Zak Mhammedi

Postdoctoral Associate

Massachusetts Institute of Technology

About Me

I am a Postdoctoral Associate at MIT working with Sasha Rakhlin on theoretical aspects of Reinforcement Learning among other exciting topics. Before that I was a PhD student in computer science at the Australian National University, advised by Bob Williamson and Wouter Koolen.

Interests

Online Learning
Stochastic Optimization
Reinforcement Learning
Statistical Learning (Generalization Bounds)

Education

PhD in Computer Science, 2021
Australian National University
MPhil in Computer Science, 2017
The University of Melbourne
MEng in Mechanical Engineering, 2015
The University of Melbourne
BSc, 2012
Ecole Centrale Paris

Publications

Efficient Projection-Free Online Convex Optimization with Membership Oracle

Risk-Monotonicity in Statistical Learning

NeurIPS 2021 (Oral Presentation${}^{\dagger}$)

Learning the Linear Quadratic Regulator from Nonlinear Observations

Zak Mhammedi, Dylan J. Foster, Max Simchowitz, Dipendra Misra, Wen Sun, Akshay Krishnamurthy, Alexander Rakhlin, John Langford

Pre-Print Poster

Lipschitz and Comparator-Norm Adaptivity in Online Learning

Zak Mhammedi, Wouter M. Koolen

PAC-Bayesian Bound for the Conditional Value at Risk

NeurIPS 2020 (Spotlight Presentation${}^{\star}$)

Zak Mhammedi, Benjamin Guedj, Robert C. Williamson

Pre-Print Poster

Lipschitz Adaptivity with Multiple Learning Rates in Online Learning

Zak Mhammedi, Wouter M. Koolen, Tim Van Erven

PAC-Bayes Un-Expected Bernstein Inequality

Zak Mhammedi, Peter Grünwald, Benjamin Guedj

Geometry Aware Constrained Optimization Techniques for Deep Learning

Soumava Kumar Roy, Zak Mhammedi, Mehrtash Harandi

Constant Regret, Generalized Mixability, and Mirror Descent

NeurIPS 2018 (Spotlight Presentation${}^{\star}$)

Zak Mhammedi, Robert C Williamson

Efficient orthogonal parametrisation of Recurrent Neural Networks using householder reflections

Zak Mhammedi, Andrew Hellicar, Ashfaqur Rahman, James Bailey

Adversarial Generation of Real-time Feedback with Neural Networks for Simulation-based Training

Xingjun Ma, Sudanthi Wijewickrema, Shuo Zhou, Yun Zhou, Zakaria Mhammedi, Stephen O'Leary, James Bailey

${}^\dagger$About 1% of submissions at NeurIPS receive an Orale.

${}^\star$About 3% of submissions at NeurIPS receive a spotlight.

In all papers, authors are ordered according to their contribution.

Experience

Applied Scientist Intern

Amazon Berkeley

Jul 2021 – Oct 2021 Remote

Developed new machine learning techniques for customer query auto-complete.

Research Vendor

Microsoft Research NYC (via Aquent AUS)

Apr 2021 – Jun 2021 Remote

Developed and implemented online parameter-free algorithms such as FreeGrad to be used in a real-world Contextual Bandits application.

Research Intern

Microsoft Research

Feb 2020 – Jun 2020 New York, USA

Worked and published research on a problem of continuous control with rich observations.

Research Intern

Centrum Wiskunde & Informatica (CWI)

Jan 2019 – Mar 2019 Amsterdam, Netherlands

Worked on developing a new and improved PAC-Bayesian generalization bound.

Research Visit

Centrum Wiskunde & Informatica (CWI)

May 2018 – Aug 2018 Amsterdam, Netherlands

Worked on adding Lipschitz adaptivity to certain second-order adaptive online learning algorithms.

Research Assistant

CSIRO

Jul 2016 – Dec 2016 Tasmania, Australia

Published research on an efficient parametrization of orthogonal recurrent neural networks.

Research Intern

IBM Research

Jul 2015 – Oct 2015 Melbourne, Australia

Worked on developing a model for the theoretical spread of citrus greening in Australia. Implemented the model in C++ and performed simulations.

Services

Program Commitee: COLT 2021

Reviewer: COLT 2019, COLT 2020, NeurIPS 2020, ICML 2020, JMLR

Teaching

Tutor

Australian National University

Aug 2017 – Oct 2018 Canberra, Australia

Taught the Information Theory course COMP2610/COMP6261.

Maths Tutor

Simply Maths

Oct 2013 – Oct 2014 Melbourne, Australia

Tutored secondary-level mathematics students.