical Analysis (Book Chapter, Elsevier)2026
Learning regularization functionals for inverse problems: A comparative study
Johannes Hertrich, Hok Shing Wong, Alexander Denker, Stanislas Ducotterd, Zhenghan Fang, Markus Haltmeier, Zeljko Kereta, Erich Kobler, Oscar Leong, Mohammad Sadegh Salehi, Carola-Bibiane Schoenlieb, Johannes Schwab, Zakhar Shumaylov, Jeremias Sulam, German Shama Wache, Martin Zach, Yasi Zhang, Matthias J Ehrhardt, Sebastian Neumayer
arXiv preprint2026
AmaraSpatial-10K: A Spatially and Semantically Aligned 3D Dataset for Spatial Computing and Embodied AI
M. S. Salehi, A. Perkins, I. Maurell, A. Dabbagh, R. Wong
SIAM Journal on Mathematics of Data Science (SIMODS)2025
An adaptively inexact first-order method for bilevel optimization with application to hyperparameter learning
M. S. Salehi, S. Mukherjee, L. Roberts, M. J. Ehrhardt
Journal of Mathematical Imaging and Vision (JMIV)2025
An Adaptively Inexact Method for Bilevel Learning Using Primal-Dual Style Differentiation
L. Bogensperger, M. J. Ehrhardt, T. Pock, M. S. Salehi, H. S. Wong
SSVM 2025 (Scale Space and Variational Methods in Computer Vision)2025
Bilevel Learning with Inexact Stochastic Gradients
M. S. Salehi, S. Mukherjee, L. Roberts, M. J. Ehrhardt
SSVM 2025 (Scale Space and Variational Methods in Computer Vision)2025
Fast Inexact Bilevel Optimization for Analytical Deep Image Priors
M. S. Salehi, T. A. Bubba, Y. Korolev
2025
Bilevel Learning via Inexact Stochastic Gradient Descent
Mohammad Sadegh Salehi, Subhadip Mukherjee, Lindon Roberts, Matthias J Ehrhardt
Research
Bilevel learning with inexact hypergradients
Adaptive inexact first-order methods for bilevel optimisation and hyperparameter learning, including MAID, ISGD, and primal-dual style differentiation. Published in SIMODS, JMIV, and SSVM 2025.
Analytical deep image priors
Fast inexact bilevel optimisation for analytical deep image priors, with an SSVM 2025 paper and accompanying code.
Learned regularisation for inverse problems
Comparative study of learning regularisation functionals for inverse problems, published as a Handbook of Numerical Analysis chapter.
Spatial 3D assets for embodied AI
AmaraSpatial-10K, a spatially and semantically aligned 3D dataset of 10.1k synthetic assets for spatial computing and embodied AI.
Neural geometric priors for UV repair
Continuous neural reparameterization as a deep geometric prior for robust fixed-chart UV repair.
Diffusion methods for inverse problems
Implementations of DiffPIR, DPS, RePaint, and DDRM for inpainting, CT, and deblurring.
Talks
26 May 2026
Bilevel Learning with Inexact Hypergradients
Recorded talk
A recorded talk on bilevel learning, inexact hypergradients, and scalable optimisation methods for hyperparameter learning.
July 2025
Inexact Stochastic Bilevel Learning
Invited minisymposium - EUROPT 2025, University of Southampton, UK
February 2025
Learned Image Priors and How to Learn Them
Workshop on Recent Advances in Learned Regularisation, UCL, London, UK
June 2024
Inexact First-Order Methods for Bilevel Learning
Invited minisymposium - EUROPT 2024, Lund, Sweden
May 2024
An Adaptively Inexact First-Order Method for Bilevel Learning
Invited talk - ICMS Workshop on Big Data Inverse Problems, Bayes Centre, Edinburgh, UK
September 2023
Scalable Methods for Bilevel Optimisation
Invited minisymposium - EUCCO 2023, Heidelberg, Germany
August 2023
Inexact Algorithms for Bilevel Learning
Invited minisymposium - ICIAM 2023, Tokyo, Japan
AmaraSpatial-10K
AmaraSpatial-10K
A spatial asset bank for embodied AI
An interactive snapshot of the AmaraSpatial-10K release: 10.1k synthetic 3D assets with semantic labels, rendered views, mesh and collision paths, metric anchors, geometry metrics, and text metadata.
10.1kassets
22top categories
159subcategories
CC-BY-4.0license