Maxime Gasse – Polytechnique Montréal, Canada
Two topics will be specifically addressed, namely i) multi-label classification using probabilistic graphical model (PGM) structure learning, and 2) deep learning for ultrasound imaging. First, we will see how several popular loss functions (Hamming loss, 0/1 loss, F-loss) affect the parameter and time complexity of the multi-label learning problem, and how these are lowered when the conditional distribution p(y|x) meets some specific independence constraints, i.e. a decomposition into irreducible label factors [1,2,3]. Second, I will present the problem of plane wave medical ultrasound imaging, which I will formulate as a supervised learning problem. I will present some specificities of the problem which call for original contributions, and show some promising results inspired by the HyperNetwork paradigm .
 On the Optimality of Multi-Label Classification under Subset Zero-One Loss for Distributions Satisfying the Composition Property, ICML 2015, Maxime Gasse, Alex Aussem, Haytham Elghazel  F-measure Maximization in Multi-Label Classification with Conditionally Independent Label Subsets, ECML 2016, Maxime Gasse, Alex Aussem  Identifying the irreducible disjoint factors of a multivariate probability distribution, PGM 2016, Maxime Gasse, Alex Aussem  HyperNetwork, ICLR 2017, David Ha, Andrew Dai, Quoc V. Le
Welcome to everyone!
Université de Montréal Campus