- Lisa_seminaires - mailman.iro.umontreal.ca

Visite et séminaire de Renato De Mori le 5 décembre à McGill
by Yoshua Bengio 26 Nov '07

26 Nov '07

Bonjour, Pour ceux qui s'intéressent aux systèmes de dialogue et à la compréhension du langage parlé, ou voudraient revoir Renato De Mori, il sera de passage à Montréal la première semaine de décembre: Begin forwarded message:

1 0

2 interesting talks this week
by Hugo Larochelle 20 Nov '07

20 Nov '07

There isn't a MITACS seminar this week, however the following talks might be of interest to you. Here is the first one at UdeM (see below for the second one at McGill): BoostCluster: Boosting Clustering by Pairwise Constraints by Rong Jin, Michigan State University Location: Pavillon Andre-Aisenstadt (UdeM), room 1177 Time: November 23rd 2007, 14h00 Data clustering is an important task in many disciplines. A large number of studies have attempted to improve clustering by using the side information that is often encoded as pairwise constraints. However, these studies focus on designing special clustering algorithms that can effectively exploit the pairwise constraints. We present a boosting framework for data clustering, termed as BoostCluster, that is able to iteratively improve the accuracy of any given clustering algorithm by exploiting the pairwise constraints. The key challenge in designing a boosting framework for data clustering is how to influence an arbitrary clustering algorithm with the side information since clustering algorithms by definition are unsupervised. The proposed framework addresses this problem by dynamically generating new data representations at each iteration that are, on the one hand, adapted to the clustering results at previous iterations by the given algorithm, and on the other hand consistent with the given side information. Our empirical study shows that the proposed boosting framework is effective in improving the performance of a number of popular clustering algorithms (K-means, partitional SingleLink, spectral clustering), and its performance is comparable to the state-of-the-art algorithms for data clustering with side information. Also, Rong Jin is giving another talk at McGill on Wednesday 21st, 11h30, McConnell Building (McGill), room 103: Title: Batch Mode Active Learning Abstract: The goal of active learning is to select the most informative examples for manual labeling. Most of the previous studies in active learning have focused on selecting a {\em single} unlabeled example in each iteration. This could be inefficient since the classification model has to be retrained for every labeled example. In this paper, we present a framework for ``{\bf batch mode active learning}'' that applies the Fisher information matrix to select a number of informative examples simultaneously. The key computational challenge is how to efficiently identify the subset of unlabeled examples that can result in the largest reduction in the Fisher information. To resolve this challenge, we propose efficient algorithms that are based on bound optimization and the property of submodular functions. Our empirical studies with text categorization and medical image classification show that the proposed batch mode active learning algorithm is more effective than the state-of-the-art algorithms for active learning.

1 0

REMINDER: UdeM-McGill-MITACS machine learning seminar Fri Nov. 16, 11:30am, MC 437
by Hugo Larochelle 15 Nov '07

15 Nov '07

This week's seminar (see http://www.iro.umontreal.ca/article.php3? id_article=107&lang=en): Bayesian Reinforcement Learning by Mohammad Ghavamzadeh, Department of Computing Science University of Alberta Location: McConnell Building (McGill), room 437 Time: November 16th 2007, 11h30 Reinforcement learning is a class of learning problems in which an agent interacts with an unfamiliar, dynamic and stochastic environment, and whose goal is to optimize some measure of its long- term performance. Despite extensive research and numerous successes in a number of different domains, there remain several fundamental obstacles hindering the widespread application of reinforcement learning methodology to real-world problems. Recent advances have shown that Bayesian approach to reinforcement learning offers viable solutions to some of these major limitations, such as the lack of confidence intervals for performance predictions, the difficulty of appropriately reconciling exploration with exploitation, and the lack of a systematic method for encoding prior knowledge and for formulating domain assumptions. Policy gradient methods are reinforcement learning algorithms that adapt a parameterized policy by following a performance gradient estimate. This talk will present two Bayesian policy gradient algorithms. These algorithms use Gaussian processes to define prior distribution over the performance gradient, and obtain closed-form expressions for its posterior distribution, conditioned on the observed data. The posterior mean serves as the policy gradient estimate and is used to update the policy, while the posterior covariance allows us to gauge the reliability of the update. In the first algorithm, the basic observable unit, upon which learning and inference are based, is a complete trajectory, allowing the algorithm to handle non-Markovian systems. The second algorithm takes advantage of the Markov property of the system trajectories and uses individual state-action-reward transitions as its basic observable unit. This helps reduce variance in the gradient estimates and facilitates handling continuing problems. _______________________________________________ Lisa_seminaires mailing list Lisa_seminaires(a)mercure.iro.umontreal.ca https://webmail.iro.umontreal.ca/mailman/listinfo/lisa_seminaires

1 0

UdeM-McGill-MITACS machine learning seminar Fri Nov. 16, 11:30am, MC 437
by Hugo Larochelle 12 Nov '07

12 Nov '07

This week's seminar (see http://www.iro.umontreal.ca/article.php3? id_article=107&lang=en): Bayesian Reinforcement Learning by Mohammad Ghavamzadeh, Department of Computing Science University of Alberta Location: McConnell Building (McGill), room 437 Time: November 16th 2007, 11h30 Reinforcement learning is a class of learning problems in which an agent interacts with an unfamiliar, dynamic and stochastic environment, and whose goal is to optimize some measure of its long- term performance. Despite extensive research and numerous successes in a number of different domains, there remain several fundamental obstacles hindering the widespread application of reinforcement learning methodology to real-world problems. Recent advances have shown that Bayesian approach to reinforcement learning offers viable solutions to some of these major limitations, such as the lack of confidence intervals for performance predictions, the difficulty of appropriately reconciling exploration with exploitation, and the lack of a systematic method for encoding prior knowledge and for formulating domain assumptions. Policy gradient methods are reinforcement learning algorithms that adapt a parameterized policy by following a performance gradient estimate. This talk will present two Bayesian policy gradient algorithms. These algorithms use Gaussian processes to define prior distribution over the performance gradient, and obtain closed-form expressions for its posterior distribution, conditioned on the observed data. The posterior mean serves as the policy gradient estimate and is used to update the policy, while the posterior covariance allows us to gauge the reliability of the update. In the first algorithm, the basic observable unit, upon which learning and inference are based, is a complete trajectory, allowing the algorithm to handle non-Markovian systems. The second algorithm takes advantage of the Markov property of the system trajectories and uses individual state-action-reward transitions as its basic observable unit. This helps reduce variance in the gradient estimates and facilitates handling continuing problems.

1 0

Fwd: COLLOQUE CRM-ISM-GERAD DE STATISTIQUE: C. Crainiceanu (Le vendredi 16 novembre 2007)
by Yoshua Bengio 05 Nov '07

05 Nov '07

Pour ceux qui ne veulent pas m'entendre radoter à McGill le 16 novembre il y a ce talk à Concordia qui pourrait être bien: Begin forwarded message: > From: CRM <activites(a)CRM.UMontreal.CA> > Date: November 5, 2007 8:29:05 AM EST (CA) > To: lofficiel <lofficiel(a)CRM.UMontreal.CA> > Subject: COLLOQUE CRM-ISM-GERAD DE STATISTIQUE: C. Crainiceanu (Le > vendredi 16 novembre 2007) > > ****************************************************************** > COLLOQUE CRM-ISM-GERAD DE STATISTIQUE > ****************************************************************** > > CONFERENCIER(S) / SPEAKER(S) : > Ciprian Crainiceanu > (Johns Hopkins University) > > > TITRE / TITLE : > "Multilevel Functional Principal Component Analysis" > > > LIEU / PLACE : > Concordia, Library Building, 1400 Blvd. de Maisonneuve O. Salle / > Room LB-921-04 > > DATE : > Le vendredi 16 novembre 2007 / Friday, November 16, 2007 > > HEURE / TIME : > 15 h 30 / 3:30 p.m. > > RESUME / ABSTRACT : > Modern research data have become increasingly complex, raising non- > traditional modeling and inferential challenges. In particular, > advancements in technology and computation have made recording and > processing of functional data possible. Examples of functional data > are time series of electroencephalographic (EEG) activity, > anatomical shape, and functional MRI. The purpose of this talk is > to describe statistical models for feature extraction from single- > level (one or multiple functions per subject at one visit) and > clustered or longitudinal (one or multiple functions per subject at > multiple visits) functional data having a large number of subjects > and large within- and between-subject heterogeneity. We introduce > the framework and inferential tools for multilevel functional data > (MFD) obtained by recording of functional characteristics at > multiple visits. Though motivated by a novel experimental setting, > the proposed methodology is general, with potential broad > applicability to many high-throughput scientific studies. A > prototypical example of MFD is the Sleep Heart Health Study (SHHS), > which contains electroencephalographic (EEG) signals for each > subject at two visits. > > ****************************************************************** > > http://www.crm.umontreal.ca/cgi/Stats.csh

1 0

séminaire de Yoshua Bengio dans 2 semaines à McGill
by Yoshua Bengio 02 Nov '07

02 Nov '07

Bonjour, Je donne un séminaire pour une audience non-spécialisée (informatique et génie électrique) dans deux semaines à McGill, sur la question de l'intelligence artificielle via l'apprentissage machine et des problèmes difficiles d'optimisation qui sont impliqués là-dedans. L'annonce est là: http://www.cse.mcgill.ca/Fall07/YB.pdf Friday Nov 16 at 14:30 in McConnell Engineering Bldg Room 103 Learning Deep Architectures for AI Theoretical results in complexity theory of circuits and in non-parametric statistics strongly suggest that in order to learn the kind of complicated functions that can represent high-level abstractions (e.g. in vision, language, and other AI-level tasks), one may need "deep architectures", which are composed of multiple levels of non-linear operations, such as in neural nets with many hidden layers. Searching the parameter space of deep architectures appears to be a fundamentally difficult optimization task, so approximate numerical optimization schemes are called for. Learning algorithms such as those for Deep Belief Networks have recently been proposed which make a dent in this difficult optimization task, beating the state-of-the-art in certain areas. This talk discusses the motivations and principles regarding learning algorithms for deep architectures and in particular for those based on unsupervised learning such as Deep Belief Networks, using as building blocks single-layer models such as Restricted Boltzmann Machines. -- Yoshua Bengio

1 0

UdeM-McGill-MITACS machine learning seminar Fri Nov. 2, 11:30am, MC 437
by Hugo Larochelle 30 Oct '07

30 Oct '07

This week's seminar (see http://www.iro.umontreal.ca/article.php3? id_article=107&lang=en): Preliminary steps toward a probabilistic, decision theoretic model of dynamic scene understanding by Nando de Freitas Computer Science Department University of british Columbia Location: Mc Gill, McConnell engineering building room 437 Time: November 2nd 2007, 11:30am Abstract: It is widely agreed that efficient visual search requires the integration of target-driven top-down information and image-driven bottom-up information. Yet the problem of gaze planning - that is, selecting the next best gaze location given the current observations - remains largely unsolved. We propose a probabilistic system that models the gaze sequence as a finite-horizon Bayesian sequential decision process. Direct policy search is used to reason about the next best gaze locations. The system integrates bottom-up saliency information, top-down target knowledge and additional context information through principled Bayesian priors. This results in proposal gaze locations that depend not only the feature visual saliency, but also on prior knowledge and the spatial likelihood of locating the target. The system has been implemented using state-of-the-art object detectors and evaluated on a real-world dataset by comparing it to gaze sequences proposed by a pure bottom-up saliency-based process and to an object detection approach that analyzes the full image. The target-directed attention system is shown to result in higher object detection precision than both competitors, to attend to more relevant targets than the bottom-up attention system, and to require significantly less computation time than the exhaustive approach. This is joint work with Julia Vogel.

1 0

REMINDER: UdeM-McGill-MITACS machine learning seminar, PAA 1409 (UdeM), Fri Oct 19, 11:30am
by Hugo Larochelle 25 Oct '07

25 Oct '07

This week's seminar (see http://www.iro.umontreal.ca/article.php3? id_article=107&lang=en): Learning the 2-D Topology of Images by Pascal Lamblin, Département d’Informatique et Recherche Opérationnelle Université de Montréal Location: Pavillon André-Aisenstadt (UdeM), room 1409 Time: October 19th 2007, 11h30 We study the following question: is the two-dimensional structure of images a very strong prior or is it something that can be learned with a few examples of natural images? If someone gave us a learning task involving images for which the two-dimensional topology of pixels was not known, could we discover it automatically and exploit it? For example suppose that the pixels had been permuted in a fixed but unknown way, could we recover the relative two-dimensional location of pixels on images? The surprising result presented here is that not only the answer is yes but that about as few as a thousand images are enough to approximately recover the relative locations of about a thousand pixels. This is achieved using a manifold learning algorithm applied to pixels associated with a measure of distributional similarity between pixel intensities. We compare different topology-extraction approaches and show how having the two-dimensional topology can be exploited.

2 1

UdeM-McGill-MITACS machine learning seminar, PAA 1409 (UdeM), Fri Oct 19, 11:30am
by Hugo Larochelle 17 Oct '07

17 Oct '07

This week's seminar (see http://www.iro.umontreal.ca/article.php3?id_article=107&lang=en) Learning the 2-D Topology of Images by Pascal Lamblin, Département d’Informatique et Recherche Opérationnelle Université de Montréal Location: Pavillon André-Aisenstadt (UdeM), room 1409 Time: October 19th 2007, 11h30 We study the following question: is the two-dimensional structure of images a very strong prior or is it something that can be learned with a few examples of natural images? If someone gave us a learning task involving images for which the two-dimensional topology of pixels was not known, could we discover it automatically and exploit it? For example suppose that the pixels had been permuted in a fixed but unknown way, could we recover the relative two-dimensional location of pixels on images? The surprising result presented here is that not only the answer is yes but that about as few as a thousand images are enough to approximately recover the relative locations of about a thousand pixels. This is achieved using a manifold learning algorithm applied to pixels associated with a measure of distributional similarity between pixel intensities. We compare different topology-extraction approaches and show how having the two-dimensional topology can be exploited.

1 0

Fwd: [dgi-seminaires] séminaire GIGL
by Yoshua Bengio 15 Oct '07

15 Oct '07

Begin forwarded message: > From: Gilles Pesant <Gilles.Pesant(a)polymtl.ca> > Date: October 15, 2007 8:25:06 AM EDT (CA) > To: dgi-seminaires(a)liste.polymtl.ca > Subject: [dgi-seminaires] séminaire GIGL > > SÉMINAIRE DU GIGL > > Titre: La création de l’avenir des multimédias interactifs > > Conférencier(ère): Christopher Pal (U. Rochester) > > Date: jeudi le 18 octobre de 11h30 à 12h20 > Lieu: L-4812, pavillon Lassonde de l'École Polytechnique > (campus de l'Université de Montréal) > > > RÉSUMÉ > > Notre capacité de capturer les informations a atteint des niveaux > sans précédent. Une manière puissante pour organiser et faciliter > les interactions avec ce déluge d'information est de créer de > nouvelles formes de multimédias interactifs. Pour cette > présentation, je démontrerai comment les modèles probabilistes et > les méthodes d'optimisation sont employés pour créer le noyau des > technologies qui permettent la création de nouvelles formes de > médias numériques (visuels) interactifs. > Je me concentrerai sur les méthodes et algorithmes pour utiliser > des représentations visuelles en tant qu'interfaces interactives > pour accéder aux informations. Je présenterai les tâches facilitant > la création des représentations interactives à haute résolution de > scènes visuelles, telles que les algorithmes pour les appareils > photo numériques interactifs ou pour les représentations virtuelles > de la Terre. J’illustrerai comment intégrer de multiples images > pour créer des représentations panoramiques de haute qualité et je > démontrerai une façon de prédire la structure tridimensionnelle de > scènes visuelles en utilisant une série d’images. En examinant les > étapes pour associer l'information sémantique des textes à ces > représentations visuelles riches, je montrerai comment des modèles > probabilistes basés sur les champs conditionnels aléatoires sont > employés pour classifier les documents et pour extraire > l'information sémantique. Je soulignerai mes contributions au > niveau de la construction de structures de modèles innovateurs et > de méthodes innovatrices pour optimiser les champs conditionnels > aléatoires. > > Pour plus d'information, > contactez Gilles Pesant (340-4711 poste 4142 ou > gilles.pesant(a)polymtl.ca) >

1 0