- Lisa_seminaires - mailman.iro.umontreal.ca

[TeaTalk] Iulian Serban (Wednesday, Aug 30) and Andrew Jesson (Friday, Sept 1)
by Dzmitry Bahdanau 01 Sep '17

01 Sep '17

Hi all, We are lucky to have two speakers this week! First,* Iulian Serban* will tell us about the progress of MILA team in Alexa challenge (*Z-209 between 14h00 - 15h00, August 30th*). Next, we will learn about deep learning with extreme class imbalance from *Andrew Jesson *from Imagia (*AA6214, 13:45 - 14:45, September 1*). Please find the details below. Dima *Speaker:* Iulian Serban *Title:* A Deep Reinforcement Learning Chatbot *Abstract*: We present MILABOT: a deep reinforcement learning chatbot developed at MILA for the Amazon Alexa Prize competition. MILABOT is capable of conversing with humans on popular small talk topics through both speech and text. The system consists of an ensemble of natural language generation and retrieval models, including sequence-to-sequence and latent variable neural network models. By applying reinforcement learning to crowd-sourced data and real-world user interactions, the system has been trained to select an appropriate response from the models in its ensemble. The system has been evaluated through A/B testing with real-world users, where it performed excellent compared to competing systems. *Bio:* Iulian Serban is a PhD candidate at MILA, where he investigates deep learning and reinforcement learning methods for dialogue and other natural language processing applications. Iulian is supervised by Yoshua Bengio, Aaron Courville and Joelle Pineau. He holds an MSc in machine learning from University College London, and a BSc in theoretical mathematics and statistics from Copenhagen University. Previously, Iulian worked at DigiCorpus ApS, where he developed a physiotherapy computer vision system, now deployed in clinics across Denmark. *Speaker:* Andrew Jesson *Title:* CASED: Curriculum Adaptive Sampling for Extreme Data Imbalance *Abstract: *We introduce CASED, a novel curriculum sampling algorithm that facilitates the optimization of deep learning segmentation or detection models on data sets with extreme class imbalance. We evaluate the CASED learning framework on the task of lung nodule detection in chest CT. In contrast to two-stage solutions wherein nodule candidates are first proposed by a segmentation model and then refined by a second detection stage, CASED improves the training of deep nodule segmentation models (e.g. UNet) to the point where state of the art results are achieved using only a trivial detection stage. CASED improves the optimization of deep segmentation models by allowing them to first learn how to distinguish nodules from their immediate surroundings, while continuously adding a greater proportion of difficult-to-classify global context, until uniformly sampling from the empirical data distribution. Using CASED during training yields a minimalist proposal to the lung nodule detection problem competitive the LUNA16 nodule detection benchmark with an average sensitivity score of 88.7%. Furthermore, we find that models trained using CASED are robust to nodule annotation quality by showing that comparable results can be achieved when only a point and radius for each ground truth nodule are provided during training. Finally, the CASED learning framework makes no assumptions with regard to imaging modality or segmentation target and so should generalize to other medical imaging problems where class imbalance is a persistent problem. *Bio:* Andrew Jesson is the leader of the new A.I. Research team at Imagia. He is currently subitting his Master's thesis in electrical engineering with a focus on medical imaging at McGill University. Andrew has spent the last 12 years in Montreal where he has aslo earned a B.A. in philosophy from Concordia University and a B.Eng from McGill.

2 6

[TeaTalk] Yaniv Romano (Thursday, Aug 24) and Georgy Derevyanko (Friday, Aug 25)
by Dzmitry Bahdanau 25 Aoû '17

25 Aoû '17

Hi all, We have two tea-talks this week! First, at *13:30 on Thursday, August 24, at Z-209*, we will have a presentation by Yaniv Romano, who is a PhD student at Technion, Israel. Second, at *13:45 on Friday, August 25, at AA6214, Georgy Derevyanko*, who is currently a post-doc at Concordia and MILA, will share his research with us. Please find more details below, and hope to see you in great numbers! Dima *Speaker 1:* Yaniv Romano *Title: *A Quest for a Universal Model for Signals: From Sparsity to ConvNets *Abstract: *The celebrated sparse representation model assumes that a signal can be represented as a linear combination of a few columns, also called atoms, taken from a matrix, termed a dictionary. When dealing with high-dimensional signals, addressing the dictionary learning problem becomes computationally infeasible due to the curse of dimensionality. Traditionally, this problem was circumvented by learning a local sparse model on small overlapping patches, extracted from the global signal, and processing (e.g. denoising) these independently. We will start this talk by proposing various approaches to bridge the gap between the efficient independent local processing and the need to model the global signal at hand. In particular, we will describe novel image restoration algorithms, leading not only to state-of-the-art results, but also to a systematic and generic way to boost the performance of many existing algorithms. A different approach to treat high dimensional signals is the convolutional sparse coding (CSC). This global model assumes that a signal can be represented as a superposition of a few local atoms, or small filters, shifted to different positions. A recent work suggested a novel theoretical analysis of this global model, which is based on the observation that while being global, the CSC can be characterized and analyzed locally. We will extend this local-global relation by showing how one can efficiently solve the pursuit problem and train the filters involved, while operating locally on image patches. Armed with these new insights, we proceed by proposing a multi-layer extension of this model, ML-CSC, in which signals are assumed to emerge from a cascade of CSC layers. This, in turn, is shown to be tightly connected to Convolutional Neural Networks (CNN), so much so that the forward-pass of the CNN is in fact the Thresholding pursuit serving the ML-CSC model. This connection brings a fresh view to CNN, as we are able to attribute to this architecture theoretical claims such as uniqueness of the representations throughout the network, and their stable estimation, all guaranteed under simple local sparsity conditions. Lastly, identifying the weaknesses in the above scheme, we propose an alternative to the forward-pass algorithm, which is both tightly connected to deconvolutional and recurrent neural networks, and has better theoretical guarantees. *Bio: *Yaniv Romano received his B.Sc. degree from the Department of Electrical Engineering, Technion – Israel Institute of Technology, in 2012, where he is currently pursuing his Ph.D.. He received the 2015 Zeff fellowship, the 2017 Andrew and Erna Finci Viterbi fellowship, and the 2017 Irwin and Joan Jacobs fellowship. In parallel to his studies, he has been working in the industry since 2011 as an Image Processing Algorithm Developer. The super-resolution technology he invented as an intern in Google Research was launched in 2017, leading to significant bandwidth savings of billions of images. *Speaker 2: *Georgy Derevyanko *Title:* Protein folding project *Abstract:* Protein folding and structure prediction is a 50 years old problem. The solution to it will immediately change the whole industry of drug discovery and biology itself. This talk is a call for collaboration to advance the field of protein structure prediction using deep learning techniques. I will give an overview of the state-of-art algorithms in the field and their assessment procedures. Also, I will present some of our publication-ready and preliminary results. Moreover, I will present the library of differentiable transformations of a protein chemical structure and pre-processed datasets in order to lower the problem entry barrier*.* *Bio:* -2014 PhD in "Physics for life sciences" from Universite Joseph Fourier (Grenoble, France) -2016 PostDoc in experimental structural biology at Forshungszentrum Juelich (Juelich, Germany) -present PostDoc at Concordia CERMM & MILA

2 4

Yacine Jernite, August 4, AA6214, 13:45
by Junyoung Chung 04 Aoû '17

04 Aoû '17

Hi all, our next speaker is Yacine Jernite from NYU. The talk will be at AA6214 on August 4, 13:45. Title: Learning Representations of Language from Text Abstract: Learning good general-purpose representations of written language can be useful for a wide range of applications, but there are still a number of open questions in the field, from model architecture to learning strategies. In this talk, we start by briefly going over these questions, then review some recent work related to two of them: the choice of the reading level for text data and of a learning objective for unsupervised training. First, we describe two recent architectures which look at sub-word information to build word and text representations. Then, considering the prevalence of language modeling as an unsupervised objective, we propose an adaptive hierarchical approach which speeds up learning while avoiding some of the pitfalls of comparable methods. Finally, we propose an alternative discriminative learning objective for sentence representation systems based on discourse modeling. Bio: Yacine is currently pursuing a PhD at NYU in David Sontag's Clinical Machine Learning group, where his research interests include graphical models, applying machine learning to medical data, and language understanding. His current focus is on using unsupervised methods to take advantage of the vast amounts of freely available text data when learning language representation systems. Yacine holds a Diplome d'ingenieur from Ecole Polytechnique (2011) and an M.Sc. in Machine Learning from ENS Cachan (2012). Best, --Junyoung

2 3

[TeaTalk] James Wright, July 28, AA6214, 13:45
by Dzmitry Bahdanau 28 Jul '17

28 Jul '17

Hi all, Our next speaker is James Wright from Microsoft Research NYC. The talk will take place on July 28, at AA6214, 13:45. Hope to see many of there! Title: Deep Learning for Predicting Human Strategic Behavior Abstract: Predicting the behavior of human participants in strategic settings is an important problem in many domains. Most existing work either assumes that participants are perfectly rational, or attempts to directly model each participant's cognitive processes based on insights from cognitive psychology and experimental economics. In this work, we present an alternative, a deep learning approach that automatically performs cognitive modeling without relying on such expert knowledge. We introduce a novel architecture that allows a single network to generalize across different input and output dimensions by using matrix units rather than scalar units, and show that its performance significantly outperforms that of the previous state of the art, which relies on expert-constructed features. Bio: James Wright's research focuses on problems at the intersection of economics, behavioral modeling, and machine learning, with a focus on modeling and predicting human behavior in strategic settings. Prior to joining Microsoft Research NYC, he completed a Ph.D. in Computer Science at the University of British Columbia, where he was advised by Kevin Leyton-Brown. Dima

1 2

Eugene Vorontsov, July 21, AA6214, 13:45
by Junyoung Chung 21 Jul '17

21 Jul '17

Hi all, our next speaker is Eugene Vorontsov. Hope to see you many of you there. When: 13:45, July 21 Where: AA6214 Title: On orthogonality and learning recurrent networks with long term dependencies Abstract: It is well known that it is challenging to train deep neural networks and recurrent neural networks for tasks that exhibit long term dependencies. The vanishing or exploding gradient problem is a well known issue associated with these challenges. One approach to addressing vanishing and exploding gradients is to use either soft or hard constraints on weight matrices so as to encourage or enforce orthogonality. Orthogonal matrices preserve gradient norm during backpropagation and may therefore be a desirable property. This paper explores issues with optimization convergence, speed and gradient stability when encouraging or enforcing orthogonality. To perform this analysis, we propose a weight matrix factorization and parameterization strategy through which we can bound matrix norms and therein control the degree of expansivity induced during backpropagation. We find that hard constraints on orthogonality can negatively affect the speed of convergence and model performance. Bio: I am a PhD student with professors Chris Pal and Samuel Kadoury at École Polytechnique de Montréal and MILA. I am working on medical image segmentation models and have recently begun developing an interest in optimization and regularization of deep neural networks. Prior to MILA, I studied Engineering Science at the University of Toronto, specializing in biomedical engineering. An aspect of computer science that appealed to me is the fast production of experimental results -- unfortunately, I also like big models. --Junyoung

1 3

[TeaTalk] Ethan Perez, July 14, AA6214, 13:45
by Dzmitry Bahdanau 14 Jul '17

14 Jul '17

Hi all, Our next speaker is Ethan Perez, who is currently interning at MILA with Aaron Courville. Please come to AA6214 on July 14 at 13:45! *Title:* Learning Visual Reasoning Without Strong Priors *Abstract:* Achieving artificial visual reasoning - the ability to answer image-related questions which require a multi-step, high-level process - is an important step towards artificial general intelligence. This multi-modal task requires learning a question-dependent, structured reasoning process over images from language. Standard deep learning approaches tend to exploit biases in the data rather than learn this underlying structure, while leading methods learn to visually reason successfully but are hand-crafted for reasoning. We show that a general-purpose, Conditional Batch Normalization method achieves state-of-the-art results on the Compositional Language and Elementary Visual Reasoning (CLEVR) task with a 2.4% error rate. We outperform the next best end-to-end method (4.5%) which uses data augmentation and even methods that use extra supervision (3.1%). We probe our model to shed light on how it reasons, showing it has learned a question-dependent, multi-step process. Previous work has operated under the assumption that visual reasoning calls for a specialized architecture, but we show that a general architecture with proper conditioning can learn to visually reason effectively. *Bio:* Ethan Perez is a rising 4th year computer science undergrad at Rice University. He is currently interning at MILA working with Aaron Courville on Visual Reasoning. Previously, he has researched on deep semi-supervised learning methods and built machine learning models for location detection at Google Maps and fraud detection at Uber. Dima

2 2

Seminar in statistics and Monte Carlo Thursday July 13, 10h45
by Pierre L'Ecuyer 12 Jul '17

12 Jul '17

A seminar of interest for people in machine learning and in statistics: https://www.gerad.ca/fr/events/1479 -- Pierre L'Ecuyer, Professeur Titulaire Chaire du Canada en Simulation et Optimisation Stochastique CIRRELT, GERAD, and DIRO, Université de Montréal, Canada http://www.iro.umontreal.ca/~lecuyer

1 0

[TeaTalk] Rémi Leblond & Jean-Baptiste Alayrac, July 7, AA6214, 13:45
by Dzmitry Bahdanau 07 Jul '17

07 Jul '17

Hi all, For those of wondering if there will be a tea-talk this Friday I have good news - the answer is yes! Our next speakers will be Rémi Leblond and Jean-Baptiste Alayrac, both PhD students at INRIA, working with Simon Lacoste-Julien. Please come to AA6214 on July 7 at 13:45! Please find more information below: 1) Searnn, Rémi Leblond *Title:* SeaRNN: Training RNNs with Global-Local Losses *Abstract:* We propose SEARNN, a novel training algorithm for recurrent neural networks (RNNs) inspired by the “learning to search” (L2S) approach to structured prediction. RNNs have been widely successful in structured prediction applications such as machine translation or parsing, and are commonly trained using maximum likelihood estimation (MLE). Unfortunately, this training loss is not always an appropriate surrogate for the test error: by only maximizing the ground truth probability, it fails to exploit the wealth of information offered by structured losses. Further, it introduces discrepancies between training and predicting (such as exposure bias) that may hurt test performance. Instead, SEARNN leverages test-alike search space exploration to introduce global-local losses that are closer to the test error. We demonstrate improved performance over MLE on three different tasks: OCR, spelling correction and text chunking. Finally, we propose a subsampling strategy to enable SEARNN to scale to large vocabulary sizes. *Bio:* Rémi is a second year Ph.D. student at INRIA in the SIERRA team under the supervision of Simon Lacoste-Julien <http://www.di.ens.fr/~slacoste/>. He's interested in very large scale optimization for new complex machine learning models. His current focus is on parallel and distributed algorithms and structured prediction using neural networks. After graduating from Ecole Polytechnique <https://www.polytechnique.edu/> and Corps des Mines (French MPA), he worked in the Big Data field for three years, as a Data Scientist at SpotRight <http://spotright.com/> in the US (2011-2012) and as a Data Engineer at the Ministry of Defense (2013-2015). 2) Joint discovery of objects and manipulation actions, JB Alayrac *Title:* Joint discovery of objects and manipulation actions *Abstract:* Many human activities involve object manipulations aiming to modify the object state. Examples of common state changes include full/empty bottle, open/closed door, and attached/detached car wheel. In this work, we seek to automatically discover the states of objects and the associated manipulation actions. Given a set of videos for a particular task, we propose a joint model that learns to identify object states and to localize state-modifying actions. Our model is formulated as a discriminative clustering cost with constraints. We assume a consistent temporal order for the changes in object states and manipulation actions, and introduce new optimization techniques to learn model parameters without additional supervision. We demonstrate successful discovery of seven manipulation actions and corresponding object states on a new dataset of videos depicting real-life object manipulations. We show that our joint formulation results in an improvement of object state discovery by action recognition and vice versa. *Bio:* Jean-Baptiste Alayrac received the MS degree in computer science in Ecole Normale Superieure (ENS), in Paris in 2014. He is currently working toward the PhD degree in the research teams WILLOW and SIERRA at INRIA Paris under the supervision of Josef Sivic, Ivan Laptev and Simon Lacoste-Julien. His research focuses on structured prediction from vision and natural language. Dima

1 2

Alexandros Dimakis, June 15th 11AM, AA6214
by Junyoung Chung 15 Jui '17

15 Jui '17

Hi all, we have another tea-talk next week. This talk will be given by Professor Alexandros Dimakis from the University of Texas at Austin. The talk will take place at *AA6214* at *11AM *on June 15th (*Thursday*). Please NOTE that there is a tea-talk on 13th (*Tuesday*) by Jackie C. K. Cheung (McGill University) at *Z109* (Pavillon Claire-McNicoll) at *2PM*. Check Dima's email for more detailed information about the talk. *Alexandros' talk.* *Title: *Generative Models and Compressed Sensing *Abstract: *The goal of compressed sensing is to estimate a vector from an underdetermined system of noisy linear measurements, by making use of prior knowledge on the structure of vectors in the relevant domain. For almost all results in this literature, the structure is represented by sparsity in a well-chosen basis. We show how to achieve guarantees similar to standard compressed sensing but without employing sparsity at all. Instead, we suppose that vectors lie near the range of a generative model, e.g. a GAN or a VAE. We show how the problems of image inpainting and super-resolution are special cases of our general framework. We show how to generalize the RIP condition for generative models and that random gaussian measurement matrices have this property with high probability. A Lipschitz condition for the generative neural network is a key technical condition. We will also discuss on-going work for adding causality and distributed training to these models. (based on joint work with Ashish Bora, Ajil Jalal and Eric Price) Code: https://github.com/AshishBora/csgm Homepage: users.ece.utexas.edu/~dimakis *Bio:* Alex Dimakis is an Associate Professor at the ECE department, University of Texas at Austin. He received his Ph.D. in 2008 from UC Berkeley and the Diploma degree from the National Technical University of Athens in 2003. During 2009 he was a CMI postdoctoral scholar at Caltech. He received an NSF Career award, a Google faculty research award and the Eli Jury dissertation award. He is the co-recipient of several best paper awards including the joint Information Theory and Communications Society Best Paper Award in 2012. He is currently serving as an associate editor for IEEE Transactions on Information Theory. His research interests include information theory, coding theory and machine learning. --Junyoung

1 3

[TeaTalk] Jackie Cheung, June 13, Z-209, 14:00
by Dzmitry Bahdanau 13 Jui '17

13 Jui '17

Hi all, Our next tea-talk will be given by Jackie C. K. Cheung, who is an assistant professor at McGill university. It will take place at Z-209 (Pavillon Claire-McNicoll), on June 13, at 14:00. Hope you to see you in great numbers! *Please note the unusual day and time of the talk!* *Title: *Leveraging External Knowledge for Machine Comprehension of Rare Entities *Abstract: *Thanks to the "long tail" of natural language, entities that only appear a handful of times in a training corpus account for the vast majority of entities overall. Standard techniques for training word embeddings such as word2vec and GloVe do not scale well to modelling these entities, because they depend on having many entity or word occurrences for reliable word co-occurrence statistics. We propose to use external knowledge in the form of short entity definitions as a source of information to train entity representations. We describe experiments in two settings. In the first, we present a simple scheme for initializing entity representations for knowledge base completion. In the second, we introduce a new machine comprehension dataset, the Wikilinks Rare Entities data set, and an accompanying entity cloze task which requires predicting rare named entities in context. We describe a hierarchical double encoder model which reads a text passage and an entity definition in order to predict which entity fits in context. Our model substantially outperforms a number of strong baselines which do not utilize entity definitions, demonstrating the impact of external knowledge in making inferences about rare entities. *Bio:* Jackie CK Cheung is an assistant professor in the School of Computer Science at McGill University. He received his Ph.D. at the University of Toronto, and was awarded a Facebook Fellowship for his doctoral research. He conducts research in computational semantics and natural language generation, with a focus on topics such as inducing event structure from distributional semantics, and automatic summarization. His work is funded by the Natural Sciences and Engineering Research Council of Canada (NSERC), and the Fonds de recherche du Québec - Nature et technologies (FRQNT). Dima

1 2