Mahdiyar Molahasani

Mahdiyar Molahasani

Senior ML Research Scientist at Captura

About me

Models are trained and judged on averages, and they fail on what the average hides: rare classes, spurious backgrounds, shifted clients, tight compute budgets. I study why, prove what I can, and build fixes that hold up, from a theory connecting long-tailed recognition to continual learning (NeurIPS 2026) to debiasing CLIP using only LLM-written text (ICCV 2025).

At Captura in Vancouver, I work on efficient inference and adaptation for vision-language models that ship in our products. I did my PhD at Queen's University with Ali Etemad and Michael Greenspan.

NeurIPS 2026 · Interactive note Long tails as continual learning Train a model in your browser and watch long-tailed training land next to a model that never saw the rare classes. Guide The Hitchhiker’s Guide to Theoretical ML Where to begin when you have to write a proof: start from the claim, then reach for the tools that usually prove it. ICCV 2025 · Interactive note Fixing CLIP's background shortcut Watch a projection learned only from LLM-written sentences fix how CLIP classifies images. AAAI 2025 · Interactive note Domain shift as gradient disagreement See why domain shift between federated clients shows up in their gradients, and watch one round of gradient alignment.

News

  • Launched The Hitchhiker’s Guide to Theoretical Machine Learning, a field guide to writing proofs in ML papers.
  • Our paper bridging long-tailed recognition and continual learning was accepted to NeurIPS 2026.
  • PRISM was accepted to ICCV 2025.
  • Joined Captura as a Senior Machine Learning Research Scientist.
  • Completed my PhD at Queen's University.

Selected publications

Generalization under imbalance

Bias and fairness in vision-language models

  • PRISM: Reducing Spurious Implicit Biases in Vision-Language Models with LLM-Guided Embedding Projection Mahdiyar Molahasani*, Azadeh Motamedi*, Michael Greenspan, Il-Min Kim, Ali Etemad ICCV 2025 [Paper, arXiv, Code, Interactive note]

Privacy-preserving generalization

  • Federated Unsupervised Domain Generalization using Global and Local Alignment of Gradients Farhad Pourpanah*, Mahdiyar Molahasani*, Milad Soltany*, Michael Greenspan, Ali Etemad AAAI 2025 [Paper, arXiv, Code, Interactive note]
  • A Theoretical Framework for Federated Domain Generalization with Gradient Alignment Mahdiyar Molahasani*, Milad Soltany*, Farhad Pourpanah*, Michael Greenspan, Ali Etemad NeurIPS 2024 Workshop on Mathematics of Modern Machine Learning [Paper]

* Equal contribution

All publications →

Experience & Education

  • Senior Machine Learning Research Scientist, Captura Efficient inference and adaptation for vision-language models (2025–)
  • Graduate Research Assistant, Ingenuity Labs Research Institute Queen's University (2021–2025)
  • Ph.D., Electrical and Computer Engineering, Queen's University With Ali Etemad and Michael Greenspan; thesis on generalization in deep representation learning (2021–2025)
  • M.Sc., Electrical and Computer Engineering, University of Saskatchewan (2019–2021)
  • B.Sc., Electrical Engineering, Iran University of Science and Technology (2013–2018)

Recognition & service

Night sky

The night sky has been my passion for years, and it even found its way into my research.

See the gallery →