Reyhane Askari Hemmat

I'm a PhD candidate at Mila. I work under the supervision of Ioannis Mitliagkas and Nicolas Le Roux. Prior to my PhD, I received my Masters in Computer Science and Bachelor in Computer Engineering from Université de Montréal and Amirkabir University of Technology (Tehran Polytechnic), respectively.

My research interests are in the intersection of machine learning, large scale optimization and generative models.

I am the winner of Borealis AI Graduate Fellowship, the FRQNT three years 'Bourse de doctorat en recherche' and the NSERC's three years Postgraduate PhD Scholarship (PGS-D).

I co-organized the Bridging Game Theory and Deep Learning workshop at NeurIPS 2019, Mila's Deep Learning Theory Group and MTL MLOpt.

Twitter  /  GitHub  /  Google Scholar

LEAD: Least-Action Dynamics for Min-Max Optimization

We propose LEAD (Least-Action Dynamics), a second-order optimizer that uses the principle of least-action from physics to discover an efficient optimizer for min-max games. We subsequently provide convergence analysis of our optimizer in quadratic minmax games using the Lyapunov theory.

Negative Momentum for Improved Game Dynamics

In this paper, we analyze gradient-based methods with momentum on simple games. We prove that alternating updates are more stable than simultaneous updates. Next, we show both theoretically and empirically that alternating gradient updates with a negative momentum term achieves convergence in a difficult toy adversarial problem, but also on the notoriously difficult to train, saturating GANs.

Oríon: Experiment Version Control for Efficient Hyperparameter Optimization

Oríon is a new black-box optimization tool currently in development that is designed to adapt to the workflow of machine learning researchers for minimal obstruction.

Auto Encoders in PyTorch

A quick implementation of Auto Encoder, Denoising Auto Encoders and Variational Auto Encoders in PyTorch.

This was cool :)