February 2017 - Lisa_seminaires - mailman.iro.umontreal.ca

Fwd: Colloque du DIRO, jeudi 23 fév. 15h30
by Yoshua Bengio 20 Feb '17

20 Feb '17

Last ML faculty job talk of the season. Please don't miss it. Also, I would appreciate any feedback on the 6 speakers which we will have gone through (every Monday and Thursday for the last 3 weeks, including this one). One of these could become your collaborator or co-supervisor ;-) ---------- Forwarded message ---------- From: Pierre Lecuyer <lecuyer(a)iro.umontreal.ca> Date: 2017-02-20 17:48 GMT-05:00 Subject: Colloque du DIRO, jeudi 23 fév. 15h30 To: seminaires <seminaires(a)iro.umontreal.ca> Cc: gerad <gerad(a)crt.umontreal.ca> *Principled Tuning for Large-Scale ML Systems* par *Ioannis Mitliagkas* Statistics and Computer Science Departments Stanford University *Jeudi 23 février, 15:30-16:30*, *Salle 6214* Pavillon André-Aisenstadt, Université de Montréal, 2920 Chemin de la Tour Café de 15:00 à 15:30 *Cette conférence sera donnée en anglais.* *Résumé / Abstract:* Modern machine learning systems rely on complex and distributed pipelines that require extensive hyperparameter tuning to achieve the desired performance. Careful tuning can result in significant speedups and improvements in solution quality. However, the dimensionality of the hyperparameter space often makes the use of brute-force search prohibitive. To make things worse, components can interact in unexpected ways and make joint tuning necessary. These challenges preclude non-experts from fully utilizing the potential of modern machine learning tools and call for a deeper understanding of the effect hyperparameters have on the quality and performance of a system. In this talk, I will discuss examples of tuning large-scale learning and inference systems. I will focus on recent work that reveals a previously unknown interaction between system and algorithm dynamics when running an asynchronous learning system. Asynchronous methods are widely used for their superior throughput, but have limited theoretical justification when applied to non-convex problems. I will show that running stochastic gradient descent (SGD) in an asynchronous manner can be viewed as adding a momentum-like term to the SGD iteration. This result does not assume convexity of the objective function, so is applicable to deep learning systems. Furthermore, using a hybrid parallel architecture we can control the level of asynchrony, a new hyperparameter. Theory then implies that jointly tuning momentum and the level of asynchrony can significantly reduce the number of iterations, necessary for an asynchronous system to achieve the same solution. This line of work provides a number of necessary components for realizing the vision of an automated machine learning pipeline. *Courte biographie / Short Bio:* Ioannis Mitliagkas is a Postdoctoral Scholar with the departments of Statistics and Computer Science at Stanford university. He obtained his Ph.D. from the department of Electrical and Computer Engineering at The University of Texas at Austin. His research focuses on understanding and optimizing the scan order for Gibbs sampling, as well as understanding the interaction between optimization and the dynamics of large-scale learning systems. In the past he has worked on high-dimensional streaming problems and fast algorithms and computation for large graph problems. -- Pierre L'Ecuyer, Professeur Titulaire Chaire du Canada en Simulation et Optimisation Stochastique CIRRELT, GERAD, and DIRO, Université de Montréal, Canadahttp://www.iro.umontreal.ca/~lecuyer

1 0

[TeaTalk] Jean-Marc Rousseau, Feb 17, 13:30, AA6214
by Dzmitry Bahdanau 17 Feb '17

17 Feb '17

Hi all, The next tea-talk will be given by Jean-Marc Rousseau on February 17, at 1:30pm, room AA6214. Jean-Marc is the Technology Transfer Director of IVADO and will tell us about entrepreneurship and IVADO. *Title:* Starting a start up; what is involved Best, Dima

2 2

Fwd: Colloque DIRO, lundi 20 février, Xiaodan Zhu (NRC et U Ottawa)
by Yoshua Bengio 16 Feb '17

16 Feb '17

Yet another job talk on Monday. ---------- Forwarded message ---------- From: Pierre McKenzie <mckenzie(a)iro.umontreal.ca> Date: 2017-02-16 20:03 GMT-05:00 Subject: Colloque DIRO, lundi 20 février, Xiaodan Zhu (NRC et U Ottawa) To: seminaires(a)iro.umontreal.ca *Deep Learning for Meaning in Natural Language* par * Xiaodan Zhu * University of Ottawa *Lundi 20 février, 15:30-16:30*, *Salle 6214*, Pavillon André-Aisenstadt Université de Montréal, 2920 Chemin de la Tour Café avant 15:00-15:30 * Cette conférence sera donnée en anglais.* *Résumé:* Text and speech are ubiquitous in our life and are essentially part of the big data we are facing. Developing machine learning models to understand and analyze such data should continue to have significant values. However, modeling human language involves many hard problems in artificial intelligence and is very challenging. this talk I will present our recent deep learning models for learning and representing the meaning of natural language, spanning from words to documents, which involve representing the basic components (e.g., words), compositionality (representing sentences), and document-level modeling. I will present some state-of-the-art results we have achieved on a range of problems and discuss the limitations of the current models. In addition, I would like to briefly discuss some important applications in medical information extraction, social media analysis, and spoken document understanding. *Biographie:* Xiaodan Zhu is a Researcher of National Research Council Canada and an Adjunct Professor of EECS, University of Ottawa. His research interests include natural language processing, deep learning, social media analysis, medical information extraction, and spoken language understanding. Xiaodan received his Ph.D. from the Department of Computer Science at the University of Toronto in 2010 and M.S. from the Department of Computer Science of Tsinghua University in 2000. Xiaodan has research experience with IBM T.J. Watson Research Center, Google, Intel, and Microsoft.

1 0

Fwd: Colloque DIRO, jeudi 16 février, Ted Meeds (Vrije U, Amsterdam)
by Yoshua Bengio 14 Feb '17

14 Feb '17

Another ML professor job talk on Thursday. There might be some time for students or postdocs interested in meeting our visitor, so let Linda know if you are interested. ---------- Forwarded message ---------- From: Pierre McKenzie <mckenzie(a)iro.umontreal.ca> Date: 2017-02-14 14:23 GMT-05:00 Subject: Colloque DIRO, jeudi 16 février, Ted Meeds (Vrije U, Amsterdam) To: seminaires(a)iro.umontreal.ca Cc: veronique.sage(a)umontreal.ca *Modeling science with machine learning: from inference in simulation-based models to understanding cancer.* par * Ted Meeds * Vrije Universiteit Amsterdam *Jeudi 16 février, 15:30-16:30*, *Salle 6214*, Pavillon André-Aisenstadt Université de Montréal, 2920 Chemin de la Tour Café avant 15:00-15:30 * Cette conférence sera donnée en anglais.* *Résumé:* In this talk I will explore the role that machine learning models, algorithms, and inference procedures play in several areas of science. Likelihood-free inference, or approximate Bayesian computation (ABC) is a framework for performing Bayesian inference in a wide variety of simulation-based sciences, from population genetics to computational psychology. I will first describe recent algorithmic improvements in ABC inference which are due in part from viewing inference from a machine learning perspective. I will then describe ongoing work using deep neural networks to model heterogeneous pan-cancer data and challenging tumour segmentation tasks whose long term goal is targeted immunotherapy treatment. Finally I will touch upon recent work on compressing neural networks using simple machine learning techniques, resulting in sparse and quantized parameters which, when applied to scientific domains, provides parsimonious models that are simple enough to be understood by experts.

1 0

Alex Shestopaloff, Feb 10, 13:30, AA6214
by Junyoung Chung 10 Feb '17

10 Feb '17

Hi all, we will have a talk by Alex Shestopaoff on this Friday. Hope to see many of you there! :) *Title:* Sampling latent states for high-dimensional non-linear state space models with the embedded HMM method *Abstract:* We propose a new scheme for selecting pool states for the embedded Hidden Markov Model (HMM) Markov Chain Monte Carlo (MCMC) method. This new scheme allows the embedded HMM method to be used for efficient sampling of state sequences in state-space models where the state can be high-dimensional. Previously, embedded HMM methods were only applied to models with a one-dimensional state space. We demonstrate that using our proposed pool state selection scheme, an embedded HMM sampler can have similar performance to a well-tuned sampler that uses a combination of Particle Gibbs with Backward Sampling (PGBS) and Metropolis updates. The scaling to higher dimensions is made possible by selecting pool states locally near the current value of the state sequence. The proposed pool state selection scheme also allows each iteration of the embedded HMM sampler to take time linear in the number of the pool states, as opposed to quadratic as in the original embedded HMM sampler. We also consider a model with a multimodal posterior, and show how a technique we term ``mirroring'' can be used to efficiently move between the modes. (Joint work with Radford M. Neal). *Bio:* Alex has graduated from U. of T (Prof. Radford Neal's group). Best, --Junyoung

1 1

Fwd: Colloque DIRO, lundi 13 février, Jian Tang (Michigan)
by Yoshua Bengio 09 Feb '17

09 Feb '17

---------- Forwarded message ---------- From: Pierre McKenzie <mckenzie(a)iro.umontreal.ca> Date: 2017-02-09 17:46 GMT-05:00 Subject: Colloque DIRO, lundi 13 février, Jian Tang (Michigan) To: seminaires(a)iro.umontreal.ca *Learning representations of large-scale networks * par * Jian Tang * University of Michigan *Lundi 13 février, 15:30-16:30*, *Salle 6214*, Pavillon André-Aisenstadt Université de Montréal, 2920 Chemin de la Tour Café avant 15:00-15:30 *Cette présentation sera donnée en anglais.* *Résumé:* Information networks (e.g., social networks, citation networks, World Wide Web ) are ubiquitous in real world, covering a variety of applications. Traditionally, networks are usually represented as adjacency matrices. However, this type of representation is very sparse and high-dimensional, which does not facilitate computation of network analysis and network understanding. In this talk, I will introduce our recent work on learning low-dimensional representations of large-scale networks. The representations are able to facility a variety of applications such as node classification, node clustering, link prediction, recommendation, and network visualization.

1 0

[TeaTalk] Prof. Hervé Lombaert, Feb 3, 13:30, AA6214
by Dzmitry Bahdanau 03 Feb '17

03 Feb '17

Hi all, You are invited to attend the talk by Prof. Hervé Lombaert this Friday, Feb 3, at 13:30, room AA6214. *Title:* Spectral Matching & Learning of Surface Data - Example on Brain Surfaces *Abstract:* How to analyze complex shapes, such as of the highly folded surface of the brain? In this talk, I will show how spectral representations of shapes can benefit learning problems where data lives on surfaces. Key operations, such as segmentation and registration, typically need a common mapping of surfaces, often obtained via slow and complex mesh deformations in a Euclidean space. Here, we exploit spectral coordinates derived from the Laplacian eigenfunctions of shapes and also address the inherent instability of spectral shape decompositions. Spectral coordinates have the advantage over Euclidean coordinates, to be geometry aware and to parameterize surfaces explicitly. This change of paradigm, from Euclidean to spectral representations, enables a classifier to be applied *directly* on surface data, via spectral coordinates. The talk will focus, first, on spectral representations of shapes, with an example on brain surface matching, and second, on the learning of surface data, with an example on automatic brain surface parcellation. *Speaker's bio: *Hervé Lombaert is a Starting Research Scientist at Inria Sophia-Antipolis, France, and Associate Professor at ETS, Montreal - with research interests in Statistics on Shapes, Data & Medical Images. He had the chance to work in multiple centers, including Microsoft Research (Cambridge, UK), Siemens Corporate Research (Princeton, NJ), Inria Sophia-Antipolis (France), McGill University (Canada), and Polytechnique Montreal (Canada). He is also a recipient of the François Erbsmann Prize, a top prize in Medical Image Analysis, earned the select NSERC Postdoctoral Fellowship and the FQRNT Étudiant-Chercheur Étoile - more at [ http://cim.mcgill.ca/~lombaert] Best, Dima

1 1

Fwd: Colloque DIRO, lundi 6 février, Devon Hjelm (UdeM)
by Yoshua Bengio 03 Feb '17

03 Feb '17

ML job (faculty) talk! Mark on your calendar. ---------- Forwarded message ---------- From: Pierre McKenzie <mckenzie(a)iro.umontreal.ca> Date: 2017-02-03 11:45 GMT-05:00 Subject: Colloque DIRO, lundi 6 février, Devon Hjelm (UdeM) To: seminaires(a)iro.umontreal.ca *Learning underlying structure with neuroimaging data and training generative models with iterative refinement* par * Devon Hjelm * Université de Montréal *Lundi 6 février, 15:30-16:30*, *Salle 6214*, Pavillon André-Aisenstadt Université de Montréal, 2920 Chemin de la Tour Café avant 15:00-15:30 *Résumé:* Generative models can be used to infer latent structure of observed data for the purpose of advancing domain-specific goals. This is demonstrable with functional and structural magnetic resonance imaging (fMRI / sMRI), where inferred structure can reveal brain function and aid in diagnosis of disease. Further advances in training and inference will increase the applicability of machine learning as a tool for scientific analysis, and the considerable flexibility and capacity allowed by deep learning greatly favor these goals. While deep, continuously-differentiable functions trained by back-propagation have been very successful, local, iterative inference in directed graphical models and generative adversarial networks (GANs) can aid in training, expanding beyond the model's default capacity.

1 0