Distributed conditional computation

L'objectif de cette thèse est de présenter différentes applications du programme de recherche de calcul conditionnel distribué. On espère que ces applications, ainsi que la théorie présentée ici, mènera à une solution générale du problème d'intelligence artificielle, en particulier en ce qui a trait à la nécessité d'efficience. La vision du calcul conditionnel distribué consiste à accélérer l'évaluation et l'entraînement de modèles profonds, ce qui est très différent de l'objectif usuel d'améliorer sa capacité de généralisation et d'optimisation. Le travail présenté ici a des liens étroits avec les modèles de type mélange d'experts. Dans le chapitre 2, nous présentons un nouvel algorithme d'apprentissage profond qui utilise une forme simple d'apprentissage par renforcement sur un modèle d'arbre de décisions à base de réseau de neurones. Nous démontrons la nécessité d'une contrainte d'équilibre pour maintenir la distribution d'exemples aux experts uniforme et empêcher les monopoles. Pour rendre le calcul efficient, l'entrainement et l'évaluation sont contraints à être éparse en utilisant un routeur échantillonnant des experts d'une distribution multinomiale étant donné un exemple. Dans le chapitre 3, nous présentons un nouveau modèle profond constitué d'une représentation éparse divisée en segments d'experts. Un modèle de langue à base de réseau de neurones est construit à partir des transformations éparses entre ces segments. L'opération éparse par bloc est implémentée pour utilisation sur des cartes graphiques. Sa vitesse est comparée à deux opérations denses du même calibre pour démontrer le gain réel de calcul qui peut être obtenu. Un modèle profond utilisant des opérations éparses contrôlées par un routeur distinct des experts est entraîné sur un ensemble de données d'un milliard de mots. Un nouvel algorithme de partitionnement de données est appliqué sur un ensemble de mots pour hiérarchiser la couche de sortie d'un modèle de langage, la rendant ainsi beaucoup plus efficiente. Le travail présenté dans cette thèse est au centre de la vision de calcul conditionnel distribué émis par Yoshua Bengio. Elle tente d'appliquer la recherche dans le domaine des mélanges d'experts aux modèles profonds pour améliorer leur vitesse ainsi que leur capacité d'optimisation. Nous croyons que la théorie et les expériences de cette thèse sont une étape importante sur la voie du calcul conditionnel distribué car elle cadre bien le problème, surtout en ce qui concerne la compétitivité des systèmes d'experts.

Efficient training of RBF networks for classification.

Radial Basis Function networks with linear outputs are often used in regression problems because they can be substantially faster to train than Multi-layer Perceptrons. For classification problems, the use of linear outputs is less appropriate as the outputs are not guaranteed to represent probabilities. We show how RBFs with logistic and softmax outputs can be trained efficiently using the Fisher scoring algorithm. This approach can be used with any model which consists of a generalised linear output function applied to a model which is linear in its parameters. We compare this approach with standard non-linear optimisation algorithms on a number of datasets.

Bayesian classification with Gaussian processes

We consider the problem of assigning an input vector bfx to one of m classes by predicting P(c|bfx) for c = 1, ldots, m. For a two-class problem, the probability of class 1 given bfx is estimated by s(y(bfx)), where s(y) = 1/(1 + e^-y). A Gaussian process prior is placed on y(bfx), and is combined with the training data to obtain predictions for new bfx points. We provide a Bayesian treatment, integrating over uncertainty in y and in the parameters that control the Gaussian process prior; the necessary integration over y is carried out using Laplace's approximation. The method is generalized to multi-class problems (m >2) using the softmax function. We demonstrate the effectiveness of the method on a number of datasets.

Efficient training of RBF networks for classification

Radial Basis Function networks with linear outputs are often used in regression problems because they can be substantially faster to train than Multi-layer Perceptrons. For classification problems, the use of linear outputs is less appropriate as the outputs are not guaranteed to represent probabilities. In this paper we show how RBFs with logistic and softmax outputs can be trained efficiently using algorithms derived from Generalised Linear Models. This approach is compared with standard non-linear optimisation algorithms on a number of datasets.

Bayesian classification with Gaussian processes

We consider the problem of assigning an input vector to one of m classes by predicting P(c|x) for c=1,...,m. For a two-class problem, the probability of class one given x is estimated by s(y(x)), where s(y)=1/(1+e-y). A Gaussian process prior is placed on y(x), and is combined with the training data to obtain predictions for new x points. We provide a Bayesian treatment, integrating over uncertainty in y and in the parameters that control the Gaussian process prior the necessary integration over y is carried out using Laplace's approximation. The method is generalized to multiclass problems (m>2) using the softmax function. We demonstrate the effectiveness of the method on a number of datasets.

Disruption prediction at JET [Joint European Torus]

The sudden loss of the plasma magnetic confinement, known as disruption, is one of the major issue in a nuclear fusion machine as JET (Joint European Torus), Disruptions pose very serious problems to the safety of the machine. The energy stored in the plasma is released to the machine structure in few milliseconds resulting in forces that at JET reach several Mega Newtons. The problem is even more severe in the nuclear fusion power station where the forces are in the order of one hundred Mega Newtons. The events that occur during a disruption are still not well understood even if some mechanisms that can lead to a disruption have been identified and can be used to predict them. Unfortunately it is always a combination of these events that generates a disruption and therefore it is not possible to use simple algorithms to predict it. This thesis analyses the possibility of using neural network algorithms to predict plasma disruptions in real time. This involves the determination of plasma parameters every few milliseconds. A plasma boundary reconstruction algorithm, XLOC, has been developed in collaboration with Dr. D. Ollrien and Dr. J. Ellis capable of determining the plasma wall/distance every 2 milliseconds. The XLOC output has been used to develop a multilayer perceptron network to determine plasma parameters as ?i and q? with which a machine operational space has been experimentally defined. If the limits of this operational space are breached the disruption probability increases considerably. Another approach for prediction disruptions is to use neural network classification methods to define the JET operational space. Two methods have been studied. The first method uses a multilayer perceptron network with softmax activation function for the output layer. This method can be used for classifying the input patterns in various classes. In this case the plasma input patterns have been divided between disrupting and safe patterns, giving the possibility of assigning a disruption probability to every plasma input pattern. The second method determines the novelty of an input pattern by calculating the probability density distribution of successful plasma patterns that have been run at JET. The density distribution is represented as a mixture distribution, and its parameters arc determined using the Expectation-Maximisation method. If the dataset, used to determine the distribution parameters, covers sufficiently well the machine operational space. Then, the patterns flagged as novel can be regarded as patterns belonging to a disrupting plasma. Together with these methods, a network has been designed to predict the vertical forces, that a disruption can cause, in order to avoid that too dangerous plasma configurations are run. This network can be run before the pulse using the pre-programmed plasma configuration or on line becoming a tool that allows to stop dangerous plasma configuration. All these methods have been implemented in real time on a dual Pentium Pro based machine. The Disruption Prediction and Prevention System has shown that internal plasma parameters can be determined on-line with a good accuracy. Also the disruption detection algorithms showed promising results considering the fact that JET is an experimental machine where always new plasma configurations are tested trying to improve its performances.