Mahta Ramezanian-Panahi (@mahtaao)
- Final-year Ph.D. Candidate in Computer Science at the Mila - Quebec AI Institute & Université de Montréal (DIRO, UdeM)
- Lab: AAI CERC (Prof. Irina Rish)
- M.Sc. in Physics and Astronomy from the University of Waterloo
- B.Sc. in Physics from Sharif University of Technology
%20a07e1e52678f47ec9c7fd85df7ed5812/Untitled.png)
Academic Interests:
- Representation learning and identifiability
- Foundation models and tokenization for multivariate time series
- Dynamical systems and world models
- Complex systems
Professional Experience:
- Scientist in Residence, Mila Applied Machine Learning Research Team (AMLRT) & Seoul AI Hub, Montreal [Jun. - Sep. 2026]: 15-week residency in partnership with the City of Seoul; designed personalized, data-driven product recommendation and analysis systems for a consumer product line.
- Resident Scientist, SEVO Bioscience, Quebec & Nova Scotia [Sep. - Dec. 2025]: protein language models; alignment, overlay, and conservation-scoring tools for large-scale sequence modeling; active-learning framework feeding experimental results back into the training pipeline.
- Research Engineer, Hexoskin, Montreal [Jun. - Aug. 2023]: production ML for personalized product analytics with automated preprocessing, quality control, time-series feature extraction, and model retraining.
- Student Researcher, Mila [Apr. 2020 - Jan. 2022]: representation learning and generative models for multivariate time series; latent differential-equation models in Julia; organized and moderated the Dynamical Systems Reading Group.
- Teaching assistant for graduate-level courses at UdeM: Towards AGI (by Irina Rish) and Representation Learning (by Aaron Courville)
- Graduate Research, Teaching & Lab Assistant, University of Waterloo [2017 - 2019]: statistical modeling and computational methods; ran lab and tutorial sessions and mentored 300+ undergraduate students.
- Research Assistant, Sharif University of Technology [2016 - 2017]: multi-scale analysis of positional time series.
Research & Publications:
- Effective Latent Differential Equation Models via Attention and Multiple Shooting, TMLR. video · code
- Generative Models of Brain Dynamics, Frontiers in Artificial Intelligence, 2022. link
- A Position on Causal Representation Learning Without Intervention, single-author position paper, under review at the NeurIPS 2026 Position Paper Track. Argues that identifiability claims from observational data are unfalsifiable without synthetic-ground-truth or interventional benchmarks.
- Validity Conditions for Geometry-Aware EEG Tokenization, ML4Health 2026 (proceedings track). In geometry-aware tokenization of multichannel time series, spatial validity and quantization matter more than backbone pretraining; discrete tokens feeding a Qwen3 causal LM, with language-pretrained vs. random backbones.
- NeuroBuilder: Auditable Agentic Data Curation for Neuro-Foundation Models, co-first author; submitted to SMASH 2026 and a NeurIPS 2026 workshop. LLM agents that turn heterogeneous scientific datasets into validated, auditable data objects.
- Inverse Dynamics Enables Cross-Modal Generalization in World Models, working paper.
- Simulation-Augmented Classification of Interacting Dynamical Systems, working paper. Interpretable random forests over features from large-scale simulated interacting dynamics.
Talks & Presentations:
- OmnEEG: Unifying Spatial Embedding for Scalable Cross-Dataset EEG Foundation Models. Contributed talk; also chaired the Foundation Models I session. International Symposium on Forecasting (ISF), Montreal, 2026.
- Interoperable Foundation Models for Neural Time Series. Contributed talk, UNIQUE Scientific Retreat, Oct. 2025.
- Dynamics Identification for Multivariate Time Series. Lightning talk, Simons Collaboration Kickoff Workshop, Stanford, 2025.
- Time-contrastive Unsupervised Representation Learning for EEG Data. Accepted presentation, INCF Neuroinformatics Assembly, 2024.
- GOKU-UI: Ubiquitous Inference through Attention and Multiple Shooting. Oral spotlight, DLDE III Workshop, NeurIPS 2023, New Orleans.
- Generative Models for Neural Dynamics. Talk, UNIQUE Student Symposium, Université de Montréal, Jun. 2023. video
- Generative Models with Latent Differential Equations in Julia. Lightning talk (co-author), JuliaCon 2021.
Open-Source Software:
- OmnEEG: geometry-aware tokenization for multichannel time series. Converts variable-layout multichannel signals into fixed-length token sequences via parameter-free spatial bases (3D spherical harmonics / 2D spatial interpolation) that preserve channel topology; unifies heterogeneous datasets into one manifest-driven PyTorch pipeline with spatial-layout visualization.
Awards & Honors:
- Mila Scientist in Residence (Seoul AI Hub & City of Seoul partnership) [2026]
- Top-10 team (as a solo participant), EEG Foundation Challenge, NeurIPS 2025 Competition Track [2025]
- IVADO Catalyst Grant, Regroupement 1, 80,000 CAD (foundation models for complex time-series dynamics) [2025]
- Simons Collaboration Travel Award [2025]
- Mila NeurIPS 2025 Travel Scholarship [2025]
- Financement d'UdeM International [2025]
- MoML @ MIT, selected admission award [2025]
- UNIQUE Excellence Scholarship [2023]
- Upper Bound Talent Bursary, North America [2023]
- Amii Talent Bursary [2022]
- UdeM Exemption Scholarship for International Students [2021]
- International Master's and Doctoral Student Awards [2017-2019]
- Marie Curie Graduate Student Award [2017-2019]
- Science Graduate Award [2017-2019]
- RoboCup Competitions, Rescue Simulation League [2nd place, 2011]
- Khwarizmi National Competitions [1st place, 2011]
- Amir Kabir International Competitions [3rd place, 2011]
Volunteer Experience:
- Co-organizer, 8th Neural Scaling Workshop @ NeurIPS 2024: forum on challenges and advances in scaling foundation models, with speakers from OpenAI, MBZUAI, Vector Institute, and Snowflake.
- Student representative, Research Computing Committee (formerly IDT Committee), Mila [Jan. 2025 - Present]
- Reviewer for Workshop on Deep Learning and Differential Equations, NeurIPS 2023
- Reviewer for New in ML Workshop, NeurIPS 2023
- Social Media Coordinator, Data for Good, Waterloo [2019 - 2021]: coordinated regional events and record-breaking volunteer recruitment for datathons and data nights.
- English Instructor, Maison de l'amitié, Montreal [2020 - Present]: placement tests for incoming students and conversation classes.
Personal Interests:
- Fluent in EN, FR, FA
- Committed to causes related to AI governance, fighting climate change, freedom, women's rights, and education justice
Online Presence:
✉️ Contact form
🐈⬛ GitHub
✖️ Twitter
I like to make learning more interpretable. Feel free to message me.