Improving sampling, optimization and feature extraction in Boltzmann machines

L’apprentissage supervisé de réseaux hiérarchiques à grande échelle connaît présentement un succès fulgurant. Malgré cette effervescence, l’apprentissage non-supervisé représente toujours, selon plusieurs chercheurs, un élément clé de l’Intelligence Artificielle, où les agents doivent apprendre à partir d’un nombre potentiellement limité de données. Cette thèse s’inscrit dans cette pensée et aborde divers sujets de recherche liés au problème d’estimation de densité par l’entremise des machines de Boltzmann (BM), modèles graphiques probabilistes au coeur de l’apprentissage profond. Nos contributions touchent les domaines de l’échantillonnage, l’estimation de fonctions de partition, l’optimisation ainsi que l’apprentissage de représentations invariantes. Cette thèse débute par l’exposition d’un nouvel algorithme d'échantillonnage adaptatif, qui ajuste (de fa ̧con automatique) la température des chaînes de Markov sous simulation, afin de maintenir une vitesse de convergence élevée tout au long de l’apprentissage. Lorsqu’utilisé dans le contexte de l’apprentissage par maximum de vraisemblance stochastique (SML), notre algorithme engendre une robustesse accrue face à la sélection du taux d’apprentissage, ainsi qu’une meilleure vitesse de convergence. Nos résultats sont présent ́es dans le domaine des BMs, mais la méthode est générale et applicable à l’apprentissage de tout modèle probabiliste exploitant l’échantillonnage par chaînes de Markov. Tandis que le gradient du maximum de vraisemblance peut-être approximé par échantillonnage, l’évaluation de la log-vraisemblance nécessite un estimé de la fonction de partition. Contrairement aux approches traditionnelles qui considèrent un modèle donné comme une boîte noire, nous proposons plutôt d’exploiter la dynamique de l’apprentissage en estimant les changements successifs de log-partition encourus à chaque mise à jour des paramètres. Le problème d’estimation est reformulé comme un problème d’inférence similaire au filtre de Kalman, mais sur un graphe bi-dimensionnel, où les dimensions correspondent aux axes du temps et au paramètre de température. Sur le thème de l’optimisation, nous présentons également un algorithme permettant d’appliquer, de manière efficace, le gradient naturel à des machines de Boltzmann comportant des milliers d’unités. Jusqu’à présent, son adoption était limitée par son haut coût computationel ainsi que sa demande en mémoire. Notre algorithme, Metric-Free Natural Gradient (MFNG), permet d’éviter le calcul explicite de la matrice d’information de Fisher (et son inverse) en exploitant un solveur linéaire combiné à un produit matrice-vecteur efficace. L’algorithme est prometteur: en terme du nombre d’évaluations de fonctions, MFNG converge plus rapidement que SML. Son implémentation demeure malheureusement inefficace en temps de calcul. Ces travaux explorent également les mécanismes sous-jacents à l’apprentissage de représentations invariantes. À cette fin, nous utilisons la famille de machines de Boltzmann restreintes “spike & slab” (ssRBM), que nous modifions afin de pouvoir modéliser des distributions binaires et parcimonieuses. Les variables latentes binaires de la ssRBM peuvent être rendues invariantes à un sous-espace vectoriel, en associant à chacune d’elles, un vecteur de variables latentes continues (dénommées “slabs”). Ceci se traduit par une invariance accrue au niveau de la représentation et un meilleur taux de classification lorsque peu de données étiquetées sont disponibles. Nous terminons cette thèse sur un sujet ambitieux: l’apprentissage de représentations pouvant séparer les facteurs de variations présents dans le signal d’entrée. Nous proposons une solution à base de ssRBM bilinéaire (avec deux groupes de facteurs latents) et formulons le problème comme l’un de “pooling” dans des sous-espaces vectoriels complémentaires.

Verfahren zur echtzeitfähigen Simulation dynamischer Emissionsverläufe ausgewählter Schadstoffe von Ottomotoren

Im Rahmen dieser Arbeit werden Modellbildungsverfahren zur echtzeitfähigen Simulation wichtiger Schadstoffkomponenten im Abgasstrom von Verbrennungsmotoren vorgestellt. Es wird ein ganzheitlicher Entwicklungsablauf dargestellt, dessen einzelne Schritte, beginnend bei der Ver-suchsplanung über die Erstellung einer geeigneten Modellstruktur bis hin zur Modellvalidierung, detailliert beschrieben werden. Diese Methoden werden zur Nachbildung der dynamischen Emissi-onsverläufe relevanter Schadstoffe des Ottomotors angewendet. Die abgeleiteten Emissionsmodelle dienen zusammen mit einer Gesamtmotorsimulation zur Optimierung von Betriebstrategien in Hybridfahrzeugen. Im ersten Abschnitt der Arbeit wird eine systematische Vorgehensweise zur Planung und Erstellung von komplexen, dynamischen und echtzeitfähigen Modellstrukturen aufgezeigt. Es beginnt mit einer physikalisch motivierten Strukturierung, die eine geeignete Unterteilung eines Prozessmodells in einzelne überschaubare Elemente vorsieht. Diese Teilmodelle werden dann, jeweils ausgehend von einem möglichst einfachen nominalen Modellkern, schrittweise erweitert und ermöglichen zum Abschluss eine robuste Nachbildung auch komplexen, dynamischen Verhaltens bei hinreichender Genauigkeit. Da einige Teilmodelle als neuronale Netze realisiert werden, wurde eigens ein Verfah-ren zur sogenannten diskreten evidenten Interpolation (DEI) entwickelt, das beim Training einge-setzt, und bei minimaler Messdatenanzahl ein plausibles, also evidentes Verhalten experimenteller Modelle sicherstellen kann. Zum Abgleich der einzelnen Teilmodelle wurden statistische Versuchs-pläne erstellt, die sowohl mit klassischen DoE-Methoden als auch mittels einer iterativen Versuchs-planung (iDoE ) generiert wurden. Im zweiten Teil der Arbeit werden, nach Ermittlung der wichtigsten Einflussparameter, die Model-strukturen zur Nachbildung dynamischer Emissionsverläufe ausgewählter Abgaskomponenten vor-gestellt, wie unverbrannte Kohlenwasserstoffe (HC), Stickstoffmonoxid (NO) sowie Kohlenmono-xid (CO). Die vorgestellten Simulationsmodelle bilden die Schadstoffkonzentrationen eines Ver-brennungsmotors im Kaltstart sowie in der anschließenden Warmlaufphase in Echtzeit nach. Im Vergleich zur obligatorischen Nachbildung des stationären Verhaltens wird hier auch das dynami-sche Verhalten des Verbrennungsmotors in transienten Betriebsphasen ausreichend korrekt darge-stellt. Eine konsequente Anwendung der im ersten Teil der Arbeit vorgestellten Methodik erlaubt, trotz einer Vielzahl von Prozesseinflussgrößen, auch hier eine hohe Simulationsqualität und Ro-bustheit. Die Modelle der Schadstoffemissionen, eingebettet in das dynamische Gesamtmodell eines Ver-brennungsmotors, werden zur Ableitung einer optimalen Betriebsstrategie im Hybridfahrzeug ein-gesetzt. Zur Lösung solcher Optimierungsaufgaben bieten sich modellbasierte Verfahren in beson-derer Weise an, wobei insbesondere unter Verwendung dynamischer als auch kaltstartfähiger Mo-delle und der damit verbundenen Realitätsnähe eine hohe Ausgabequalität erreicht werden kann.

A benchmark method for global optimization problems in structure determination from powder diffraction data

Quasi-Newton-Raphson minimization and conjugate gradient minimization have been used to solve the crystal structures of famotidine form B and capsaicin from X-ray powder diffraction data and characterize the chi(2) agreement surfaces. One million quasi-Newton-Raphson minimizations found the famotidine global minimum with a frequency of ca 1 in 5000 and the capsaicin global minimum with a frequency of ca 1 in 10 000. These results, which are corroborated by conjugate gradient minimization, demonstrate the existence of numerous pathways from some of the highest points on these chi(2) agreement surfaces to the respective global minima, which are passable using only downhill moves. This important observation has significant ramifications for the development of improved structure determination algorithms.

GSHMC: an efficient method for molecular simulation

The hybrid Monte Carlo (HMC) method is a popular and rigorous method for sampling from a canonical ensemble. The HMC method is based on classical molecular dynamics simulations combined with a Metropolis acceptance criterion and a momentum resampling step. While the HMC method completely resamples the momentum after each Monte Carlo step, the generalized hybrid Monte Carlo (GHMC) method can be implemented with a partial momentum refreshment step. This property seems desirable for keeping some of the dynamic information throughout the sampling process similar to stochastic Langevin and Brownian dynamics simulations. It is, however, ultimate to the success of the GHMC method that the rejection rate in the molecular dynamics part is kept at a minimum. Otherwise an undesirable Zitterbewegung in the Monte Carlo samples is observed. In this paper, we describe a method to achieve very low rejection rates by using a modified energy, which is preserved to high-order along molecular dynamics trajectories. The modified energy is based on backward error results for symplectic time-stepping methods. The proposed generalized shadow hybrid Monte Carlo (GSHMC) method is applicable to NVT as well as NPT ensemble simulations.

Partial spectral projected gradient method with active-set strategy for linearly constrained optimization

A method for linearly constrained optimization which modifies and generalizes recent box-constraint optimization algorithms is introduced. The new algorithm is based on a relaxed form of Spectral Projected Gradient iterations. Intercalated with these projected steps, internal iterations restricted to faces of the polytope are performed, which enhance the efficiency of the algorithm. Convergence proofs are given and numerical experiments are included and commented. Software supporting this paper is available through the Tango Project web page: http://www.ime.usp.br/similar to egbirgin/tango/.

Continuous GRASP with a local active-set method for bound-constrained global optimization

Global optimization seeks a minimum or maximum of a multimodal function over a discrete or continuous domain. In this paper, we propose a hybrid heuristic-based on the CGRASP and GENCAN methods-for finding approximate solutions for continuous global optimization problems subject to box constraints. Experimental results illustrate the relative effectiveness of CGRASP-GENCAN on a set of benchmark multimodal test functions.

Outer Trust-Region Method for Constrained Optimization

Given an algorithm A for solving some mathematical problem based on the iterative solution of simpler subproblems, an outer trust-region (OTR) modification of A is the result of adding a trust-region constraint to each subproblem. The trust-region size is adaptively updated according to the behavior of crucial variables. The new subproblems should not be more complex than the original ones, and the convergence properties of the OTR algorithm should be the same as those of Algorithm A. In the present work, the OTR approach is exploited in connection with the ""greediness phenomenon"" of nonlinear programming. Convergence results for an OTR version of an augmented Lagrangian method for nonconvex constrained optimization are proved, and numerical experiments are presented.

Structured minimal-memory inexact quasi-Newton method and secant preconditioners for augmented Lagrangian optimization

Augmented Lagrangian methods for large-scale optimization usually require efficient algorithms for minimization with box constraints. On the other hand, active-set box-constraint methods employ unconstrained optimization algorithms for minimization inside the faces of the box. Several approaches may be employed for computing internal search directions in the large-scale case. In this paper a minimal-memory quasi-Newton approach with secant preconditioners is proposed, taking into account the structure of Augmented Lagrangians that come from the popular Powell-Hestenes-Rockafellar scheme. A combined algorithm, that uses the quasi-Newton formula or a truncated-Newton procedure, depending on the presence of active constraints in the penalty-Lagrangian function, is also suggested. Numerical experiments using the Cute collection are presented.

Optimization of a method for determination of phenolic acids in exotic fruits by capillary electrophoresis

In this work, the separation of nine phenolic acids (benzoic, caffeic, chlorogenic, p-coumaric, ferulic, gallic, protocatechuic, syringic, and vanillic acid) was approached by a 32 factorial design in electrolytes consisting of sodium tetraborate buffer(STB) in the concentration range of 10-50 mmol L(-1) and methanol in the volume percentage of 5-20%. Derringer`s desirability functions combined globally were tested as response functions. An optimal electrolyte composed by 50 mmol L(-1) tetraborate buffer at pH 9.2, and 7.5% (v/v) methanol allowed baseline resolution of all phenolic acids under investigation in less than 15 min. In order to promote sample clean up, to preconcentrate the phenolic fraction and to release esterified phenolic acids from the fruit matrix, elaborate liquid-liquid extraction procedures followed by alkaline hydrolysis were performed. The proposed methodology was fully validated (linearity from 10.0 to 100 mu g mL(-1), R(2) > 0.999: LOD and LOQ from 1.32 to 3.80 mu g mL(-1) and from 4.01 to 11.5 mu g mL(-1), respectively; intra-day precision better than 2.8% CV for migration time and 5.4% CV for peak area; inter-day precision better than 4.8% CV for migration time and 4.8-11% CV for peak area: recoveries from 81% to 115%) and applied successfully to the evaluation of phenolic contents of abiu-roxo (Chrysophyllum caimito), wild mulberry growing in Brazil (Morus nigra L.) and tree tomato (Cyphomandra betacea). Values in the range of 1.50-47.3 mu g g(-1) were found, with smaller amounts occurring as free phenolic acids. (C) 2009 Elsevier B.V. All rights reserved.

Determination of sorbate and benzoate in beverage samples by capillary electrophoresis - Optimization of the method with inspection of ionic mobilities

The aim of this study was to develop a fast capillary electrophoresis method for the determination of benzoate and sorbate ions in commercial beverages. In the method development the pH and constituents of the background electrolyte were selected using the effective mobility versus pH curves. As the high resolution obtained experimentally for sorbate and benzoate in the studies presented in the literature is not in agreement with that expected from the ionic mobility values published, a procedure to determine these values was carried out. The salicylate ion was used as the internal standard. The background electrolyte was composed of 25 mmol L(-1) tris(hydroxymethyl)aminomethane and 12.5 mmol L(-1) 2-hydroxyisobutyric acid, atpH 8.1.Separation was conducted in a fused-silica capillary(32 cm total length and 8.5 cm effective length, 50 mu m I.D.), with short-end injection configuration and direct UV detection at 200 nm for benzoate and salicylate and 254 nm for sorbate ions. The run time was only 28 s. A few figures of merit of the proposed method include: good linearity (R(2) > 0.999), limit of detection of 0.9 and 0.3 mg L(-1) for benzoate and sorbate, respectively, inter-day precision better than 2.7% (n =9) and recovery in the range 97.9-105%. Beverage samples were prepared by simple dilution with deionized water (1:11, v/v). Concentrations in the range of 197-401 mg L(-1) for benzoate and 28-144 mg L(-1) for sorbate were found in soft drinks and tea. (c) 2008 Elsevier B.V. All rights reserved.

Design/Evaluation of A Methodology For Performance Optimization Of Indexable Carbide Inserts

In this project, two broad facets in the design of a methodology for performance optimization of indexable carbide inserts were examined. They were physical destructive testing and software simulation.For the physical testing, statistical research techniques were used for the design of the methodology. A five step method which began with Problem definition, through System identification, Statistical model formation, Data collection and Statistical analyses and results was indepthly elaborated upon. Set-up and execution of an experiment with a compression machine together with roadblocks and possible solution to curb road blocks to quality data collection were examined. 2k factorial design was illustrated and recommended for process improvement. Instances of first-order and second-order response surface analyses were encountered. In the case of curvature, test for curvature significance with center point analysis was recommended. Process optimization with method of steepest ascent and central composite design or process robustness studies of response surface analyses were also recommended.For the simulation test, AdvantEdge program was identified as the most used software for tool development. Challenges to the efficient application of this software were identified and possible solutions proposed. In conclusion, software simulation and physical testing were recommended to meet the objective of the project.