Ouais Alsharif
I'm a M.Sc. Candidate at McGill University's School of Computer Science
Proud member of the formidable RL lab. Supervised by Joelle Pineau.
The questions that usually haunt me these days involve learning representations, feature generation and all questions that emanate from these two basic ideas. During my masters, I had the opportunity to work on a diverse set of projects that coalesce under the above title. I started by working on what is, as far as I know, the most accurate text recognition pipeline. Then I created a novel lifelong learning framework that seems to be useful to many problems. I am actively involved in many projects at McGill. Including domain adaptation, deep reinforcement learning and model compression. If you are interested in any of these ideas, or any other ideas that you think might be interesting, just email me.
I was born in Damascus, Syria where I spent 23 years of my life. After graduating first on Damascus University's department of Computer Science Engineering, I left Syria in 2012 to start my masters degree at McGill University in Machine Learning.
At Damascus University. I worked with prof.Nada Ghneim on sentiment analysis for Arabic and opinions mining from product reviews and with prof.Firas Alshamaa on static and dynamic gesture recognition and finger tracking. While there, I was a member of Damascus University's ACM team.
Damascus University Valedictorian Award for ranking 1st on about 250 students in the department of Computer Science Engineering.
Damascus University top student award for four years for ranking 1st on the department.
NSERC Discovery Grant Funding for graduate work
McGill University Differential Fee Waiver for academic standing
Fikra contest 3rd place for innovative projects, awarded startup funding (declined)
Representation as a Service [pdf]
In this paper, we present the idea of a Machine Learning Service Provider (MLSP) facing a continuous stream of problems and tasked with finding rapid functions that solve each problem. This problem is a flavour of what has been called lifelong learning. We frame this problem in the representation learning framework and show that an empirical inter-task generalization error can be parameterized as a function of a representation. Then, we show how that representation can be optimized on new tasks. This leads to a new algorithm that can be used for supervised transfer. We show experiments that validate our formulation on single task transfer, multitask learning and lifelong learning.
Ouais Alsharif, Philip Bachman, Joelle Pineau
End-to-End Text Recognition with Hybrid HMM Maxout Models [pdf] [video]
This is my masters thesis with Joelle Pineau. This work presents a pipeline for constructing a fast and accurate end-to-end text recognition system. We beat previous benchmarks on character recognition, word recognition and end-to-end detection and recognition.
Ouais Alsharif, Joelle Pineau
xkcd | space comics | comics