Les colloques - Département d'informatique et de recherche opérationnelle https://diro.umontreal.ca/notre-departement/colloques/ Colloques récents fr-CA Wed, 05 Aug 2020 20:20:45 -0400 Wed, 05 Aug 2020 20:20:45 -0400 news-74308 Mon, 10 Aug 2020 15:00:00 -0400 Learning at your fingertips - Vikash Kumar https://diro.umontreal.ca/departement/colloques/colloque/news/detail/News/learning-at-your-fingertips-vikash-kumar/ Learning at your fingertips

Par

Vikash Kumar

Google AI

 

Lundi 10 août, 15:00, sur Bluejeans

Résumé:

While the last decade saw incredible progress in the field of robotic locomotion, with robots increasingly capable of navigating different terrains both indoors and outdoors, current robotic manipulation abilities are not at par with tasks that await at their destination. Most common robots have limited dexterity which restricts their manipulation abilities to simple pick and place operations. Useful tasks found in common homes, shops, hospitals, etc. are far too complex, and rarely involve picking up objects (TV remotes, wallets, scissors, keys, etc.) to simply place it down without using! We need to endow our robots with enhanced dexterity and the ability to exhibit a wide range of manipulation skills to meet the challenges presented by human-centric environments. In this talk, I will outline my efforts towards imparting human-level dexterity to our robots in a scalable way. Dexterous manipulation involves solving for fine motor behaviors that leverage intermittent contact events to delicately balance the object under manipulation. Acquiring such intricate behaviors in the real world while ensuring the stability of the object has proven to be notoriously hard to solve using existing robotics methods. Recently, data-driven techniques have been quite successful in generating motor-skills in simulations and simpler systems. However, these techniques in their current form are less effective in contact-rich dexterous manipulation problems, especially on real robots. This talk will draw insight from the fields of robotics, optimal control, machine learning, and game theory to design algorithms that deliver a new state of the art in standard robotics benchmark problems. On real systems, the proposed techniques scale gracefully to high-dimensional, contact-rich problems, and learns various dexterous manipulation behaviors directly via real-world interactions providing a significant boost to robotic capability towards human-level dexterity.

Biographie :

Vikash Kumar finished his Ph.D. from the University of Washington with Prof. Emo Todorov and Prof. Sergey Levine, where his research focused on imparting human-level dexterity to anthropomorphic robotic hands. He continued his research as a post-doctoral fellow with Prof. Sergey Levine at Univ. of California Berkeley where he further developed his methods to work on low-cost scalable systems. He also spent time as a Research Scientist at OpenAI and Google-Brain where he diversified his research on low-cost scalable systems to the domain of multi-agent locomotion. He has also been involved with the development of the MuJoCo physics engine, now widely used in the fields of Robotics and Machine Learning. His works have been recognized with the best manipulation paper at ICRA’16, the biggest robotics conference, and have been widely covered with a wide variety of media outlets such as NewYorkTimes, Reuters, ACM, WIRED, MIT Tech reviews, IEEE Spectrum, etc.

]]>
news-70721 Thu, 23 Apr 2020 10:30:00 -0400 Show and Tell: Learning to Connect Images and Text for Natural Communication - Malihe Alikhani https://diro.umontreal.ca/departement/colloques/colloque/news/detail/News/show-and-tell-learning-to-connect-images-and-text-for-natural-communication-malihe-alikhani/ Show and Tell: Learning to Connect Images and Text for Natural Communication

Par

Malihe Alikhani

Rutgers University

 

Jeudi 23 avril, 10:30, sur Bluejeans

 

Résumé:

From the gestures that accompany speech to images in social media posts, humans effortlessly combine words with visual presentations. However, machines are not equipped to understand and generate such presentations due to people’s pervasive reliance on common-sense and world knowledge in relating words and images. I present a novel framework for modeling and learning a deeper combined understanding of text and images by classifying inferential relations to predict temporal, causal, and logical entailments in context. This enables systems to make inferences with high accuracy while revealing author expectations and social-context preferences. I proceed to design methods for generating text based on visual input that use these inferences to provide users with key requested information. The results show a dramatic improvement in the consistency and quality of the generated text by decreasing spurious information by half. Finally, I sketch my other projects on human-robot collaboration and conversational systems and describe my research vision: to build human-level communicative systems and grounded artificial intelligence by leveraging the cognitive science of language use.

Biographie :

Malihe Alikhani is a 5th year Ph.D. candidate in the department of computer science at Rutgers University, advised by Prof. Matthew Stone. She is pursuing a certificate in cognitive science through the Rutgers Center for Cognitive Science and holds a BA and MA in Mathematics. Her research aims at teaching machines to understand and generate multimodal communication. She is the recipient of the fellowship award for excellence in computation and data sciences from the Rutgers Discovery Informatics Institute in 2018 and the Anita Borg student fellowship in 2019. Before joining Rutgers, she was a lecturer and an adjunct professor of Mathematics and Statistics for a year at San Diego State University and San Diego Mesa College. She has organized workshops and tutorials at ACL and EMNLP and has served as the program committee of several conferences and journals including ACL and AAAI conferences and is currently the associate editor of the Mental Note Journal.

]]>
news-68914 Mon, 20 Apr 2020 14:00:00 -0400 NLP with Heuristic Knowledge https://diro.umontreal.ca/departement/colloques/colloque/news/detail/News/nlp-with-heuristic-knowledge/ Two Case Studies of Hybrid Approaches to NLP: Collective Social Text Mining and Enhanced Attention Mechanisms for Neural Machine Translation

Par

Xiaohua Liu

 

Lundi 20 avril, 14:00, sur Bluejeans

Pour assister au colloque sur Bluejeans, merci de compléter ce formulaire:
forms.gle/xNmNmvK9pjdPtCHc8

Résumé:

NLP tasks that employ statistical or deep learning heavily depend on annotated data, which is not often available. Properly integrating heuristic knowledge into NLP represents a promising way to attack the lack of annotated data. This talk introduces two groups of pilot studies that leverage heuristic knowledge to boost the performance of NLP with a limited amount of annotated data: 1) A framework of collective social text mining motivated by heuristic knowledge on a social network. It aggregates similar Tweets into a macro context and runs KNNs+CRFs or SVMs+Graph Random Walks to consider the current Tweet and the macro context; and 2) Coverage, context gate and reconstruction for neural machine translation motivated by several well adopted practices in statistical machine translation. They are implemented as auxiliary attention/reconstruction sub neural networks and can be seamlessly integrated into the encoder-decoder-attention framework.

Biographie :

Xiaohua Liu has 14 years of research experiences on NLP and its applications as a principal researcher of Huawei and project lead researcher of Microsoft. He has published 50+ papers on social text mining, machine translation and language learning. He has made contributions to several widely used productions including Bing Dictionary and Xiaodu Zaijia (Baidu Smart Speakers). He received his PhD from Harbin Institute of Technologies in 2011.

]]>
news-70311 Fri, 17 Apr 2020 10:30:00 -0400 Fairness and Algorithmic Discrimination https://diro.umontreal.ca/departement/colloques/colloque/news/detail/News/fairness-and-algorithmic-discrimination/ Fairness and Algorithmic Discrimination

Par

Golnoosh Farnadi

Mila / Université de Montréal

Vendredi 17 avril, 10:30-11:45, sur Bluejeans

Pour assister au colloque sur Bluejeans, merci de compléter ce formulaire:
forms.gle/5KB7KFJXScrnjR6g9

Résumé:

AI and machine learning tools are being used with increasing frequency for decision making in domains that affect peoples' lives such as employment, education, policing and loan approval. These uses raise concerns about biases and algorithmic discrimination and have motivated the development of fairness-aware mechanisms in the machine learning (ML) community and the operations research (OR) community, independently. In this talk, I will show how to ensure that the inference and predictions produced by a learned model are fair. Moreover, I will presents methods to ensure fairness in solutions of an optimization problem. I will conclude my talk with my research agenda to build on the complementary strengths of fairness methods in ML and OR and integrate ideas from them into a single system to build trustworthy AI.

Biographie :

Golnoosh obtained her PhD from KU Leuven and UGent in 2017. During her PhD, she addressed several problems in user modeling by applying and developing statistical machine learning algorithms. She later joined the LINQS group of Lise Getoor at UC Santa Cruz, to continue her work on learning and inference in relational domains. She is currently a post-doctoral IVADO fellow at UdeM and Mila, working with professor Simon Lacoste-Julien and professor Michel Gendreau on fairness-aware AI. She has been a visiting scholar at multiple institutes such as UCLA, University of Washington, Tacoma, Tsinghua University, and Microsoft Research, Redmond. She has had successful collaborations that are reflected in her several publications in international conferences and journals. She has also received two paper awards for her work on statistical relational learning frameworks. She has been an invited speaker and a lecturer in multiple venues and the scientific director of IVADO/Mila "Bias and Discrimination in AI" online course.

]]>
news-70310 Fri, 10 Apr 2020 10:30:00 -0400 Two-player Games in the Era of Machine Learning https://diro.umontreal.ca/departement/colloques/colloque/news/detail/News/two-player-games-in-the-era-of-machine-learning/ Two-player Games in the Era of Machine Learning

Par

Gauthier Gidel

mila / Université de Montréal
https://gauthiergidel.github.io/

Vendredi 10 avril, 10:30-11:45, sur Bluejeans

 

Pour assister au colloque sur Bluejeans, merci de remplir ce formulaire :

https://forms.gle/NB5KfbTpf19ukZ3x6

 

Résumé:

Adversarial training, a special case of multiobjective optimization, is an increasingly useful tool in machine learning. For example, two-player zero-sum games are important for generative modeling (GANs) and for mastering games like Go or Poker via self-play. A classic result in Game Theory states that one must mix strategies, as pure equilibria may not exist. Surprisingly, machine learning practitioners typically train a single pair of agents – instead of a pair of mixtures – going against Nash’s principle. In this talk I will put learning in two-player zero-sum games on a firm theoretical foundation. I will first introduce a notion of limited-capacity-equilibrium for games played with neural networks for which, as capacity grows, optimal agents – not mixtures – can learn increasingly expressive and realistic behaviors. I will then focus on the training of such games by countering some common misconceptions about the difficulties of two-player games optimization and proposing to extend techniques designed for variational inequalities to the training of GANs.

Biographie :

Gauthier Gidel is a PhD candidate at Université de Montréal as a member of Mila and DIRO and has been a research intern at ElementAI and DeepMind. His research is a the interface between machine learning, game theory and optimization on which he co-organized NeurIPS workshops in 2018 and 2019. Gauthier is a recipient of the Borealis AI graduate fellowship as well as the DIRO excellence grant.

]]>
news-68912 Fri, 13 Mar 2020 10:30:00 -0400 Much Morphology, Little Data - Garrett Nicolai https://diro.umontreal.ca/departement/colloques/colloque/news/detail/News/much-morphology-little-data-garrett-nicolai/ Much Morphology, Little Data

Par

Garrett Nicolai

University of British Columbia

 

Vendredi 13 mars 2020, 10:30-12:00Salle 3195, Pavillon André-Aisenstadt

    Université de Montréal, 2920 Chemin de la Tour

Résumé:

Computational Linguistics has seen a renaissance in the past 5 years with the advent of deep learning. New benchmarks are regularly set and broken as traditional tasks are supplemented with powerful neural models such as BERT. These methods, however, are very data-hungry, often requiring many millions of lines of text to train. The sparsity imposed by rich inflectional systems exacerbates this problem significantly. 

In this talk, I will describe the current status of computational inflectional morphology, and describe some recent work I have done in creating and exploiting resources in the low resource sphere. 

Biographie :

Garrett Nicolai is a post-doctoral fellow in the Department of Linguistics at UBC, working on problems related to computational linguistics, particularly when data is small. He received his PhD from the Department of Computing Science at the University of Alberta in 2017, before working as a post-doc with the LORELEI project at Johns Hopkins University. 

]]>
news-68126 Fri, 28 Feb 2020 10:30:00 -0500 Causality : From Learning to Generative Models - Murat Kocaoglu https://diro.umontreal.ca/departement/colloques/colloque/news/detail/News/causality-from-learning-to-generative-models-murat-kocaoglu/ Causality: From Learning to Generative Models

Par

Murat Kocaoglu

MIT-IBM Watson AI Lab

 

Vendredi 28 février, 10:30-12:00Salle 3195, Pavillon André-Aisenstadt

    Université de Montréal, 2920 Chemin de la Tour

Résumé:

Causal inference is fundamental for multiple disciplines ranging from medical research to engineering, statistics and economics. It is also central in machine learning and is now becoming a core component of artificial intelligence research. Although causal inference has been studied for a long time in various fields under different frameworks, today we need tools that can process a large number of variables to handle modern large datasets. The graphical approach to probabilistic causation advocated by Judea Pearl and others provides a way to compactly represent the causal relations using directed acyclic graphs and paves the way for the design of algorithms that can answer causal questions for many variables.

In this talk, I first provide a friendly introduction to causality and explain why causal understanding is important. As my first contribution, I propose a framework called entropic causal inference for inferring the causal direction between two variables from data. I show that entropy can be used to capture the complexity of a causal mechanism. Further, if the true direction has a simple mechanism, we can identify it from data. The entropic causal inference framework leverages tools from information theory for causal inference. As my second contribution, I show how we can apply causality in deep generative models - deep neural networks used for modeling complex data. I demonstrate how to define and train a causal deep generative model, called CausalGAN for generating images with labels. As an extension of generative adversarial networks (GANs), CausalGAN allows sampling not only from the observed data distribution but also from the interventional distributions of images. I conclude with future directions for causal inference and its applications in supervised learning and reinforcement learning. 

Biographie :

I received my B.S. degree in Electrical - Electronics Engineering with a minor degree in Physics from the Middle East Technical University in 2010, and M.S. degree from the Koc University, Turkey in 2012 under the supervision of Prof. Ozgur B. Akan and Ph.D. degree from The University of Texas at Austin in 2018 under the supervision of Prof. Alex Dimakis and Prof. Sriram Vishwanath. I am currently a Research Staff Member in the MIT-IBM Watson AI Lab in IBM Research, Cambridge, Massachusetts. My current research interests include causal inference, generative adversarial networks, and information theory. 

]]>
news-68125 Thu, 27 Feb 2020 10:30:00 -0500 Une technologie vocale plus humaine - Ewan Dunbar https://diro.umontreal.ca/departement/colloques/colloque/news/detail/News/une-technologie-vocale-plus-humaine-ewan-dunbar/ Une technologie vocale plus humaine

Par

Ewan Dunbar

Université de Paris

 

Jeudi 27 février, 10:30-12:00Salle 3195, Pavillon André-Aisenstadt

    Université de Montréal, 2920 Chemin de la Tour

Résumé:

Grâce à l’application d’architectures neuronales performantes à des bases de données énormes, la technologie vocale a fait d'énormes progrès au cours des cinq dernières années (reconnaissance automatique de la parole, synthèse de la parole—tâches de traitement de la parole « superficielles » qui peuvent être effectuées avec peu ou zéro compréhension de la signification des mots). Bien que les tâches de reconnaissance et de génération de la parole ne soient pas entièrement « résolues » (i.e., donnant un résultat indiscernable de celui de l’humain), pour de nombreux cas d’usage, dans plusieurs langues, les résultats sont suffisamment impressionnants pour que les assistants numériques à commande vocale continuent à bien se vendre à des clients satisfaits. En ce qui concerne la science fondamentale, par contre, le travail ne fait que commencer. L’objectif qui motive les projets de recherche de mon groupe est de comprendre intégralement comment les humains traitent la parole. Le problème ne sera pas résolu tant que nous n'aurons pas un modèle computationnel précis de l'humain, qui, au minimum, se comportera exactement de la même manière que l'être humain—non seulement aux tâches quotidiennes, comme écouter des phrases, mais aussi aux tâches expérimentales psychoacoustiques spécifiquement construites et contrôlées pour révéler le fonctionnement des mécanismes sous-jacents. Nous cherchons donc à faire de la rétro-ingénierie de l’humain. Je présenterai des résultats expérimentaux et de modélisation montrant que nous sommes encore loin de cet objectif. Je présenterai des benchmarks et des tâches qui ont pour objectif de propulser le domaine dans cette direction—entre autres, le Zero Resource Speech Challenge, un challenge de machine learning qui cherche à réduire la dépendance de la technologie de la parole aux bases de données de parole labélisées avec du texte, afin de passer à un apprentissage plus autonome et plus semblable à celui de l’humain. De nombreuses applications pratiques seraient possibles grâce à une technologie vocale plus humaine, non seulement une amélioration de la performance à des tâches existantes dans les cas où les systèmes actuels sont encore faillibles, mais aussi à des tâches innovantes, telles que de nouvelles formes de technologie adaptative et éducative. 

Biographie :

Ewan Dunbar est maître de conférences à l'Université de Paris et chercheur au sein de l'équipe Cognitive Machine Learning (CoML) de l'École Normale Supérieure/Inria. Il a obtenu son doctorat de l'Université du Maryland, College Park, en 2013, après des études de baccalauréat et de maîtrise à l'Université de Toronto. Sa recherche porte principalement sur la parole, avec des thématiques en perception, en modélisation computationnelle, et en technologie vocale.

]]>
news-66758 Mon, 17 Feb 2020 10:30:00 -0500 Structure and function in neural networks - David Rolnick https://diro.umontreal.ca/departement/colloques/colloque/news/detail/News/structure-and-function-in-neural-networks-david-rolnick/ Structure and function in neural networks

Par

David Rolnick

University of Pennsylvania

 

Lundi 17 février, 10:30-12:00Salle 3195, Pavillon André-Aisenstadt

    Université de Montréal, 2920 Chemin de la Tour

Résumé:

Despite the success of deep learning, the design of neural networks is still based largely on empirical observation and is limited by a lack of underlying theory. In this talk, I provide theoretical insights into how the architecture of a neural network affects the functions that it can express and learn. I prove that deep networks require exponentially fewer parameters than shallow networks to approximate even simple polynomials, but that there is a massive gap between the maximum complexity of the functions that a network can express and the expected complexity of the functions that it learns in practice. Using these results, I demonstrate how to reverse-engineer the weights and structure of a neural network merely by querying it.

 

Biographie :

David Rolnick is an NSF Mathematical Sciences Postdoctoral Research Fellow at the University of Pennsylvania. His research combines tools from machine learning, mathematics, and neuroscience to provide insight into the behavior of neural networks. He also leads the Climate Change AI initiative, dedicated to helping tackle climate change with machine learning. David received his PhD in Applied Math from MIT as an NSF Graduate Research Fellow; he has also conducted research at Google and DeepMind, and as a Fulbright Scholar.

]]>
news-66756 Thu, 13 Feb 2020 10:30:00 -0500 Natural Language Processing and Text Mining with Graph-Structured Representations - Bang Liu https://diro.umontreal.ca/departement/colloques/colloque/news/detail/News/natural-language-processing-and-text-mining-with-graph-structured-representations-bang-liu/ Natural Language Processing and Text Mining with Graph-Structured Representations

Par

Bang Liu

PhD, University of Alberta

 

Jeudi 13 février, 10:30-12:00Salle 3195, Pavillon André-Aisenstadt

    Université de Montréal, 2920 Chemin de la Tour

Résumé:

In this talk, I will share with the audience my research experiences on a range of NLP tasks, including text matching, text mining, and text generation. I will demonstrate that the graph is a natural way to capture the connections between different text objects, such as words, entities, sentences, and documents. By combining graph-structured representations of text objects at various granularities with Graph Neural Networks (GNNs), significant benefits can be brought to various NLP tasks. Finally, I will share my experience in deploying our algorithms in industry applications, such as Tencent QQ Browser, Mobile QQ and WeChat, for hot event discovery, query and document understanding, as well as news feeds recommendation.

 

Biographie :

Bang Liu received his MSc and PhD degrees in Computer Engineering from the University of Alberta (Canada), and his B.E. degree in Electrical Engineering from University of Science and Technology of China. His research interests primarily lie in the areas of natural language processing (NLP), data mining, and applied machine learning. Bang has produced visible values to both academia and industry. His innovations have been deployed in real-world applications, serving over a billion daily active users. He has 15 papers published or accepted by top conferences and journals such as SIGMOD, ACL, KDD, WWW, ICDM, CIKM, TKDD, etc., as well as multiple manuscripts under submission. His research on NLP had helped his co-supervisor's research team win the Extraordinary Achievement Award for 2016-2017 CCF-Tencent Rhino Bird Open Grant

]]>
news-56868 Wed, 26 Jun 2019 10:00:00 -0400 Irina Rish - Towards the Synergy Between AI and Neuroscience https://diro.umontreal.ca/departement/colloques/colloque/news/detail/News/irina-rish-towards-the-synergy-between-ai-and-neuroscience/ Towards the Synergy Between AI and Neuroscience

par

Irina Rish

AI Foundations department
IBM T.J. Watson Research Center

 

 

Mercredi 26 juin 2019, 10:00-11:00, Salle 3195, Pavillon André-Aisenstadt

Université de Montréal, 2920 Chemin de la Tour

La conférence sera présentée en anglais

 

Résumé:

Despite its remarkable recent advances, AI is still far from achieving human-level intelligence, and making further progress in that direction may require developing fundamentally new approaches to move us from today’s mostly “narrow”/task-specific AI towards a “broad”/multi-functional, continually learning, and self-evolving AI, and eventually to general/human-level AI. One promising avenue of research which is believed to have a potential for revolutionizing the field is to explore more biologically-inspired mechanisms behind the human intelligence. This approach has already proved to be useful before, giving rise to modern deep learning and reinforcement learning, but there is a wealth of untapped knowledge about the brain functioning accumulated in neuroscience, psychology and related disciplines, that can help us bring AI to the next level. On the other hand, introducing AI ideas, models and techniques to neuroscience, psychology and mental health also has a potential of revolutionizing those fields, so there is a clear need for more synergy between the studies of artificial and natural intelligence. In this talk, I plan to provide an overview of several ongoing efforts in our lab on the intersection between those fields, from (1) developing more biologically plausible alternatives to backpropagation, (2) tackling continual learning with adaptive, neurogenetic architectures and novel learning algorithms, (3) reward-driven attention in online decision making and more bio-inspired reward models in reinforcement learning, to (4) novel dialog generation approach for mental health (therapy), as well as (5) coupled nonlinear dynamical models for both brain activity modeling in neuroimaging and for improving recurrent neural nets performance in applications with limited training data.

Bio :

Irina Rish is a researcher at the AI Foundations department of the IBM T.J. Watson Research Center. She received MS in Applied Mathematics from Moscow Gubkin Institute, Russia, and PhD in Computer Science from the University of California, Irvine. Her areas of expertise include artificial intelligence and machine learning, with a particular focus on probabilistic graphical models, sparsity and compressed sensing, active learning, and their applications to various domains, ranging from diagnosis and performance management of distributed computer systems (“autonomic computing”) to predictive modeling and statistical biomarker discovery in neuroimaging and other biological data. Irina has published over 80 research papers, several book chapters, two edited books, and a monograph on Sparse Modeling, taught several tutorials and organized multiple workshops at machine-learning conferences, including NIPS, ICML and ECML. She holds over 60 patents and several IBM awards. Irina currently serves on the editorial board of the Artificial Intelligence Journal (AIJ). As an adjunct professor at the EE Department of Columbia University, she taught several advanced graduate courses on statistical learning and sparse signal modeling.

]]>
news-47446 Fri, 22 Mar 2019 10:30:00 -0400 Automatic Human Behavior Analysis and Recognition for Research and Clinical Use - Zakia Hammal https://diro.umontreal.ca/departement/colloques/colloque/news/detail/News/automatic-human-behavior-analysis-and-recognition-for-research-and-clinical-use-zakia-hammal/ Automatic Human Behavior Analysis and Recognition for Research and Clinical Use

par

Zakia Hammal

Senior Project Scientist, The Robotics Institute, Carnegie Mellon University

Vendredi 22 mars, 10:30-12:00Salle 3195, Pavillon André-Aisenstadt

    Université de Montréal, 2920 Chemin de la Tour

Résumé:

Nonverbal behavior is multimodal and interpersonal. In several studies, I addressed the dynamics of facial expression and head movement for emotion communication, social interaction, and clinical applications. By modeling multimodal and interpersonal communication my work seeks to inform affective computing and behavioral health informatics. In this talk, I will address some of my recent work that has addressed computational methods for affect communication in children with facial abnormalities, automatic measurement of pain intensity, and depression severity assessment. I will conclude my talk by sketching my efforts to pursue my research in using stateof-the art approaches from both AI and machine learning in building new multimodal artificial intelligence focused around understanding and modeling human emotion, physical, and cognitive states.

 

Biographie :

Zakia Hammal is a senior project scientist at the Robotics Institute at Carnegie Mellon University. Her training is in computer vision, machine learning, and signal and image processing. Through her early work, she developed leading approaches for automatic tracking of face motion and the detection of the occurrence and timing of facial expressions. Much of her recent work has addressed computational methods for automatic detection of pain and pain intensity, automatic assessment of treatment outcomes in psychiatric disorders, automatic measurement of behavioral markers in autism spectrum disorder, and modelling the dynamics of nonverbal behavior in social interaction. She has made extensive and sustained contributions to the social and organizational betterment of affective computing. She organized successful workshops in Interpersonal Synchrony and Influence (INTERPERSONAL at ICMI 2015), and in Face and Gesture Analysis for Health Informatics (FGAHI at CVPR 2019, FG 2018). To promote the critical importance of context in affect recognition, she leads a series of six successful Context-Based Affect Recognition workshops at premier IEEE conferences in computer vision, affective computing, social communication, and multimedia (CBAR at FG 2019, ACII 2017, CVPR 2016, FG 2015, ACII 2013, and SocialCom 2012). She served as Publication Chair of ACM ICMI 2014, Area Chair of IEEE FG 2017 and FG 2019, and served as mentor in several doctoral consortia (FG 2013, ACII 2015, FG 2018). Her honors include an outstanding paper award at ACM ICMI 2012, Best paper ward at IEEE ACII 2015, and Outstanding Reviewer Award at IEEE FG 2015.

Website: https://www.ri.cmu.edu/personal-pages/ZakiaHammal/.

 

 

]]>
news-42210 Thu, 21 Mar 2019 10:30:00 -0400 Agency in the Era of Learning Systems - Jakob Foerster https://diro.umontreal.ca/departement/colloques/colloque/news/detail/News/agency-in-the-era-of-learning-systems-jakob-foerster/ Agency in the Era of Learning Systems

par

Jakob Foerster

PhD in AI at the University of Oxford

Jeudi 21 mars, 10:30-12:00Salle 3195, Pavillon André-Aisenstadt

    Université de Montréal, 2920 Chemin de la Tour

Résumé:

We commonly think of machine learning problems, such as machine translation, as supervised tasks consisting of a static set of inputs and desired outputs. Even reinforcement learning, which tackles sequential decision making, typically treats the environment as a stationary black box.

However, as machine learning systems are deployed in the real world, these systems start having impact on each other and their users, turning their decision making into a multi-agent problem. It is time we start thinking of these problems as such, by directly accounting for the agency of other learning systems in the environment. In this talk we look at recent advances in the field of multi-agent learning, where accounting for agency can have drastic effects.

As a case study we present the “Bayesian Action Decoder” (BAD), which allows agents to directly reason over the beliefs of other agents in order to learn communication protocols in settings with limited public knowledge and actions that can be used to share information. BAD can be seen as a step towards a kind of “theory of mind” for AI agents and achieves a new state-of-the-art on the cooperative, partial-information, card-game Hanabi (“Spiel des Jahres” in 2013), an exciting new benchmark for measuring AI progress.
 

Biographie :

Jakob Foerster recently obtained his PhD in AI at the University of Oxford, under the supervision of Shimon Whiteson. Using deep reinforcement learning (RL) he studies how accounting for agency can address multi-agent problems, ranging from the emergence of communication to non-stationarity, reciprocity and multi-agent credit-assignment. His papers have gained prestigious awards at top machine learning conferences (ICML, AAAI) and have helped push deep multi-agent RL to the forefront of AI research. During his PhD Jakob interned at Google Brain, OpenAI, and DeepMind. Prior to his PhD Jakob obtained a first-class honours Bachelor’s and Master’s degree in Physics from the University of Cambridge and also spent four years working at Goldman Sachs and Google. Previously he has also worked on a number of research projects in systems neuroscience, including work at MIT and research at the Weizmann Institute.

Website: www.jakobfoerster.com.

 

 

]]>
news-47445 Tue, 19 Mar 2019 10:30:00 -0400 Apprentissage d'actions temporellement abstraites - Pierre-Luc Bacon https://diro.umontreal.ca/departement/colloques/colloque/news/detail/News/apprentissage-dactions-temporellement-abstraites-pierre-luc-bacon/ Apprentissage d'actions temporellement abstraites

Par

Pierre-Luc Bacon

Stanford AI for Human Impact Lab

 

Mardi 19 mars, 10:30-12:00Salle 3195, Pavillon André-Aisenstadt

    Université de Montréal, 2920 Chemin de la Tour

Résumé:

Comment peut-on arriver à planifier et prendre des décisions complexes ayant des conséquences à long terme ? Cette question se pose depuis les débuts de l'intelligence artificielle : des premiers exploits d'Arthur Samuel en IA dans le jeu de dames, en passant par les algorithmes de planification classiques des années 70, jusqu'aux progrès les plus récents en apprentissage par renforcement. Une solution efficace trouvée à ce problème consiste à éliminer les détails du « comment » par un processus d'abstraction temporel. Bien que nous sachions comment planifier et apprendre à partir d'actions temporellement abstraites données, la question de comment découvrir ces abstractions automatiquement s'est avérée plus difficile. Dans cette présentation, je vais développer les idées maîtresses derrière l'architecture « option-critic » (Bacon et al., 2017) ayant permis une des premières percée sur ce problème en apprentissage par renforcement. Je vais ensuite expliquer comment la notion de « rationalité limitée » de Simon (1957) peut nous aider régulariser les solutions apprises par notre approche. Cette perspective sera mise en correspondance avec le problème de construction de bons préconditionneurs de matrices en algèbre linéaire par la notion de « matrix splitting » de Varga (1962). Je vais finalement conclure avec les plans d'une nouvelle approche permettant de planifier dans un continuum de « buts » à différentes portées dans le temps : un problème d'optimisation à deux niveaux avec un point fixe au plus bas niveau.

 

Biographie :

Pierre-Luc Bacon a obtenu son doctorat en science informatique en 2018 sous la supervision de Doina Precup à l'Université McGill. Il est actuellement chercheur postdoctoral dans le « Stanford AI for Human Impact Lab » sous la direction d'Emma Brunskill. Ses efforts de recherche en apprentissage par renforcement se concentrent autour du problème d'apprentissage sur de longues portées dans le temps basé sur le cadre théorique des actions temporellement abstraites de Sutton et al. (1999).

Website: http://pierrelucbacon.com.

 

 

]]>
news-42209 Fri, 15 Mar 2019 10:30:00 -0400 Towards Literate Artificial Intelligence https://diro.umontreal.ca/departement/colloques/colloque/news/detail/News/towards-literate-artificial-intelligence/ Towards Literate Artificial Intelligence

par

Mrinmaya Sachan

Ph.D. candidate in the Machine Learning Department in the School of Computer Science at Carnegie Mellon University

Vendredi 15 mars, 10:30-12:00Salle 3195, Pavillon André-Aisenstadt

    Université de Montréal, 2920 Chemin de la Tour

Résumé:

Over the past decade, the field of artificial intelligence (AI) has seen striking developments. Yet, today’s AI systems sorely lack the essence of human intelligence i.e. our ability to (a) understand language and grasp its meaning, (b) assimilate common-sense background knowledge of the world, and (c) draw inferences and perform reasoning. Before we even begin to build AI systems that possess the aforementioned human abilities, we must ask an even more fundamental question: How would we even evaluate an AI system on the aforementioned abilities? In this talk, I will argue that we can evaluate AI systems in the same way as we evaluate our children - by giving them standardized tests. Standardized tests are administered to students to measure the knowledge and skills gained by them. Thus, it is natural to use these tests to measure the intelligence of our AI systems. Then, I will describe Parsing to Programs (P2P), a framework that combines ideas from semantic parsing and probabilistic programming for situated question answering. We used P2P to build systems that can solve pre-university level Euclidean geometry and Newtonian physics examinations. P2P achieves a performance at least as well as the average student on questions from textbooks, geometry questions from previous SAT exams, and mechanics questions from Advanced Placement (AP) exams. I will conclude by describing implications of this research and some ideas for future work.

 

Biographie :

Mrinmaya Sachan is a Ph.D. candidate in the Machine Learning Department in the School of Computer Science at Carnegie Mellon University. His research is in the interface of machine learning, natural language processing, knowledge discovery and reasoning. He received an outstanding paper award at ACL 2015, multiple fellowships (IBM fellowship, Siebel scholarship and CMU CMLH fellowship) and was a finalist for the Facebook fellowship. Before graduate school, he graduated with a B.Tech. in Computer Science and Engineering from IIT Kanpur with an Academic Excellence Award.

Website: sites.google.com/site/mrinsachan.

 

 

]]>
news-42208 Wed, 13 Mar 2019 10:30:00 -0400 Kernel distances for distinguishing and sampling from probability distributions https://diro.umontreal.ca/departement/colloques/colloque/news/detail/News/kernel-distances-for-distinguishing-and-sampling-from-probability-distributions/ Kernel distances for distinguishing and sampling from probability distributions

par

Dougal Sutherland

Postdoctoral researcher at the Gatsby Computational Neuroscience Unit, University College London

Mercredi 13 mars, 10:30-12:00Salle 3195, Pavillon André-Aisenstadt

    Université de Montréal, 2920 Chemin de la Tour

Résumé:

Probability distributions are the core object of statistical machine learning, and one of the basic properties we can consider is distances between them. In this talk, we will consider using these distances for two important tasks, and show how to design distances which will be useful for each. First, we study the problem of two-sample testing, where we wish to determine whether (and how) two different datasets meaningfully differ. We then study this framework in the setting of training generative models, such as generative adversarial networks (GANs), which learn to sample from complex distributions such as those of natural images.

The distances used are defined in terms of kernels, but we parameterise these kernels as deep networks for flexibility. This combination gives both theoretical and practical benefits over staying purely in either framework, and we obtain state-of-the-art results for unsupervised image generation on CelebA and ImageNet with our novel Scaled MMD GAN.

 

Biographie :

Dougal Sutherland is a postdoctoral researcher at the Gatsby Computational Neuroscience Unit, University College London, working with Arthur Gretton. He received his PhD in 2016 from Carnegie Mellon University, advised by Jeff Schneider. His research focuses on problems of learning about distributions from samples, including training implicit generative models, density estimation, two-sample testing, and distribution regression. His work combines kernel frameworks with deep learning, and aims for theoretical grounding of practical results.

]]>
news-42207 Thu, 07 Mar 2019 10:30:00 -0500 Better understanding of modern paradigms in probabilistic models https://diro.umontreal.ca/departement/colloques/colloque/news/detail/News/better-understanding-of-modern-paradigms-in-probabilistic-models/ Better understanding of modern paradigms in probabilistic models

par

Andrej Risteski

Norbert Wiener Fellow at the Institute for Data Science and Statistics (IDSS) and an Instructor of Applied Mathematics at MIT

Jeudi 7 mars, 10:30-12:00Salle 3195, Pavillon André-Aisenstadt

    Université de Montréal, 2920 Chemin de la Tour

Résumé:

In recent years, one of the areas of machine learning that has seen the most exciting progress is unsupervised learning, namely learning in the absence of labels or annotation. An integral part of these advances have been complex probabilistic models for high-dimensional data, capturing different types of intricate latent structure. As a consequence, a lot of statistical and algorithmic issues have emerged, stemming from all major aspects of probabilistic models: representation (expressivity and interpretability of the model), learning (fitting a model from raw data) and inference (probabilistic queries and sampling from a known model).

A common theme is that the models that are used in practice are often intractable in the worst-case (either computationally or statistically), yet even simple algorithms are, to borrow from Wigner, unreasonably effective in practice. It thus behooves us to ask why this happens.

I will showcase some of my research addressing this question, in the context of (i) computationally efficient inference using Langevin dynamics in the presence of multimodality; (ii) statistical guarantees for Learning distributions using GANs (Generative Adversarial etworks); iii) explaining surprising properties of vector representations of words (word embeddings).

 

Biographie :

Andrej Risteski holds a joint position as the Norbert Wiener Fellow at the Institute for Data Science and Statistics (IDSS) and an Instructor of Applied Mathematics at MIT.

Before MIT, he was a PhD student in the Computer Science Department at Princeton University,
working under the advisement of Sanjeev Arora. Prior to that he received his B.S.E. degree at Princeton University as well.

His work lies in the intersection of machine learning and theoretical computer science. The broad goal of his research is theoretically understanding statistical and algorithmic phenomena and problems arising in modern machine learning.

Website: math.mit.edu/~risteski.

 

 

]]>
news-42206 Tue, 05 Mar 2019 10:30:00 -0500 Humans teaching machines and machines teaching humans https://diro.umontreal.ca/departement/colloques/colloque/news/detail/News/humans-teaching-machines-and-machines-teaching-humans/ Humans teaching machines and machines teaching humans

par

Oisin Mac Aodha

Postdoc at Caltech Computational Vision Lab

Mardi 5 mars, 10:30-12:00Salle 3195, Pavillon André-Aisenstadt

    Université de Montréal, 2920 Chemin de la Tour

Résumé:

Our current machine learning solutions are rigid (i.e. we collect, train, and deploy). In contrast, many real world problem domains are not structured in this way. We need flexible systems that enable each stakeholder (e.g. experts, annotators, and model consumers) to interact and iterate in order to efficiently reach consensus for the task at hand. This will result in the creation of living knowledge bases that are empowered by experts and available to all. To achieve this goal we need to build on ideas from active learning, machine teaching, representation learning, crowdsourcing, and interpretable machine learning.

In this talk I will discuss our recent attempts to develop methods that unite the complementary strengths of humans and machines. I will present work on the automatic teaching of visual concepts to human learners. Our proposed model provides automatically generated interpretable feedback to learners and models how they update their beliefs in light of this information. Through empirical evaluation, I will show that this results in a significant reduction in the time required to teach new concepts in varied domains such as species identification and medical diagnosis. Finally, I will also discuss the self-supervised learning of deep representations from raw unlabeled image data. I will show that rich representations that encode information about the shape and structure of the world can be extracted from image sequences without requiring any explicit supervision at training time. These types of representations offer the potential to act as a powerful initialization signal for other downstream tasks.

Biographie :

Oisin Mac Aodha is a postdoctoral scholar with Prof. Pietro Perona in the Computational Vision Lab at the California Institute of Technology (Caltech). He obtained his PhD from University College London (UCL) with Prof. Gabriel Brostow. He is a recipient of the Travelling Studentship in the Sciences from the National University of Ireland. His current research interests are broadly in the areas of machine learning, computer vision, and human-in-the-loop methods such as active learning and machine teaching. More information and a list of publications can be found on his website: www.oisin.info.

]]>
news-42205 Mon, 04 Mar 2019 10:30:00 -0500 Learning adaptive language interfaces through interaction https://diro.umontreal.ca/departement/colloques/colloque/news/detail/News/learning-adaptive-language-interfaces-through-interaction/ Learning adaptive language interfaces through interaction

par

Sida I. Wang

Research instructor at Princeton University and Institute for Advanced Study

Lundi 4 mars, 10:30-12:00Salle 3195, Pavillon André-Aisenstadt

    Université de Montréal, 2920 Chemin de la Tour

Résumé:

The interactivity and adaptivity of natural language have the potential to allow people to better communicate with increasingly AI-driven computer systems. However, current natural language interfaces are mostly static and fall short of their potential. In this talk, I will cover two systems that can quickly learn from interactions, adapt to users, and simultaneously give feedback so that users can adapt to the system. The first system learns from scratch from users in real time. The second starts with a programming language and then learns to naturalize the programming language by interacting with users. Finally, I will present how these ideas can be combined to build a natural language interface for data visualization and discuss my work on modeling interactive language learning more rigorously. 

Biographie :

Sida Wang is a research instructor at Princeton University and Institute for Advanced Study working in the areas of natural language processing and machine learning. He holds a Ph.D in computer science from Stanford University and a B.A.Sc. from the University of Toronto. He received an outstanding paper award at ACL 2016 and the NSERC Postgraduate Scholarship.

Website: www.sidaw.xyz.

 

 

]]>
news-42204 Fri, 01 Mar 2019 10:30:00 -0500 Bringing Computer Vision to Robotics https://diro.umontreal.ca/departement/colloques/colloque/news/detail/News/bringing-computer-vision-to-robotics/ Bringing Computer Vision to Robotics

par

Sajad Saeedi

Dyson Research Fellow at Imperial College London UK

Vendredi 1 mars, 10:30-12:00Salle 3195, Pavillon André-Aisenstadt

    Université de Montréal, 2920 Chemin de la Tour

Résumé:

Recent advances in AI and machine learning have revolutionized computer vision and robotics. Robots are gradually moving beyond carefully controlled manufacturing facilities into households, executing tasks such as vacuum cleaning and lawn mowing. To extend the capabilities of these systems, robots need to move beyond just reporting ‘what’ is ‘where’ in an image to having the spatial awareness, necessary to interact usefully with their environment. Therefore, there is an urgent need for novel robotic perception systems that can deal with many real-world constraints such as limited resources and uncertain environments. 

In this brief technical talk, several recent projects related to robotics and machine perception are presented. Recent developments such as autonomous quadrotor aircraft, multi-robot systems, and accelerated inference on focal-plane sensor-processor arrays are introduced. These developments have significant economic and scientific impacts on our society and will open up new possibilities for real-time and reliable utilization of AI and computer vision algorithms in robotic systems.

At the end of the talk, future research directions will be outlined. The main goal for future research will be developing reliable, high-speed, and low-power robotic visual perception systems that can be deployed in real-world applications. It has been hypothesized that while machine learning algorithms will give us the required reliability, data processing in the focal plane will help us to achieve the desired energy consumption and run-time speed limits. Reliable, fast, and low-power computation for scene understanding and spatial awareness will be of great interest not only to the robotics community, but also other fields, such as Internet of Things (IoT), privacy-aware devices, and networked-visual devices. These research directions will help entrepreneurs and academic researchers to identify new opportunities in machine learning and its application in robotics and computer vision.

 

Biographie :

Sajad Saeedi is a Dyson Research Fellow at Imperial College London UK, Department of Computing, and an Associate Fellow of Higher Education Academy. He received his PhD in Electrical and Computer Engineering from the University of New Brunswick, Canada. In a joint £5 million British Research Council project with the University of Manchester and the University of Edinburgh, he was actively involved in the design of the future architecture for robotic perception and accelerated inference. He has developed large-scale datasets, benchmarking frameworks, and high-speed/low-power visual odometry algorithms for visual perception. Additionally, in his collaborations with industry, he has developed several successful products including omnidirectional and stereoscopic vision systems. He is currently working on semantic perception, bringing deep learning advances to computer vision and robotic systems. His research interests span over the design of multi-agent systems, aerial/marine robotics, and machine learning and its applications in computer vision, robotics, and control systems.

Website: www.sajad-saeedi.ca.

 

 

]]>
news-42203 Fri, 22 Feb 2019 10:30:00 -0500 Learning Hierarchical Control Programs https://diro.umontreal.ca/departement/colloques/colloque/news/detail/News/learning-hierarchical-control-programs/ Learning Hierarchical Control Programs

par

Roy Fox

Postdoc at UC Berkeley AI Research

Vendredi 22 février, 10:30-12:00Salle 3195, Pavillon André-Aisenstadt

    Université de Montréal, 2920 Chemin de la Tour

Résumé:

Machine learning is, in a sense, data-driven programming: the execution of coded programs is
replaced by the evaluation of learned models. In this talk, I will propose leveraging decades of
hard-earned programming wisdom by imposing a hierarchical structure on deep-learnable models. I will introduce Parametrized Hierarchical Procedures (PHP), a neural network architecture with improved data efficiency, interpretability, and reusability — in analogy to similar benefits for human coders who adhere to the procedural programming paradigm. This line of work focuses on two application domains, robot control and algorithmic program synthesis, in which the control program interacts with physical environments and memory structures, respectively, through a sequence of API calls to readers (sensors) and writers (motors). We collect data demonstrating correct interaction, and use it to learn a program, represented as a PHP, that can perform the task.

I will present three learning methods for PHPs: (1) hierarchical behavior cloning, which can
utilize data consisting of complete execution traces of the intended program; (2) exact inference, which can discover program structure that is latent in the data; and (3) variational inference, which can learn a more expressive class of models. I will demonstrate the superior data efficiency and generalization of these methods in learning several robotic and algorithmic tasks. Finally, I will discuss the need and opportunity to reuse data and models between related tasks, and show how hierarchical models can enhance reusability. In current work, we scale this up to a cloud service for sharing data and models between large libraries of control procédures.

Biographie :

Roy Fox is a postdoc at UC Berkeley AI Research (BAIR), working with Ion Stoica in the RISELab and with Ken Goldberg in the AUTOLAB. His research interests include reinforcement learning, dynamical systems, information theory, and robotics. His current research focuses on data-driven discovery of hierarchical control structures in deep reinforcement and imitation learning of robotic tasks. Roy has a MSc in Computer Science with Moshe Tennenholtz at the Technion, and a PhD in Computer Science with Naftali Tishby at the Hebrew University. He was an exchange PhD student with Larry Abbott and Liam Paninski at Columbia University, and a research intern at Microsoft Research.

]]>
news-42202 Mon, 18 Feb 2019 10:30:00 -0500 Towards Embodied Visual Intelligence https://diro.umontreal.ca/departement/colloques/colloque/news/detail/News/towards-embodied-visual-intelligence/ Towards Embodied Visual Intelligence

par

Dinesh Jayaraman

Postdoc at UC Berkeley

Lundi 18 février, 10:30-12:00, Salle 3195, Pavillon André-Aisenstadt

Université de Montréal, 2920 Chemin de la Tour

Résumé :

What would it mean for a machine to see the world? Computer vision has recently made great progress on problems such as finding categories of objects and scenes, and poses of people in images. However, studying such tasks in isolated disembodied contexts, divorced from the physical source of their images, is insufficient to build intelligent visual agents. My research focuses on remarrying vision to action, by asking: how might vision benefit from the ability to act in the world, and vice versa? Could embodied visual agents teach themselves through interaction and experimentation? Are there actions they might perform to improve their visual perception? How might they construct visual plans to achieve long-term action goals? In my talk, I will set up the context for these questions, and cover some strands of my work addressing them, proposing approaches for self-supervised learning through proprioception, visual prediction for decomposing complex control tasks, and active perception. Finally, I will discuss my long-term vision and directions that I hope to work on in the next several years.

Biographie :

Dinesh Jayaraman is a postdoctoral scholar at UC Berkeley. He received his PhD from UT Austin (2017) and B. Tech from IIT Madras (2011). His research interests are broadly in computer vision, robotics, and machine learning. In the last few years, he has worked on visual prediction, active perception, self-supervised visual learning, visuo-tactile robotic manipulation, semantic visual attributes, and zero-shot categorization. He has received an ACCV Best Application Paper Award (2016), a Samsung PhD Fellowship (2016), a UT Austin Graduate Dean's Fellowship (2016), and a Microelectronics and Computer Development Fellowship Award (2011). He has published in and reviewed for conferences and journals in computer vision, machine learning, and robotics, received a CVPR Outstanding Reviewer Award (2016), is as an Area Chair for NeurIPS (2018 & 2019).

]]>
news-42201 Fri, 08 Feb 2019 10:30:00 -0500 Visual Question Answering and Beyond https://diro.umontreal.ca/departement/colloques/colloque/news/detail/News/visual-question-answering-and-beyond/ Visual Question Answering and Beyond

par

Aishwarya Agrawal

School of Interactive Computing at Georgia Tech

Vendredi 8 février, 10:30-12:00Salle 3195, Pavillon André-Aisenstadt

    Université de Montréal, 2920 Chemin de la Tour

Résumé:

In this talk, I will present our work on a multi-modal AI task called Visual Question Answering (VQA) -- given an image and a natural language question about the image (e.g., “What kind of store is this?”, “Is it safe to cross the street?”), the machine’s task is to automatically produce an accurate natural language answer (“bakery”, “yes”). Applications of VQA include -- aiding visually impaired users in understanding their surroundings, aiding analysts in examining large quantities of surveillance data, teaching children through interactive demos, interacting with personal AI assistants, and making visual social media content more accessible.

Specifically, I will provide a brief overview of the VQA task, dataset and baseline models,
highlight some of the problems with existing VQA models, and talk about how to fix some of these problems by proposing -- 1) a new evaluation protocol, 2) a new model architecture, and 3) a novel objective function.

Most of my past work has been towards building agents that can ‘see’ and ‘talk’. However, for a lot of practical applications (e.g., physical agents navigating inside our houses executing natural
language commands) we need agents that can not only ‘see’ and ‘talk’ but can also take actions.
Towards the end of the talk, I will present future directions towards generalizing vision and
language agents to be able to take actions.

Biographie:

Aishwarya Agrawal is a fifth year Ph.D. student in the School of Interactive Computing at Georgia Tech, working with Dhruv Batra and Devi Parikh. Her research interests lie at the intersection of computer vision, deep learning and natural language processing. The Visual Question Answering (VQA) work by Aishwarya and her colleagues has witnessed tremendous interest in a short period of time.

Aishwarya is a recipient of the Facebook Fellowship 2019-2020 (declined) and NVIDIA Graduate
Fellowship 2018-2019. Aishwarya was selected for the Rising Stars in EECS 2018. She was also a finalist of the Foley Scholars Award 2018 and Microsoft and Adobe Research Fellowships 2017-2018. As a research intern Aishwarya has spent time at Google DeepMind, Microsoft Research and Allen Institute for Artificial Intelligence.

Aishwarya led the organization of the first VQA challenge and workshop at CVPR 2016 and
co-organized the second and the third VQA challenges and workshops at CVPR 2017 and CVPR 2018. As a reviewer, she has served on the program committee of various conferences 
(CVPR, ICCV, ECCV, NIPS, ICLR) and a journal (IJCV). She was awarded an Outstanding Reviewer award twice (NIPS 2017 and CVPR 2017).

Aishwarya received her bachelor's degree in Electrical Engineering with a minor in Computer Science and Engineering from Indian Institute of Technology (IIT) Gandhinagar in 2014.

]]>
news-42200 Tue, 05 Feb 2019 10:30:00 -0500 Olivier Lichtarge : Making Personal Sense of Disease: Machine Learning and a New Calculus of Fitness https://diro.umontreal.ca/departement/colloques/colloque/news/detail/News/olivier-lichtarge-making-personal-sense-of-disease-machine-learning-and-a-new-calculus-of-fitness/  Making Personal Sense of Disease: Machine Learning and a New Calculus of Fitness

par

Olivier Lichtarge, MD, PhD

 Computational and Integrative Biomedical Research Center, Baylor College of Medicine, Houston, TX

Mardi 5 février, 10:30-12:00Salle 3195, Pavillon André-Aisenstadt

    Université de Montréal, 2920 Chemin de la Tour

 

Résumé:

The relationship between genotype and phenotype shapes evolution in the long run, and human health every day. Although a complete solution may seem intractable, we will show two complementary approaches that provide insights into disease mechanisms and drug targets. First, we apply machine learning to networks of data and text. In p53 biology, malaria, and for drug discovery, these networks connect genes to chemicals and to diseases in ways that are novel, predictive, and which illustrate how data integration may yield automated discovery of new therapeutic paths. For precision medicine, however, we must also understand the role of individual genome variants in disease. For this we show a second approach: a new calculus of fitness landscapes that measures the co-evolution of DNA and fitness. A differential part quantifies the impact of mutations on proteins, patients and populations, including morbidity and mortality. An integral part solves which genes drive a phenotype, such as, antibiotic resistance in bacteria, tumors in cancers, and cognition in autism or in Alzheimer’s disease. Much work remains, but, so far, both approaches appear to be general, unbiased, scalable and complementary. Together they point to new disease genes and potentially will help guide clinical decisions tailored to each patient.

Biographie:

Olivier Lichtarge is Director of the Computational and Integrative Biomedical Research Center at Baylor College of Medicine, where he holds the Cullen Chair as a Professor in the Department of Molecular and Human Genetics. His computational Laboratory is at the interface between bioinformatics, machine learning and evolutionary theory and works on applications from bacterial to cancer biology. An early contribution was the Evolutionary Trace (ET) method to predict protein functional sites and associated studies of the allosteric mechanism of signal transduction in G protein coupled receptors. Building on this work, his laboratory developed more recently a formalism describing the impact of individual coding mutations on protein function. This Evolutionary Action (EA) theory formulates a general and computable differential equation for the co-evolution of genotype and phenotype. This equation is notable for consistently performing well against the most advanced machine learning methods at the blinded CAGI challenges, and recovering the distribution of fitness effect predicted by Fisher, in 1930. In Mendelian diseases, and some cancers, the EA equation is also predictive of morbidity and mortality. Other contributions are analyses of heterogeneous networks and of the literature by text mining leading to the automated generation of hypotheses. He was trained in Mathematics and Physics at McGill (B.Sc. first class joint honors), in Biophysics and Medicine at Stanford (M.D., Ph.D., with Oleg Jardetzky and Bruce Buchanan), and in Internal Medicine and Endocrinology at UCSF, where he also completed a postdoc in Pharmacology (with Fred Cohen and Henry Bourne). He learned most, however, from his spouse of 30 years, Karen Urbani M.D., and their three children.

]]>
news-42199 Mon, 21 Jan 2019 10:30:00 -0500 Marco Bijvank : Asymptotic Optimality of Order-Up-To Replenishment Policies for Serial Inventory Systems with Lost Sales https://diro.umontreal.ca/departement/colloques/colloque/news/detail/News/marco-bijvank-asymptotic-optimality-of-order-up-to-replenishment-policies-for-serial-inventory-systems-with-lost-sales/ Asymptotic Optimality of Order-Up-To Replenishment Policies for Serial Inventory Systems with Lost Sales

par

Marco Bijvank

 .....

Lundi 21 janvier, 10:30-11:30Salle 4488, Pavillon André-Aisenstadt

    Université de Montréal, 2920 Chemin de la Tour

 

Résumé:


We study the optimal policy for a serial inventory system under periodic review when excess demand at the retailer (i.e., the most downstream stage) is lost. When excess demand is backordered, the optimal policy is a base-stock policy with base-stock levels calculated using the algorithm of Clark and Scarf (1960). From the literature it is know that base-stock policies are asymptotically optimal for single-echelon inventory systems with lost sales as the penalty cost for lost demand grows large. In this paper, we extend this result to serial inventory systems. First, we show that an (S–1,S) inventory system is asymptotically optimal. We also show that this result is robust in the following sense: There is a large family of choices for the base-stock levels such that the asymptotic optimality continues to hold. Next, we generalize this result for (R,nQ) replenishment policies. In these policies, an order is placed when the inventory position is below a re-order level R and the order size is an integer multiple of Q units such that the inventory position after ordering equals or exceeds R. When Q=1, this equals the (S–1,S) system. Our theoretical results open up two interesting questions which we also study: (a) How cost-effective are the best echelon base-stock policy at moderate service levels (that is, 75% to 99%)? (b) Given that there is a large family of asymptotically optimal echelon base-stock policies, how can we pick one which offers good performance across a wide range of problem parameters?

]]>
news-38540 Fri, 30 Nov 2018 10:00:00 -0500 Sevag Gharibian : Efficient algorithms for quantum constraint satisfaction problems https://diro.umontreal.ca/departement/colloques/colloque/news/detail/News/sevag-gharibian-efficient-algorithms-for-quantum-constraint-satisfaction-problems/ Quantum Machine Learning: from nearest neighbor classifiers to quantum neural networks

par

Nathan Wiebe

quantum architectures and computing group at Microsoft in Redmond

Mercredi 28 novembre, 10:30-12:00, Salle 3195, Pavillon André-Aisenstadt

    Université de Montréal, 2920 Chemin de la Tour

La conférence sera présentée en anglais.

Résumé :

In recent years quantum computing has captivated people’s imaginations by promising speedups for computational tasks.  In recent years the promise of applying these methods to machine learning has led to a flurry of increasingly sophisticated results that show how quantum computers can be used to solve problems in machine learning and artificial intelligence faster (or better) than their classical brethren can.  I will provide a summary of some of the most important recent developments in this space that myself and collaborators have provided including quantum nearest neighbor classifiers, quantum perceptrons and algorithms for training variational autoencoders.  Finally, I will discuss a recent approach that collaborators and I have produced that shows how to use quantum Boltzmann machines to efficiently model quantum states and I will show that these quantum neural networks cannot be simulated by classical computers under reasonable complexity theoretic conjectures.  This is important because it addresses the longstanding conjecture that quantum neural networks can have greater expressive power than classical neural networks.

Biographie :

Nathan Wiebe is a researcher in quantum computing who focuses on quantum methods for machine learning and simulation of physical systems.   His work has provided the first quantum algorithms for deep learning, least squares fitting, quantum simulations using linear-combinations of unitaries, Hamiltonian learning, efficient Bayesian phase estimation and also has pioneered the use of particle filters for characterizing quantum devices as well as many other contributions ranging from the foundations of thermodynamics to adiabatic quantum computing and quantum chemistry simulation.   He received his PhD in 2011 from the university of Calgary studying quantum computing before accepting a post-doctoral fellowship at the University of waterloo and then finally joining Microsoft Research in 2013.  He is currently a researcher in the Microsoft Research Station Q, Quantum Architectures and Computing Group (QuArC) in Redmond WA.

]]>
news-38542 Wed, 28 Nov 2018 10:30:00 -0500 Nathan Wiebe : Quantum Machine Learning: from nearest neighbor classifiers to quantum neural networks https://diro.umontreal.ca/departement/colloques/colloque/news/detail/News/nathan-wiebe-quantum-machine-learning-from-nearest-neighbor-classifiers-to-quantum-neural-networks/  Quantum Machine Learning: from nearest neighbor classifiers to quantum neural networks

par

Nathan Wiebe

quantum architectures and computing group at Microsoft in Redmond

Mercredi 28 novembre, 10:30-12:00, Salle 3195, Pavillon André-Aisenstadt

    Université de Montréal, 2920 Chemin de la Tour

 

La conférence sera présentée en anglais.

Résumé :

In recent years quantum computing has captivated people’s imaginations by promising speedups for computational tasks.  In recent years the promise of applying these methods to machine learning has led to a flurry of increasingly sophisticated results that show how quantum computers can be used to solve problems in machine learning and artificial intelligence faster (or better) than their classical brethren can.  I will provide a summary of some of the most important recent developments in this space that myself and collaborators have provided including quantum nearest neighbor classifiers, quantum perceptrons and algorithms for training variational autoencoders.  Finally, I will discuss a recent approach that collaborators and I have produced that shows how to use quantum Boltzmann machines to efficiently model quantum states and I will show that these quantum neural networks cannot be simulated by classical computers under reasonable complexity theoretic conjectures.  This is important because it addresses the longstanding conjecture that quantum neural networks can have greater expressive power than classical neural networks.

Biographie :

Nathan Wiebe is a researcher in quantum computing who focuses on quantum methods for machine learning and simulation of physical systems.   His work has provided the first quantum algorithms for deep learning, least squares fitting, quantum simulations using linear-combinations of unitaries, Hamiltonian learning, efficient Bayesian phase estimation and also has pioneered the use of particle filters for characterizing quantum devices as well as many other contributions ranging from the foundations of thermodynamics to adiabatic quantum computing and quantum chemistry simulation.   He received his PhD in 2011 from the university of Calgary studying quantum computing before accepting a post-doctoral fellowship at the University of waterloo and then finally joining Microsoft Research in 2013.  He is currently a researcher in the Microsoft Research Station Q, Quantum Architectures and Computing Group (QuArC) in Redmond WA.

]]>
news-38543 Mon, 26 Nov 2018 10:30:00 -0500 Frédéric Dupuis : Informatique quantique: théorie de l'information et cryptographie https://diro.umontreal.ca/departement/colloques/colloque/news/detail/News/frederic-dupuis-informatique-quantique-theorie-de-linformation-et-cryptographie/ Informatique quantique: théorie de l'information et cryptographie

par

Frédéric Dupuis

CNRS, LORIA, Nancy, France

Lundi 26 novembre, 10:30-12:00, Salle 3195, Pavillon André-Aisenstadt

Université de Montréal, 2920 Chemin de la Tour

Résumé :

La théorie de l'information quantique est un domaine interdisciplinaire en plein essor, qui étudie le comportement de l'information lorsqu'elle est encodée dans un système régi par les lois de la mécanique quantique.  Plusieurs phénomènes surprenants font alors leur apparition: par exemple, il devient impossible, même en principe, de déterminer avec certitude l'état d'un système inconnu, et ces phénomènes peuvent être exploités pour accomplir des tâches impossibles dans le monde classique. Dans cet exposé, je présenterai un bref survol de ce domaine de recherche ainsi que certains des problèmes auxquels je me suis attaqué dans mes propres travaux, notamment en théorie de l'information et en cryptographie.

Biographie :

Frédéric Dupuis a soutenu sa thèse en 2009 au DIRO à l'Université de Montréal. Il fut ensuite chercheur postdoctoral en Suisse et au Danemark, avant de devenir professeur adjoint à l'Université Masaryk à Brno, République tchèque. Il travaille actuellement au LORIA, à Nancy, depuis son recrutement au CNRS en octobre 2017.

]]>
news-38515 Thu, 22 Nov 2018 15:00:00 -0500 Supartha Podder : The Garden-Hose Model https://diro.umontreal.ca/departement/colloques/colloque/news/detail/News/supartha-podder-the-garden-hose-model-1/  The Garden-Hose Model

par

Supartha Podder

Université d'Ottawa

Jeudi 22 novembre, 15:30-16:30, Salle 3195, Pavillon André-Aisenstadt

    Université de Montréal, 2920 Chemin de la Tour

Café avant 15:00-15:30

 La conférence sera présentée en anglais.

 

Résumé:

In 2011 Harry Buhrman, Serge Fehr, Christian Schaffner and Florian Speelman proposed a new measure of complexity for finite Boolean functions, called "The Garden-hose complexity". This measure can be viewed as a type of distributed space complexity where two players with private inputs compute a Boolean function co-operatively. While its motivation is mainly in applications to position based quantum cryptography, the playful definition of the model is quite appealing in itself.

Recently there has been some work proving non-trivial upper bounds for functions like Equality, Majority, etc., in this model and establishing new connections of this model with well studied models like communication complexity, permutation branching programs, and formula size.

In this talk we will discuss these results and look at potential directions for future research.

]]>
news-38545 Thu, 22 Nov 2018 15:00:00 -0500 Supartha Podder : The Garden-Hose Model https://diro.umontreal.ca/departement/colloques/colloque/news/detail/News/supartha-podder-the-garden-hose-model/ The Garden-Hose Model

par

Supartha Podder

Université d'Ottawa

Jeudi 22 novembre, 15:30-16:30, Salle 3195, Pavillon André-Aisenstadt

Université de Montréal, 2920 Chemin de la Tour

Café avant 15:00-15:30

La conférence sera présentée en anglais.

Résumé:

In 2011 Harry Buhrman, Serge Fehr, Christian Schaffner and Florian Speelman proposed a new measure of complexity for finite Boolean functions, called "The Garden-hose complexity". This measure can be viewed as a type of distributed space complexity where two players with private inputs compute a Boolean function co-operatively. While its motivation is mainly in applications to position based quantum cryptography, the playful definition of the model is quite appealing in itself.
Recently there has been some work proving non-trivial upper bounds for functions like Equality, Majority, etc., in this model and establishing new connections of this model with well studied models like communication complexity, permutation branching programs, and formula size.
In this talk we will discuss these results and look at potential directions for future research.

]]>