Lisa_seminaires

lisa_seminaires@iro.umontreal.ca

396 discussions

Fwd: Conférencier émérite et récipiendaire de la médaille du 50e du DIRO, Yann LeCun
by Yoshua Bengio 16 Sep '16

16 Sep '16

Yann LeCun is coming here next Wednesday and will give a talk at 3:15pm, see below. Probably in French though. A good exercise for those who are unsure about their French comprehension. He will receive some kind of prize, too. ---------- Message transféré ---------- De : Houari Sahraoui <sahraouh(a)iro.umontreal.ca> Date : 16 septembre 2016 à 15:27 Objet : Conférencier émérite et récipiendaire de la médaille du 50e du DIRO, Yann LeCun À : seminaires(a)iro.umontreal.ca Bonjour, Vous êtes conviés à la prochaine conférence de la série "Les conférenciers émérites du 50e du DIRO". http://diro.umontreal.ca/departement/conferenciers-emerites/ Yann LeCun (Facebook AI Research et New York Univesity) Date : Mercredi 21 septembre, 15 h 15 Lieu : Salle S1-151 du pavillon Jean-Coutu Titre : L'apprentissage prédictif et le futur de l'intelligence artificielle La conférence sera suivie par la remise de la médaille du 50e anniversaire du département d'informatique et de recherche opérationnelle Résumé : Les progrès fulgurants de l'intelligence artificielle ces dernières années sont en grande partie dus à la réémergence de l'apprentissage profond et des réseaux neuronaux, conjuguée à la disponibilité de grandes quantités de données et d'ordinateurs de plus en plus rapides. Nous disposons maintenant de systèmes de reconnaissance d'images dont la précision rivalise celles des humains. Des révolutions dans des domaines tels que les voitures autonomes et l'analyse d'images médicales sont pressenties. Mais tous ces succès utilisent l'apprentissage supervisé dans lequel la machine est entraînée avec des données étiquetées par des humains. Le défi au cours des prochaines années sera de permettre aux machines d'apprendre à partir de données brutes non étiquetées, telles que des vidéos ou du texte. C'est ce qu'on appelle l'apprentissage prédictif ou non supervisé. Ce qui manque aux machines d’aujourd'hui, c'est le « sens commun », qui est enraciné dans notre compréhension des contraintes du monde physique. L'apprentissage non supervisé est vu par certains comme la clé de l'acquisition du sens commun par les machines. Biographie : Yann LeCun est directeur de Facebook AI Research (FAIR) et professeur à l’Université de New York (NYU), où il est affilié au Courant Institute of Mathematical Sciences, de même qu’au Center for Data Science. Diplômé de l’ESIEE Paris, école supérieure d’ingénieurs, en électronique et électrotechnique et de l’Université Pierre et Marie Curie, il devient chercheur aux Laboratoires Bell en 1988. M. LeCun est nommé directeur de département au AT&T Labs-Research en 1996 puis rejoint NYU en 2003 et Facebook à la fin de 2013. Ses recherches portent principalement sur l'intelligence artificielle, l'apprentissage machine, la vision artificielle et la robotique. Il est l’un des chefs de file de l’apprentissage profond (deep learning). Depuis quelques années, ce champ d’études a révolutionné l'intelligence artificielle, notamment la reconnaissance visuelle et vocale et le traitement de la langue naturelle. L’une des techniques développées par M. LeCun, le réseau convolutif, est utilisée par toutes les grandes sociétés pour la reconnaissance d'images et de la parole et pour l'assistance à la conduite des voitures. Yann LeCun a publié près de 200 articles et a reçu le Neural Network Pioneer Award ainsi que le PAMI Distinguished Researcher Award de l'IEEE. Pour l'année 2015-2016, il a été nommé titulaire de la chaire annuelle « Informatique et sciences numériques » au Collège de France.

1 0

Predoc presentation of Saizheng Zhang
by Aaron Courville 12 Sep '16

12 Sep '16

Greeting, Today (Monday Sept. 12th) at 1pm in AA3195 will be Saizheng Zhang predoc presentation. Please come to support a fellow lab member. Title: Understanding and Improving Recurrent Neural Networks Abstract: This predoctoral presentation will mainly focus on understanding and exploring the properties of recurrent neural networks (RNNs) which are proved to have strong capacities for modeling highly nonlinear dynamics in sequential data. I will introduce two of my finished works on RNNs: In the first work, I will talk about measuring the architectural complexity of the general RNN structure, which is strongly related to RNN's connecting topology. In the second work, I will talk about a general structural design which is applicable for any popular recurrent structures (vanilla RNN, LSTM, GRU, etc.). This structural design is extremely easy to implement and can bring consistent performance improvement on a wide range of sequential tasks. Jury: Président-rapporteur: Aaron Courville Membre: Jian-Yun Nie Directeur de Recherche: Yoshua Bengio -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com

1 0

Fwd: CFP: AISTATS 2017
by Yoshua Bengio 22 Aoû '16

22 Aoû '16

---------- Forwarded message ---------- From: Aaditya Ramdas <aramdas(a)berkeley.edu> Date: 2016-08-21 21:23 GMT-04:00 Subject: CFP: AISTATS 2017 To: Csaba Szepesvari <szepesva(a)cs.ualberta.ca>, Mark Schmidt < schmidtm(a)cs.ubc.ca>, yoshua.bengio(a)umontreal.ca Hi Csaba, Mark, Yoshua, Hope things are going well in Canada. I'm the publicity chair for AISTATS next year, with Jerry Zhu and Aarti Singh being the program chairs. We've tried to be a bit more progressive in bringing together the Stat and CS/ML/AI communities this year. Could you please help disseminate the call for papers amongst Stats, EE/CS, math, b-schools, etc in your schools and other relevant departments that do ML/Stats/AI/optimization? Thanks! AISTATS <http://aistats.org/> is an interdisciplinary gathering of researchers at the intersection of artificial intelligence, machine learning, statistics, and related areas. The 20th International Conference on Artificial Intelligence and Statistics (AISTATS <http://www.aistats.org/>) will take place in Fort Lauderdale, Florida, USA from *April 20-22, 2017*. The deadline for paper submission is *Oct 13, 2016*, with final decisions made on Jan 24, 2017. New this year: 1. *Fast-track for Electronic Journal of Statistics*: Authors of a small number of accepted papers will be invited to submit an extended version for fast-track publication in a special issue of the Electronic Journal of Statistics (EJS) after the AISTATS decisions are out. Details on how to prepare such extended journal paper submission will be announced after the AISTATS decisions. 2. *Review-sharing with NIPS*: Papers previously submitted to NIPS 2016 are required to declare their previous NIPS paper ID, and optionally supply a one-page letter of revision (similar to a revision letter to journal editors; anonymized) in supplemental materials. AISTATS reviewers will have access to the previous anonymous NIPS reviews. Other than this, all submissions will be treated equally. *Paper Submission:* Electronic submission of PDF papers is required. The main part of the paper (single PDF up to 5Mb) may be up to 8 double-column pages in length including tables/figures. References only can exceed the 8 page limit. The main part should have enough information so that reviewers are able to judge the correctness and merit of the paper. Authors may optionally submit supplementary material (up to 10Mb) as a single zip file, containing additional proofs, audio, images, video, data or source code. Reviewing any supplementary material is up to the discretion of the reviewers. *Dual Submissions Policy:* Submissions that are identical (or substantially similar) to versions that have been previously published, or accepted for publication, or that have been submitted in parallel to other conferences or journals are not appropriate for AISTATS and violate our dual submission policy. Exceptions to this rule are the following: (a) it is acceptable to submit work that has been made available as a technical report or similar, e.g., on arXiv, without citing it (to preserve anonymity). (b) Submission is permitted for papers presented or to be presented at conferences or workshops without proceedings (e.g., ICML or NIPS workshops), or with only abstracts published. The dual-submission rules apply during the whole AISTATS review period until the authors have been notified about the decision on their paper. *Double-blind review:* Papers will be selected via a rigorous double-blind peer-review process (the reviewers will not know the identities of the authors, and vice versa). It will be up to the authors to ensure the proper anonymization of their paper and supplemental materials. Violation of the above rules may lead to rejection without review. One round of author rebuttal will occur with the initial reviews available to the authors. *Evaluation Criteria:* Submissions will be judged on the basis of technical quality, novelty, potential impact, and clarity. Typical papers often (but not always) consist of a mix of algorithmic, theoretical and experimental results, in varying proportions. Results will be judged on the degree to which they have been objectively established and/or their potential for scientific and technological impact. *Publication and presentation:* All accepted papers will be presented at the conference as posters, with a few selected for additional oral presentation. All accepted papers will be treated equally when published in the AISTATS Conference Proceedings (Journal of Machine Learning Research Workshop and Conference Proceedings series). At least one author of each accepted paper must register and attend AISTATS. A small number of accepted papers will be invited to submit an extended version for fast-track publication in a special issue of the Electronic Journal of Statistics (EJS) journal after the AISTATS decisions are out. *Topics:* Since its inception in 1985, the primary goal of AISTATS has been to promote the exchange of ideas from artificial intelligence, machine learning, and statistics. We encourage the submission of all papers in keeping of this objective. Solicited topics include, but are not limited to: - Supervised, unsupervised and semi-supervised learning, kernel and Bayesian methods - Stochastic processes, hypothesis testing, causality, time-series, nonparametrics, asymptotic theory - Graphical models and inference, manifold learning and embedding, network analysis, statistical analysis of deep learning - Sparse models and compressed sensing, information theory - Reinforcement learning, planning, control, multi-agent systems, logic and probability, relational learning - Learning theory, game theoretic learning, online learning, bandits, learning for mechanism design - Convex and non-convex optimization, discrete optimization, Bayesian optimization - Algorithms and architectures for high-performance computing - Applications in biology, cognition, computer vision, natural language, neuroscience, robotics, etc. - Topological data analysis, selective inference, experimental design, interactive learning, optimal teaching, and other emerging topics --- Aaditya Ramdas www.cs.berkeley.edu/~aramdas

1 0

Fwd: [NIPS 2016] CFP "Bayesian Deep Learning" Workshop
by Yoshua Bengio 20 Aoû '16

20 Aoû '16

---------- Forwarded message ---------- From: Yarin Gal <yg279(a)cam.ac.uk> Date: 2016-08-19 4:54 GMT-04:00 Subject: [NIPS 2016] CFP "Bayesian Deep Learning" Workshop To: Yarin Gal <yg279(a)cam.ac.uk> Dear all, Could I please ask you to circulate the CFP below to people you think this might be relevant to? Thanks so much, Yarin ************************************************************* Bayesian Deep Learning workshop, NIPS 2016 Date: December 10, 2016 Location: Centre Convencions Internacional Barcelona, Barcelona, Spain http://bayesiandeeplearning.org/ ************************************************************* *1. Call for papers* We invite researchers to submit work in any of the following areas: * Probabilistic deep models for classification and regression (such as extensions and application of Bayesian neural networks), * Generative deep models (such as variational autoencoders), * Incorporating explicit prior knowledge in deep learning (such as posterior regularisation with logic rules), * Approximate inference for Bayesian deep learning (such as variational Bayes / expectation propagation / etc. in Bayesian neural networks), * Scalable MCMC inference in Bayesian deep models, * Deep recognition models for variational inference (amortised inference), * Model uncertainty in deep learning, * Bayesian deep reinforcement learning, * Deep learning with small data, * Deep learning in Bayesian modelling, * Probabilistic semi-supervised learning techniques, * Active learning and Bayesian optimisation for experimental design, * Information theory in deep learning, * Applying non-parametric methods, one-shot learning, and Bayesian deep learning in general. A submission should take the form of an extended abstract (2 pages long) in PDF format using the NIPS style. Author names do not need to be anonymised and references may extend as far as needed beyond the 2 page upper limit. If research has previously appeared in a journal, workshop, or conference (including NIPS 2016 conference), the workshop submission should extend that previous work. Submissions will be accepted as contributed talks or poster presentations. Extended abstracts should be submitted by 1 November 2016; submission details will be updated online towards the submission deadline. Final versions will be posted on the workshop website (and are archival but do not constitute a proceedings). *Key Dates:* Extended abstract submission: **1 November 2016** Acceptance notification: 16 November 2016 Travel award notification: 16 November 2016 Final paper submission: 5 December 2016 The workshop is endorsed by the International Society for Bayesian Analysis (ISBA), which will also provide a Travel Award to a graduate student or a junior researcher. *2. Description* While deep learning has been revolutionary for machine learning, most modern deep learning models cannot represent their uncertainty nor take advantage of the well studied tools of probability theory. This has started to change following recent developments of tools and techniques combining Bayesian approaches with deep learning. The intersection of the two fields has received great interest from the community over the past few years, with the introduction of new deep learning models that take advantage of Bayesian techniques, as well as Bayesian models that incorporate deep learning elements. In fact, the use of Bayesian techniques in deep learning can be traced back to the 1990s', in seminal works by Radford Neal, David MacKay, and Dayan et al.. These gave us tools to reason about deep models confidence, and achieved state-of-the-art performance on many tasks. However earlier tools did not adapt when new needs arose (such as scalability to big data), and were consequently forgotten. Such ideas are now being revisited in light of new advances in the field, yielding many exciting new results. This workshop will study the advantages and disadvantages of such ideas, and will be a platform to host the recent flourish of ideas using Bayesian approaches in deep learning and using deep learning tools in Bayesian modelling. The program will include a mix of invited talks, contributed talks, and contributed posters. Also, the historic context of key developments in the field will be explained in an invited talk, followed by a tribute talk to David MacKay's work in the field. Future directions for the field will be debated in a panel discussion. *3. Organisers* Yarin Gal (University of Cambridge) Christos Louizos (University of Amsterdam) Zoubin Ghahramani (University of Cambridge) Kevin Murphy (Google) Max Welling (University of Amsterdam)

1 0

Microsoft visit + presentation on dialogue tomorrow by Julian Serban, 10h30am, 3195
by Yoshua Bengio 29 Jui '16

29 Jui '16

Hello, Tomorrow morning we have a visit by a top brass from microsoft, Executive Vice-President Qi Lu. (1) please be at the lab at 10am, we want to show our "strength in numbers" (they arrive at my office at 10am) (2) please attend a presentation that Julian Serban will give to our visitors, at 10h30 in rm 3195, on Building Dialogue Systems Using Neural Networks Thank you, -- Yoshua

1 0

Re: [Lisa_seminaires] Pr?doc oral de Guillaume Alain: lundi 27 juin
by Aaron Courville 27 Jui '16

27 Jui '16

Reminder: Today is Guillaume Alain's predoc oral, at 11h00 in AA3195 Titre: "Denoising autoencoders, importance sampling, and visualizing training with probes" FYI: I understand the talk will be in French. Président-rapporteur: Aaron Courville Membre: Laurent Charlin Directeur de Recherche: Yoshua Bengio -- Aaron 2016-06-24 9:21 GMT-04:00 Aaron Courville <aaron.courville(a)gmail.com>: > Bonjour, > vous êtes invité au prédoc oral de Guillaume Alain. > Lundi 27 juin à 11h00 au 3195. > > Titre: "Denoising autoencoders, importance sampling, and visualizing > training with probes" > > Président-rapporteur: Aaron Courville > Membre: Laurent Charlin > Directeur de Recherche: Yoshua Bengio > > -- Aaron > > -- > Aaron C. Courville > Département d’Informatique et > de recherche opérationnelle > Université de Montréal > email:Aaron.Courville@gmail.com > -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com

1 0

Fwd: a talk by Valentine Kabanets
by Yoshua Bengio 24 Jui '16

24 Jui '16

---------- Forwarded message ---------- From: Pierre McKenzie <mckenzie(a)iro.umontreal.ca> Date: 2016-06-24 9:42 GMT-04:00 Subject: a talk by Valentine Kabanets To: Yoshua Bengio <bengioy(a)iro.umontreal.ca>, Pascal Vincent < vincentp(a)iro.umontreal.ca>, christopher.pal(a)polymtl.ca, Aaron Courville < aaron.courville(a)umontreal.ca>, roland.memisevic(a)umontreal.ca, lcharlin(a)gmail.com Cc: Sylvie Hamel <hamelsyl(a)iro.umontreal.ca> Bonjour collègues du MILA, If you think some members of your lab might be interested, please don't hesitate to distribute the announcement. Regards, Pierre McKenzie ------------------------------ Tuesday June 28 @ 11:00 in room 3195, pavillon André-Aisenstadt of U de M Speaker: Valentine Kabanets (Simon Fraser) Title: Learning Algorithms from Natural Proofs Abstract: Based on Hastad's (1986) circuit lower bounds, Linial, Mansour, and Nisan (J ACM, 1993) gave a quasipolytime learning algorithm for AC^0 (constant-depth circuits with AND, OR, and NOT gates), in the PAC model over the uniform distribution. It was an open question to get a learning algorithm (of any kind) for the class of AC^0[p] circuits (constant-depth, with AND, OR, NOT, and mod-p gates for a prime p). Our main result is a quasipolytime learning algorithm for AC^0[p] in the PAC model over the uniform distribution with membership queries. This algorithm is an application of the general connection between natural properties (in the sense of Razborov and Rudich (JCSS, 1997)) and learning algorithms that we obtain. We show that a natural circuit lower bound against any (sufficiently powerful) circuit class yields a learning algorithm for the same circuit class. Then the known lower bounds against AC ^0[p] by Razborov (1987) and Smolensky (1987), shown to be natural by Razborov and Rudich, imply our main result. Joint work with Marco Carmosino (UCSD), Russell Impagliazzo (UCSD), and Antonina Kolokolova (MUN).

1 0

ML talk tomorrow by François Laviolette
by Yoshua Bengio 14 Jui '16

14 Jui '16

Speaker: François Laviolette Wednesday (15th June) at 11h00. Room: TBA *1.**- TITLE: **Sparsity, interpretability and sample compression in Machine Learning* *2.- ABSTRACT:* We will investigate the Machine Learning subject called the Fat Data paradigm, a setting where the dimension of the feature space is much bigger than the number of training examples. This is a situation often encountered in data arising from life science. In such situations, there is a real risk of overfitting, even if one makes use of highly regularized learning algorithms. To overcome the lack of examples and even achieve good generalization performances, one solution consists in including prior knowledge of the domain into the learning algorithm. Another approach is to look for predictors that will be interpretable by experts in the domain, who in turn will help to validate the prediction. The Set Covering Machine, introduced by Mario Marchand and John Shawe-Taylor more than a decade ago, is a learning algorithm that can be used to produce such interpretable predictors. We will present the original version of this algorithm together with a new one, specialized for genomic data. We will also show some results, based on the sample compression theory, that show why such a learning algorithm can alleviate overfitting, even in the fat data situation. *3.- Bio Sketch* François Laviolette is a full Professor at the department of Computer Science and Software Engineering of Laval University. He received his doctorate in graph theory at the University of Montreal in 1995. His thesis solved an old problem that had been studied among others by the mathematician Paul Erdos. For over 10 years, his main area of research has been Machine Learning. More specifically, he develops learning algorithms to solve new types of learning problems, including problems relating to genomics and proteomics. He is currently the director of the new Big Data Research Center at Laval University.

1 1

Predoc oral de Zhouhan Lin: vendredi 27 mai (Demain) à 10h00 au AA3195
by Aaron Courville 26 Mai '16

26 Mai '16

Qui: Zhouhan Lin, MILA / DIRO Quoi: Examen prédoctoral oral Quand: vendredi, 27 mai 2016 à 10h00 Où: local 3195, André-Aisenstadt Titre: Towards Hardware Friendly Machine Learning Résumé: Le domaine de l'intelligence artificielle (AI) a connu de vastes progrès ces dernières années. La communauté de l'apprentissage machine évolue rapidement, grâce aux continuelles améliorations du matériel informatique. Les GPUs et logiciels actuels (comme Theano, Torch et TensorFlow) rendent maintenant possible d’entraîner des modèles en quelques jours, alors que leur entraînement pouvait durer plusieurs mois dans le passé. Ce gain de temps permet non seulement d'ajuster le modèle sur de beaucoup plus grands ensembles de données, mais fait également évoluer la recherche beaucoup plus rapidement. Par conséquent, les réseaux de neurones sont devenus l'état de l'art dans une grande variété d'applications. Cependant, la demande pour du calcul toujours plus rapide ne se terminera jamais. L'entraînement des modèles utilisés dans la traduction automatique, la reconnaissance vocale, etc. peut encore durer plusieurs semaines. Je propose donc de faire des adaptations aux réseaux de neurones et de les rendre plus léger et plus rapide. Je vais vous faire une revue de la recherche dans ce domaine, vous présenter plusieurs problèmes importants qui nous empêchent d'obtenir une exécution plus rapide de ces réseaux de neurones, et vous proposer des solutions pour les résoudre. Abstract: The field of Artificial Intelligence (AI) has witnessed vast advances in recent years. Advances in hardware have made the community of machine learning evolve much faster than before. With GPUs and well-developed frameworks like Theano, Torch and TensorFlow, it is now possible to train models in days which would instead need several months in the past. That speedup not only makes it possible to fit the model on much larger datasets, but also makes research evolve much faster. As a result, neural networks have become the state-of-the-art model behind a wide variety of applications. However, the demand for more efficient computation never ends. Models used in machine translation, speech recognition, etc still take weeks to train. I propose to make adaptations to neural network algorithms and make them lighter and faster. I am going to talk about several major factors blocking us from obtaining faster neural network execution, review previous researches in those ends, and propose ways to solve them. Jury: Aaron Courville (président) Yoshua Bengio (directeur) Roland Memisevic (codirecteur) Alain Tapp (membre) Tout le monde est bienvenu! -- Aaron C. Courville Département d’Informatique et de recherche opérationnelle Université de Montréal email:Aaron.Courville@gmail.com

1 0

Talk Fri. 19th: Towards bridging the gap between deep learning and biology
by Jörg Bornschein 19 Fév '16

19 Fév '16

Hi everyone, this weeks tea talk will be a bit special -- Yoshua Bengio will talk about bridging the gap between deep learning and biology; and we will have a Dutch TV crew around to film how "science is done". I expect it will be very crowded -- so maybe come a few minutes earlier. -- Title: Towards bridging the gap between deep learning and biology Who: Yoshua Bengio When: 14:30 Where: AA 3195 We explore the following crucial question: how could brains potentially perform the kind of powerful credit assignment that allows hidden layers of a very deep network to be trained and that has been so successful with backprop in deep nets recently? Global reinforcement learning signals have too much variance (scaling with the number of neurons or synapses) to be credible from a machine learning point of view. Concerns have been raised about how something like back-propagation could be implemented in brains. We present several intriguing results all aimed at answering this question and possibly providing pieces of this puzzle. We start with an update rule that yields updates similar to STDP but that is anchored in quantities such as pre-synaptic and post-synaptic firing rates and temporal rates of change. We then show that if neurons are connected symmetrically (with feedback connections) and define an energy function, (a) their behaviour corresponds both to inference, i.e., going down the energy, and propagating error gradients, (b) after a prediction is made on a sensor and an actual value is observed, the early phases of inference in this network actually propagate prediction error gradients, and (c) using the above STDP-inspired rule yields a gradient descent step on prediction error for the fixed point of the recurrent network (d) contrary to previously believed for such fixed-point networks, it is not necessary to do a full relaxation in the positive phase (perturbation propagation does the backprop job). Finally, we discuss some of the open problems we are facing to move forward, such as avoiding the negative phase fixed point relaxation (just like we got rid of the positive phase one), avoiding the forced symmetry of synaptic weights, the question of learning the full joint distribution and not just a point prediction, doing unsupervised learning, and handling time. -- Best and looking forward to see you there j

3 4

Aller à la page :

2024

2023

2022

2021

2020

2019

2018

2017

2016

2015

2014

2013

2012

2011

2010

2009

2008

2007

Lisa_seminaires