Geometry & optimization
Studying the structure of neural representations and loss landscapes to understand learning dynamics and improve optimization.
Understanding how neural networks learn, reason, and generalize.
I'm a PhD student at Université de Montréal and a researcher at Mila. I work at the intersection of deep learning theory, optimization, and mechanistic interpretability.
Supervised by Ioannis Mitliagkas and Aristide Baratin.

I connect theoretical questions about neural networks with experiments that reveal how models behave.
Studying the structure of neural representations and loss landscapes to understand learning dynamics and improve optimization.
Connecting a model's internal computations to its behavior, with a focus on vision-based agents and sequential decisions.
Designing controlled benchmarks that reveal how graph models handle memory, delayed effects, and long-range dependencies.
Combining a slowly learned loss geometry with sharpness-aware perturbations to improve neural network optimization.
Controlled synthetic tasks that test temporal graph models on repetition, delayed causal effects, and long-range dependencies.
Attention analysis and interventions in a Minecraft agent reveal clues about memory and a case of goal misgeneralization.
A research note explains what periodic dynamics can reveal about memory, alongside graph prediction failures, incomplete contraction evidence, and a bit retained through 6,144 steps.
Our new preprint explores how learned loss geometry and sharpness-aware minimization can work together.
T-GRAB is published in the ACM SIGKDD proceedings, introducing diagnostic tasks for temporal graph learning.
What attention patterns and interventions can tell us about behavior—and where interpretation needs care.
What periodic dynamics can tell us about recurrent memory—and what graph failures, contracting trajectories, and a bit retained for 6,144 steps actually show.
Before my PhD, I completed an MSc at McGill with David Rolnick and a BSc in Computer Science with a mathematics minor at Sharif University of Technology.
My earlier work spans medical imaging, statistical inference in biology, and speech recognition.
Experience & education ↗Outside research, I enjoy hiking, tennis, and making music. I play the violin, kamancheh, and saxophone.
MehrabInterested in research scientist and research engineer opportunities, and collaborations on understanding and improving learning systems.