SEMINAIRES MITACS - LISA sur l'Apprentissage Statistique

Université de Montréal
Organisateur: Hugo Larochelle


Pour accéder à notre nouvelle page web de séminaires cliquer ici.


The "efficiency sharpening" algorithms for OOM estimation [ pdf ]

Herbert Jaeger
School of Engineering and Science
International University Bremen

3 août, 10h00, 3195 PAA


Hidden Markov Models (HMMs) are widely used for blackbox modelling of symbolic, stochastic time series with memory. HMMs are usually trained using the expectation-maximization (EM) algorithm. This algorithm is note entirely satisfactory due to slow convergence and the presence of globally suboptimal solutions. Observable operator models (OOMs) present an alternative. At the surface OOMs appear almost like HMMs: both can be expressed in structurally identical matrix formalisms. However, unlike HMM representations, the matrices and state vectors of OOMs may contain negative components. This freedom in sign gives OOMs algebraic properties that radically differ from HMMs, and leads to novel learning algorithms that are fast and yield asymptotically correct model estimates. Unfortunately, the basic versions of these algorithms are statistically quite inefficient, which has so far precluded a widespread use of OOMs. This talk gives a condensed tutorial introduction to OOMs and introduces a novel approach to OOM estimation, called efficiency sharpening (ES). The ES method is iterative. In each iteration, the model estimated in the previous round is used to construct an estimator with a better statistical efficiency than the previous one. The computational load per iteration is less than an EM iteration, and only 2 to 5 iterations are typically needed. On real-life datasets, both the training and testing error of ES-trained OOMs is significantly lower than that of EM-trained HMMs. Because of the very nature of the learning algorithm (efficiency sharpening, not likelihood maximization), the bias-variance dilemma appears not to play the crucial role that it has in learning approaches based on fitting models to the training data.


Wavelet Transform for Preprocessing [ pdf ]

Alexandre Lacoste
Département d'Informatique et Recherche Opérationnelle
Université de Montréal

22 juillet, 15h00, 3195 PAA


Wavelet transforms have been succesfully used for image compression and denoising. It is therefore a good idea to try them for dimensionality reduction in data mining. Wavelets are also invariant to scale and translation which is usefull for multiresolution analysis.

For this presentation, we'll see what is a wavelet transform and how it is related to short-time fourier transform. Discrete wavelet transform is also a special case of filter banks. Finaly, we'll make an overview of different preprocession done with wavelet transform.


Stochastic Direct Reinforcement:
Policies with Recurrence [ pdf ]

John Moody
International Computer Science Institute
Berkeley & Portland

14 juillet, 15h00, 3195 PAA


A paradigm shift is underway in reinforcement learning (RL). The dominant approach to RL over the past 20 years has been based on dynamic programming, whereby RL agents learn an abstract value function (VF). An alternative approach, direct reinforcement (DR), has recently been revisited, wherein DR agents learn strategies to solve problems directly. DR can enable a simpler problem representation, mitigate Bellman's curse of dimensionality, and offer compelling advantages in efficiency.

This talk will begin with a short introduction to RL. I will briefly trace its history and origins, distinguish RL from other statistical learning approaches, and contrast value function methods with direct reinforcement.

I will then present a new algorithm called Stochastic Direct Reinforcement (SDR). This policy gradient algorithm is formulated for uncertain environments, partially-observed states and non-Markovian policies. Since SDR agents represent policies directly, they can naturally incorporate recurrent structure that is intrinsic to many potential applications and better solve the temporal credit assignment problem.

Demonstrations of DR include repeated games and trading financial markets. We show that SDR agents can learn winning strategies in simple competitive games. For the Iterated Prisoner's Dilemma, we find that non-recurrent SDR agents learn only the Nash equilibrium strategy of defection, while recurrent SDR agents can learn the global Pareto-optimal strategy of cooperation.

Time permitting, I'll give a brief survey of applications to trading financial markets.


Introduction à PLearn, Partie I: le point de vue de l'utilisateur [ html ]

Pascal Vincent, Christian Dorion, Nicolas Chapados
Département d'Informatique et Recherche Opérationnelle
Université de Montréal

23 juin, 13h30, 3195 PAA


Cette présentation a pour but une introduction au fonctionnement de l'ensemble des outils logiciels Open Source "PLearn" développés en partie au laboratoire LISA, et utilisés pour la recherche en apprentissage statistique. Elle s'adresse aussi bien aux nouveaux venus qu'aux personnes déjà familières avec PLearn (qui y apprendront à tirer partie des développements récents de la plateforme pour leur simplifier la vie!). La présentation prendra essentiellement la forme d'un tutoriel. Les concepts et classes de bases (ensemble de donnée, algorithme d'apprentissage, expérience d'apprentissage/test) y seront abordés au travers d'exemples et de manipulations avec les commandes et les scripts PLearn (.plearn et .pyplearn)

Il est à noter que cette présentation se fera en anglais. De plus, une présentation ultérieure (Partie II) sera consacrée au point de vue du développeur C++.


Apprentissage semi-supervisé à base de graphes [ pdf ]

Olivier Delalleau
Département d'Informatique et Recherche Opérationnelle
Université de Montréal

17 juin, 15h00, 3195 PAA


Cette présentation (identique à celle donnée au workshop CIAR en avril dernier) présentera (c'est original pour une présentation) plusieurs aspects de l'apprentissage semi-supervisé à base de graphes de proximité. J'essaierai de garder les choses simples pour que même Nicolas puisse comprendre, donc n'hésitez pas à venir même si vous n'y connaissez rien. Je crains qu'il n'y ait pas de beignes et café à la fin, par contre je parlerai du principe de la propagation de labels sur un graphe, ce qui est presque aussi intéressant. Les autres thèmes abordés seront le passage de la transduction à l'induction ainsi que la malédiction de la dimensionalité, sans oublier la faim dans le monde. Un beau programme en perspective, donc.


Statistical Language Learning by Human Infants [ ppt ]

Jenny Saffran
Department of Psychology
University of Wisconsin-Madison

6 mai, 15h00, 3195 PAA


Language acquisition is arguably the most complex learning problem facing developing humans. In my talk, I will consider new approaches to this problem involving the detection of statistical patterns in language input. Results from experiments with infants suggest that statistical learning may play a central role in human language acquisition.


A Bayesian account of animal learning [ pdf ]

Aaron Courville
Robotics Institute
Carnegie Mellon University

22 avril, 15h00, 3195 PAA


Configural conditioning experiments probe how animals discriminate and generalize between patterns of stimuli (such as tones and lights) that are differentially predictive of reinforcement. In this talk, I will present a Bayesian account of configural conditioning. According to our theory, an organism's learning process approximates statistical inference over a family of latent variable models in an attempt to recover the generative process that gave rise to the training data. In form, our model is reminiscent of the more phenomenological model of Pearce (1994); however, a normative grounding allows our theory to clarify seemingly arbitrary aspects of this and other previous models. Issues thought to be important in explaining configural conditioning, such as the choice of representation and generalization from previous experience , will be shown to follow from the properties of Bayesian statistical inference, principally the well known automatic Occam's razor effect. I will also show how our theory provides a novel explanation for conditioning phenomena, such as second-order conditioning and acquired equivalence effects.


Curse of Dimensionality: Local versus Non-Local Learning [ pdf ]

Yoshua Bengio
Département d'Informatique et Recherche Opérationnelle
Université de Montréal

15 avril, 14h30, Z-205 Pavillon Principal


We present a series of theoretical arguments supporting the claim that a large class of modern learning algorithms based on local kernels are sensitive to the curse of dimensionality. These include local manifold learning algorithms such as Isomap and LLE, and support vector classifiers with Gaussian or other local kernels. These algorithms are shown to be local in the sense that crucial properties of the learned function at x depend mostly on the neighbors of x in the training set. This makes them sensitive to the curse of dimensionality, well studied for classical non-parametric statistical learning. There is a large class of data distributions for which non-local solutions could be expressed compactly and potentially be learned with few examples, but which will require a large number of local bases and therefore a large number of training examples when using a local learning algorithm. We present new manifold learning algorithms that are non-local and manage to generalize far from training examples. They are like local PCA models, but in which the local directions are learnt as functions of x and of global parameters.


La transformée de Fourier pour les nuls [ pdf ]

Nicolas Le Roux
Département d'Informatique et Recherche Opérationnelle
Université de Montréal

25 mars, 14h30, 3195 PAA


La transformée de Fourier est un outil prépondérant en traitement du signal et en Mathématiques en général. Il permet de passer de manière (quasi-) bijective du domaine temporel au domaine fréquentiel. Elle jouit de nombreuses et très utiles propriétés mathématiques qui permettent de faciliter l'étude de certains types de fonctions (j'essaye d'être encore plus vague, mais là vraiment je n'y arrive pas).

Je présenterai donc tout d'abord la formule mathématique de la transformée de Fourier, sa signification en termes mathématiques et physiques, puis ses propriétés.

Pour intéresser le plus grand nombre, la présentation devrait être assez peu orientée mathématiques (enfin, un peu quand même, faut pas déconner), mais je pourrais pousser cette partie si la demande en est faite.


Inferring Dynamical Models of Gene Expression [ pdf ]

Ted Perkins
McGill Centre for Bioinformatics
McGill University

11 mars, 14h30, 3195 PAA


Identifying regulatory relationships between genes, and describing the resulting dynamics of gene activity, is a key problem facing biology today. Progress in this area impacts our understanding of normal and abnormal development, of disease, and even of evolution itself. In this talk, I will present my work on inferring continuous-time switching networks for modelling the dynamics of gene expression. Continuous-time switching networks are a special type of differential equation in which variables can change continuously but time derivatives are discrete, typically specified by logical rules.

I will present new theoretical bounds on how much data is needed to solve the inference problem, as a function of parameters such as the network size and how often data is sampled. I will also discuss the computational complexity of the problem. Perhaps the most interesting observation is that the sampling rate controls a trade-off between sample complexity and computational complexity. By collecting more data, we can ease the computational burden of inferring the model, and vice-versa. Finally, I will describe an application to modelling the gap gene system---a prototypical, but still controversial, network of genes crucial to early development in the Drosophila Melanogaster (fruit fly) embryo. The model obtained extends previous, manually constructed logical models to account for more of the observed gap gene patterns and uses simpler logical rules to do so.


"Best Of" NIPS (partie 3)

Hugo Larochelle, Olivier Delalleau, Li Gen Wang et Nicolas Chapados
Département d'Informatique et Recherche Opérationnelle
Université de Montréal

25 février, 14h30, 3195 PAA


Titre: Integrating Topics and Syntax [ pdf ]
Auteurs: Thomas L. Griffiths, Mark Steyvers, David M. Blei,Joshua B. Tenenbaum
Par: Hugo Larochelle

Titre: Adaptive Manifold Learning [ pdf ]
Auteurs: Jing Wang, Zhenyue Zhang, Hongyuan Zha
Par: Olivier Delalleau

Titre: Maximum likelihood estimation of intrinsic dimension [ ppt ]
Auteurs: Elizaveta Levina, Peter Bickel
Par: Li Gen Wang

Titre: An Investigation of Practical Approximate Nearest Neighbor Algorithms [ ppt ]
Auteurs: Ting Liu, Andrew W. Moore, Alexander Gray, Ke Yang
Par: Nicolas Chapados



"Best Of" NIPS (partie 2)

Nicolas Le Roux, Norman Casagrande et Francois Rivest
Département d'Informatique et Recherche Opérationnelle
Université de Montréal

18 février, 14h30, 3195 PAA


Titre: A Generalized Bradley-Terry Model: From Group Competition to Individual Skill [ pdf ]
Auteurs: Tzu-Kuo Huang, Chih-Jen Lin, Ruby C. Weng
Par: Nicolas Le Roux

Titre: Contextual models for object detection using boosted [ ppt ]
Auteurs: Antonio Torralba, Kevin P. Murphy et William T. Freeman
Par: Norman Casagrande

Titre: Intrinsically Motivated Reinforcement Learning [ pdf ]
Auteurs: Singh, Barto et Chentanez
Par: Francois Rivest



"Best Of" NIPS (partie 1)

Julie Carreau, Karim Aboud Moustafa et Dumitru Erhan
Département d'Informatique et Recherche Opérationnelle
Université de Montréal

11 février, 14h30, 3195 PAA


Titre: Outlier Detection with One-class Kernel Fisher Discriminants [ pdf ]
Auteur: Volker Roth
Par: Julie Carreau

Titre: Proximity graphs for clustering and manifold learning [ pdf ]
Auteurs: Miguel A. Carreira-Perpinan, Richard S. Zemel
Par: Karim Aboud Moustafa

Titre: Whos In the Picture [ pdf ]
Auteurs: Tamara L. Berg, Alexander C. Berg, Jaety Edwards, David Forsyth
Par: Dumitru Erhan




Réseaux de neurones récurrents et applications en musique [ flash ]

Jasmin Lapalme
Département d'Informatique et Recherche Opérationnelle
Université de Montréal

28 janvier, 14h30, 3195 PAA


La présentation portera sur le calcul de gradient dans les réseaux de neurones récurrents BPTT (Backpropagation through time) et LSTM(Long short term memory) et comment peut-on utiliser ces réseaux pour concevoir des modèles génératifs de musique automatique. J'élaborai sur le processus d'entraînement, choix des bases de données et représentation des notes de musique, et de la génération de mélodie. Des exemples de mélodie générés par LSTM et BPTT seront comparés et discutés.


Applications récentes d'algorithmes d'apprentissage pour la découverte de médicament [ pdf ]

Pierre-Jean L'Heureux
Département d'Informatique et Recherche Opérationnelle
Université de Montréal

21 janvier, 14h30, 3195 PAA


L'automatisation de la recherche pharmaceutique, autant du côté chimique que biologique, génère aujourd'hui de grande quantité d'information dans les premières phases de la découverte de médicament. Mise ensemble, les librairies de chimie combinatoire et les réponses biologiques de criblage haute performance (HTS) produisent des jeux de données étiquetées assez grands pour appliquer des algorithmes d'apprentissage statistique.

Notre recherche vise à produire des modèles capables de filtrer les molécules les plus susceptibles de devenir des médicaments. Dans ce séminaire, nous présenterons nos résultats sur l'utilisation d'algorithmes de réduction de dimensions non-linéaires. J'introduirai aussi de nouvelles expériences d'apprentissage multi-tâche, sous l'angle du "filtrage collaboratif".


Quantifying the Uncertainty of a Belief Net Response

Russ Greiner
Department of Computing Science and Alberta Ingenuity Centre for Machine Learning
University of Alberta

23 août 2004, 15h, 3195 PAA


A Bayesian Belief Network (BN) models a joint distribution over a set of n variables, using a DAG structure to represent the immediate dependencies between the variables, and a set of parameters (aka "CPTables") to represent the local conditional probabilities of a node, given each assignment to its parents. In many situations, these parameters are themselves random variables --- this may reflect the uncertainty of the domain expert, or may come from a training sample used to estimate the parameter values. The distribution over these "CPtable variables" induces a distribution over the response the BN will return to any "What is Pr(Q=q | E=e)?" query. This paper investigates properties of this response: showing first that it is asymptotically normal, then providing, in closed form, its mean and asymptotic variance. We then present an effective general algorithm for computing this variance, which has the same complexity as simply computing (the mean value of) the response itself --- ie, O(n 2^w), where w is the effective tree width. Finally, we provide empirical evidence that a Beta approximation works much better than the normal distribution, especially for small sample sizes, and that our algorithm works effectively in practice, over a range of belief net structures, sample sizes and queries.

This is joint work with Tim Van Allen, Ajit Singh and Peter Hooper.
http://www.cs.ualberta.ca/~greiner/BN-ErrBar.html