School of Informatics  ·  University of Edinburgh

Sahel Torkamani

Sahel Torkamani
About

Who am I?

I am a Ph.D. student in Computer Science at the University of Edinburgh and EPFL, as part of the ELLIS PhD Program. I am supervised by Professor Rik Sarkar and Professor Henry Gouk at Edinburgh, and by Professor Pascal Frossard at EPFL, where I will be a visiting researcher from February 2027. My research focuses on Theoretical Machine Learning and Geometric Deep Learning, in particular geometric tools based on metric space magnitude for understanding how neural networks learn, as well as Differential Privacy.

I completed my B.Sc. in Applied Mathematics at Sharif University of Technology, working with Professor Javad Ebrahimi on Differential Privacy and Information Theory. I also interned at IST Austria with Professor Marco Mondelli on Channel Coding and Machine Learning.

Research Interests

Theoretical Machine Learning Geometric Deep Learning Differential Privacy
Contact

Get In Touch

Work

Research Projects

Hover over a project to read more.

Generalization in Neural Networks Through the Lens of Magnitude Potential

Accepted · NeurIPS 2026

The magnitude potential ratio, computed at the logit layer, correlates with Feldman memorization scores, detects decision boundary shifts, and provides a geometric indicator of grokking.

Magnitude Distance: A Geometric Measure of Dataset Similarity

Accepted · ICML 2026

A novel distance metric on finite datasets based on the magnitude of a metric space, with a tunable scale parameter and strong theoretical guarantees.

Privacy Requires Slicing: Differentially Private Magnitude via Projections

EuroTDP 2026

Sliced magnitude (SMag), with global sensitivity bounded uniformly in scale and dimension, enabling ε-differentially private release via the Laplace mechanism.

Try the interactive demos: Magnitude Potential · Magnitude Distance

Updates

Latest News

Generalization in Neural Networks Through the Lens of Magnitude Potential

September 2026

Paper accepted at NeurIPS 2026.

Privacy Requires Slicing: Differentially Private Magnitude via Projections

September 2026

Paper accepted at the 1st European Workshop on the Theory of Differential Privacy (EuroTDP 2026), ISTA, Austria.

London Hopper Colloquium 2026

May 2026

Awarded the Research Spotlight Speaker Prize at the London Hopper Colloquium 2026, hosted by UCL & BCS Academy of Computing.

Magnitude Distance: A Geometric Measure of Dataset Similarity

May 2026

Paper accepted at ICML 2026.

PrivaDE: Privacy-preserving Data Evaluation for Blockchain-based Data Marketplaces

December 2025

Paper accepted at AsiaCCS 2026.

EurIPS 2025

December 2025

Attending the Workshop on Principles of Generative Modeling (PriGM) @ EurIPS2025.

Count on Your Elders: Laplace vs Gaussian Noise

January 2025

Paper accepted at FORC 2025.

Improved Counting under Continual Observation

August 2024

Paper accepted at TPDP 2024.

Optimal Differential Privacy via Graphs

October 2023

Paper accepted at JSAIT 2023.

Heterogeneous Differential Privacy via Graphs

April 2022

Paper accepted at IEEE ISIT'22.