cv
Basics
| Name | Mehrab Hamidi |
| mehrab[dot]hamidi[at]mila.quebec |
Work
-
2021.01 - 2022.06 Research Intern
McGill University
In this experience I got to work on a biologuical casuality problem with interatablel nature, (SNPs, genotype/phenotype data) and we developed a variational likelihood free sampling method for estimating the posterior distribution.
- Likelihood Free Methods
-
2020.01 - 2021.01 Machine Learning Scientist
AIMed, Iran
I worked on bunch of medical image problems, namely using CT-scan images for detecting Covid-19 and its variants and lung cancer and etc.
- Medical Images
-
2019.01 - 2020.01 Machine Learning Engineer
FANAP, Iran
I worked on a project about Automatic Speech Recognition using cnn-based model to capture time dependent features
- ASR
Education
-
2024.10 - Present Montreal, Quebec
-
2022.10 - 2024.10 Montreal, Quebec
M.S.c
Mcgill University, Montreal, Canada
Computer Science
- Deep Learning Theory
- Mathematical Computer Science
-
2019.10 - 2022.10 Tehran, Iran
-
2017.10 - 2022.10 Tehran, Iran
B.S.c
Sharif University of Technology, Tehran, Iran
Computer Science
- Game Theory
- Convex Optimization
- Bayesian Statistics
Awards
- 2024.10.01
Inter-Math-AI Scholarship
Inter-Math-AI
Publications
-
2024.07.16 Interpretability in Action: Exploratory Analysis of VPT, a Minecraft Agent
ICML2024MI
In this work we tried to understand the behaiviors of the OpenAI's minecraft agent, VPT, by looking at its internal components, attention head, cnns and mlps.
-
2023.11.01 Predicting survival of Iranian COVID-19 patients infected by various variants including omicron from CT Scan images and clinical data using deep neural networks
Heliyon, Elsevier
In this study, we proposed a model for predicting survival based on simple clinical features, blood tests, axial computerized tomography scan images of lungs, and the patients' planned treatment.
-
2023.07.12 Reverse Engineering Deep ReLU Networks An Optimization-based Algorithm
Arxiv
In this work, we proposed an optimization algorithm with convexity garantuess to reverse-engineer NNs, with only querying inputs/outputs without any assumption on the architecture of the netowrk.
Skills
| Programming | |
| Java | |
| C++ | |
| Python | |
| R | |
| Julia | |
| Swift | |
| Matlab | |
| Latex |
| Framworks | |
| Pytorch | |
| Numpy | |
| Networkx |
| Mathematics | |
| Topology | |
| Measure Theory | |
| Game Theory | |
| Optimization |
Languages
| Persian | |
| Native speaker |
| English | |
| Fluent |
| French | |
| Beginner |
Interests
| Interpretability and Explainability | |
| Geometircal Representation Explainibility | |
| Theorical Interpretability | |
| Mechanistic Interpretability |
| Reinforcement Learning | |
| Reward Based Mechaniistic Interpretability | |
| World Models |
| Flow Matching | |
| Energy Matching Wold Models | |
| Optimal Tansport Networks |
References
| Ioannis Mitliagkas | |
| My PhD co-supervisor |
| Aristide Baratin | |
| My PhD co-supervisor |
| David Rolnick | |
| My Master supervisor |
Projects
- 2023.12 - present
Spaghettification in Neural Network Representations
We studied architectural symmetries and geometrical causes related to neural collapse and lossles compression via a phenomena called spaghettification.
- Neural collapse
- Lossless compression
- 2024.07 - present
Synthetic Benchmarks of TGN Methods
This project we focuses on proposing a collections of synthetic tasks specifically designed to benchmark the ability of current TGL methods to capture and model archetypal sequential structures and patterns in dynamic graphs.
- Graph Neural Networks
- Temporal Analysis