Cette thèse considère l'apprentissage de motifs récurrents dans la perception multimodale. Elle s'attache à développer des modèles robotiques de ces facultés telles qu'observées chez l'enfant, et elle s'inscrit en cela dans le domaine de la robotique développementale.Elle s'articule plus précisément autour de deux thèmes principaux qui sont d'une part la capacité d'enfants ou de robots à imiter et à comprendre le comportement d'humains, et d'autre part l'acquisition du langage. A leur intersection, nous examinons la question de la découverte par un agent en développement d'un répertoire de motifs primitifs dans son flux perceptuel. Nous spécifions ce problème et établissons son lien avec ceux de l'indétermination de la traduction décrit par Quine et de la séparation aveugle de source tels qu'étudiés en acoustique.Nous en étudions successivement quatre sous-problèmes et formulons une définition expérimentale de chacun. Des modèles d'agents résolvant ces problèmes sont également décrits et testés. Ils s'appuient particulièrement sur des techniques dites de sacs de mots, de factorisation de matrices et d'apprentissage par renforcement inverse. Nous approfondissons séparément les trois problèmes de l'apprentissage de sons élémentaires tels les phonèmes ou les mots, de mouvements basiques de danse et d'objectifs primaires composant des tâches motrices complexes. Pour finir nous étudions le problème de l'apprentissage d'éléments primitifs multimodaux, ce qui revient à résoudre simultanément plusieurs des problèmes précédents. Nous expliquons notamment en quoi cela fournit un modèle de l'ancrage de mots acoustiques / This thesis focuses on learning recurring patterns in multimodal perception. For that purpose it develops cognitive systems that model the mechanisms providing such capabilities to infants; a methodology that fits into thefield of developmental robotics.More precisely, this thesis revolves around two main topics that are, on the one hand the ability of infants or robots to imitate and understand human behaviors, and on the other the acquisition of language. At the crossing of these topics, we study the question of the how a developmental cognitive agent can discover a dictionary of primitive patterns from its multimodal perceptual flow. We specify this problem and formulate its links with Quine's indetermination of translation and blind source separation, as studied in acoustics.We sequentially study four sub-problems and provide an experimental formulation of each of them. We then describe and test computational models of agents solving these problems. They are particularly based on bag-of-words techniques, matrix factorization algorithms, and inverse reinforcement learning approaches. We first go in depth into the three separate problems of learning primitive sounds, such as phonemes or words, learning primitive dance motions, and learning primitive objective that compose complex tasks. Finally we study the problem of learning multimodal primitive patterns, which corresponds to solve simultaneously several of the aforementioned problems. We also details how the last problems models acoustic words grounding.

In the studies underlying this thesis, various source tracing techniques were applied to environmental samples from the Baltic region. Comprehensive sampling and analysis of polychlorinated dibenzo-p-dioxins (PCDDs) and polychlorinated dibenzofurans (PCDFs) in surface sediments in Swedish coastal and offshore areas resulted in a unique data set for this region. Nearly 150 samples of surface sediments were analyzed for all tetra- to octa-chlorinated PCDD/Fs. The levels showed large spatial variability with hotspots in several coastal regions. Neither Sweden nor the EU has introduced guideline values for PCDD/Fs in sediment, but comparisons to available guidelines and quality standards from other countries indicate that large areas of primarily coastal sediments may constitute a risk to marine organisms. Multivariate pattern analysis techniques and receptor models, such as Principal Component Analysis (PCA) and Positive Matrix Factorization (PMF), were used to trace sources. These analyses suggested that three to six source types can explain most of the observed pattern variations found in the sediment samples. Atmospheric deposition was suggested as the most important source to offshore areas, thus confirming earlier estimates. However, spatial differences indicated a larger fraction of local/regional atmospheric sources, characterized by PCDFs, in the south. This was indicated by the identification of several patterns of atmospheric origin. In coastal areas, the influence of direct emission sources was larger, and among these, chlorophenol used for wood preservation and emissions from pulp/paper production and other wood related industry appeared to be most important. The historic emissions connected to processes involving chemical reactions with chlorine (e.g. pulp bleaching) were found to be of less importance except at some coastal sites. The analysis of PCDD/Fs in Baltic herring also revealed spatial variations in the levels and pollution patterns along the coast. The geographical match against areas with elevated sediment levels indicated that transfer from sediments via water to organisms was one possible explanation. Fugacity, a concept used to predict the net transport direction between environmental matrices, was used to explore the gas exchange of hexachlorocyclohexanes (HCHs) and polychlorinated biphenyls (PCBs) between air and water. These estimates suggested that, in the Kattegat Sea, the gaseous exchange of HCHs primarily resulted in net deposition while PCBs were net volatilized under certain environmental conditions. The study also indicated that, while the air concentrations of both PCBs and γ-HCH are mostly dependent upon the origin of the air mass, the fluctuations in α-HCH were primarily influenced by seasonal changes.

On étudie l’application des algorithmes de décomposition matricielles tel que la Factorisation Matricielle Non-négative (FMN), aux représentations fréquentielles de signaux audio musicaux. Ces algorithmes, dirigés par une fonction d’erreur de reconstruction, apprennent un ensemble de fonctions de base et un ensemble de coef- ficients correspondants qui approximent le signal d’entrée. On compare l’utilisation de trois fonctions d’erreur de reconstruction quand la FMN est appliquée à des gammes monophoniques et harmonisées: moindre carré, divergence Kullback-Leibler, et une mesure de divergence dépendente de la phase, introduite récemment. Des nouvelles méthodes pour interpréter les décompositions résultantes sont présentées et sont comparées aux méthodes utilisées précédemment qui nécessitent des connaissances du domaine acoustique. Finalement, on analyse la capacité de généralisation des fonctions de bases apprises par rapport à trois paramètres musicaux: l’amplitude, la durée et le type d’instrument. Pour ce faire, on introduit deux algorithmes d’étiquetage des fonctions de bases qui performent mieux que l’approche précédente dans la majorité de nos tests, la tâche d’instrument avec audio monophonique étant la seule exception importante. / We study the application of unsupervised matrix decomposition algorithms such as Non-negative Matrix Factorization (NMF) to frequency domain representations of music audio signals. These algorithms, driven by a given reconstruction error function, learn a set of basis functions and a set of corresponding coefficients that approximate the input signal. We compare the use of three reconstruction error functions when NMF is applied to monophonic and harmonized musical scales: least squares, Kullback-Leibler divergence, and a recently introduced “phase-aware” divergence measure. Novel supervised methods for interpreting the resulting decompositions are presented and compared to previously used methods that rely on domain knowledge. Finally, the ability of the learned basis functions to generalize across musical parameter values including note amplitude, note duration and instrument type, are analyzed. To do so, we introduce two basis function labeling algorithms that outperform the previous labeling approach in the majority of our tests, instrument type with monophonic audio being the only notable exception.

Sentiment Analysis is an area of Computer Science that deals with the impact a document makes on a user. The very field is further sub-divided into Opinion Mining and Emotion Analysis, the latter of which is the basis for the present work. Work on songs is aimed at building affective interactive applications such as music recommendation engines. Using song lyrics, we are interested in both supervised and unsupervised analyses, each of which has its own pros and cons. For an unsupervised analysis (clustering), we use a standard probabilistic topic model called Latent Dirichlet Allocation (LDA). It mines topics from songs, which are nothing but probability distributions over the vocabulary of words. Some of the topics seem sentiment-based, motivating us to continue with this approach. We evaluate our clusters using a gold dataset collected from an apt website and get positive results. This approach would be useful in the absence of a supervisor dataset. In another part of our work, we argue the inescapable existence of supervision in terms of having to manually analyse the topics returned. Further, we have also used explicit supervision in terms of a training dataset for a classifier to learn sentiment specific classes. This analysis helps reduce dimensionality and improve classification accuracy. We get excellent dimensionality reduction using Support Vector Machines (SVM) for feature selection. For re-classification, we use the Naive Bayes Classifier (NBC) and SVM, both of which perform well. We also use Non-negative Matrix Factorization (NMF) for classification, but observe that the results coincide with those of NBC, with no exceptions. This drives us towards establishing a theoretical equivalence between the two.

La séparation aveugle de sources audio dans le cas sous-déterminé est un problème mathématique complexe dont il est aujourd'hui possible d'obtenir une solution satisfaisante, à condition de sélectionner la méthode la plus adaptée au problème posé et de savoir paramétrer celle-ci soigneusement. Afin d'automatiser cette étape de sélection déterminante, nous proposons dans cette thèse de recourir au principe de fusion. L'idée est simple : il s'agit, pour un problème donné, de sélectionner plusieurs méthodes de résolution plutôt qu'une seule et de les combiner afin d'en améliorer la solution. Pour cela, nous introduisons un cadre général de fusion qui consiste à formuler l'estimée d'une source comme la combinaison de plusieurs estimées de cette même source données par différents algorithmes de séparation, chaque estimée étant pondérée par un coefficient de fusion. Ces coefficients peuvent notamment être appris sur un ensemble d'apprentissage représentatif du problème posé par minimisation d'une fonction de coût liée à l'objectif de séparation. Pour aller plus loin, nous proposons également deux approches permettant d'adapter les coefficients de fusion au signal à séparer. La première formule la fusion dans un cadre bayésien, à la manière du moyennage bayésien de modèles. La deuxième exploite les réseaux de neurones profonds afin de déterminer des coefficients de fusion variant en temps. Toutes ces approches ont été évaluées sur deux corpus distincts : l'un dédié au rehaussement de la parole, l'autre dédié à l'extraction de voix chantée. Quelle que soit l'approche considérée, nos résultats montrent l'intérêt systématique de la fusion par rapport à la simple sélection, la fusion adaptative par réseau de neurones se révélant être la plus performante. / Underdetermined blind source separation is a complex mathematical problem that can be satisfyingly resolved for some practical applications, providing that the right separation method has been selected and carefully tuned. In order to automate this selection process, we propose in this thesis to resort to the principle of fusion which has been widely used in the related field of classification yet is still marginally exploited in source separation. Fusion consists in combining several methods to solve a given problem instead of selecting a unique one. To do so, we introduce a general fusion framework in which a source estimate is expressed as a linear combination of estimates of this same source given by different separation algorithms, each source estimate being weighted by a fusion coefficient. For a given task, fusion coefficients can then be learned on a representative training dataset by minimizing a cost function related to the separation objective. To go further, we also propose two ways to adapt the fusion coefficients to the mixture to be separated. The first one expresses the fusion of several non-negative matrix factorization (NMF) models in a Bayesian fashion similar to Bayesian model averaging. The second one aims at learning time-varying fusion coefficients thanks to deep neural networks. All proposed methods have been evaluated on two distinct corpora. The first one is dedicated to speech enhancement while the other deals with singing voice extraction. Experimental results show that fusion always outperform simple selection in all considered cases, best results being obtained by adaptive time-varying fusion with neural networks.

Collaborative filtering is a popular technique to use behavior data in the form of user’s interactions with, or ratings of, items in a system to provide personalized recommendations of items to the user. This study compares three different state-of-the-art Recommendation System models that implement this technique, Matrix Factorization, Multi-layer Perceptron and Neural Matrix Factorization, using behavior data from a digital book platform for children. The field of Recommendation Systems is growing, and many platforms can benefit of personalizing the user experience and simplifying the use of the platforms. To perform a more complex comparison and introduce a new take on the models, this study proposes a new way to represent the behavior data as input to the models, i.e., to use the Term Frequency-Inverse Document Frequency (TFIDF) of occurrences of interactions between users and books, as opposed to the traditional binary representation (positive if there has been any interaction and negative otherwise). The performance is measured by extracting the last book read for each user, and evaluating how the models would rank that book for recommendations to the user. To assess the value of the models for the children’s reading platform, the models are also compared to the existing Recommendation System on the digital book platform. The results indicate that the Matrix Factorization model performs best out of the three models when using children’s reading behavior data. However, due to the long training process and larger set of hyperparameters to tune for the other two models, these may not have reached an optimal hyperparameter tuning, thereby affecting the comparison among the three state-of-the-art models. This limitation is further discussed in the study. All three models perform significantly better than the current system on the digital book platform. The models with the proposed representation using TF-IDF values show notable promise, performing better than the binary representation in almost all numerical metrics for all models. These results can suggest future research work on more ways of representing behavior data as input to these types of models. / Kollaborativ filtrering är en populär teknik för att använda beteendedata från användare i form av t.ex. interaktioner med, eller betygsättning av, objekt i ett system för att ge användaren personliga rekommendationer om objekt. I den här studien jämförs tre olika modeller av moderna rekommendationssystem som tillämpar denna teknik, matrisfaktorisering, flerlagersperceptron och neural matrisfaktorisering, med hjälp av beteendedata från en digital läsplattform för barn. Rekommendationssystem är ett växande område, och många plattformar kan dra nytta av att anpassa användarupplevelsen utifrån individen och förenkla användningen av plattformen. För att utföra en mer komplex jämförelse och introducera en ny variant av modellerna, föreslår denna studie ett nytt sätt att representera beteendedata som indata till modellerna, d.v.s. att använda termfrekvens med omvänd dokumentfrekvens (TF- IDF) av förekomster av interaktioner mellan användare och böcker, i motsats till den traditionella binära representationen (positiv om en tidigare interaktion existerar och negativ i annat fall). Prestandan mäts genom att extrahera den senaste boken som lästs för varje användare, och utvärdera hur högt modellerna skulle rangordna den boken i rekommendationer till användaren. För att värdesätta modellerna för plattformen med digitala böcker, så jämförs modellerna också med det befintliga rekommendationssystemet på plattformen. Resultaten tyder på att matrisfaktorisering-modellen presterar bäst utav de tre modellerna när man använder data från barns läsbeteende. På grund av den långa träningstiden och fler hyperparametrar att optimera för de andra två modellerna, kan det dock vara så att de inte har nått en optimal hyperparameterinställning, vilket påverkar jämförelsen mellan de tre moderna modellerna. Denna begränsning diskuteras ytterligare i studien. Alla tre modellerna presterar betydligt bättre än det nuvarande systemet på läsplattformen. Modellerna med den föreslagna representationen av TFIDF-värden visar sig mycket lovande och presterar bättre än den binära representationen i nästan alla numeriska mått för alla modeller. Dessa resultat kan ge skäl för framtida forskning av fler sätt att representera beteendedata som indata till denna typ av modeller.

非負値行列因子分解(NMF)は，全要素が非負であるデータ行列に対する行列分解法である．本論文では，実在するデータ行列に頻繁に見られる特徴や解釈容易性の向上を考慮に入れ，探索的にデータ分析を行うためのNMFの拡張について論じている．具体的には，零過剰行列や外れ値を含む行列を扱うための確率分布やダイバージェンス，さらには分解結果である因子行列の数や因子行列への直交制約について述べている． / Nonnegative matrix factorization (NMF) is a matrix decomposition technique to analyze nonnegative data matrices, which are matrices of which all elements are nonnegative. In this thesis, we discuss extensions of NMF for exploratory data analysis considering common features of a real nonnegative data matrix and an easy interpretation. In particular, we discuss probability distributions and divergences for zero-inflated data matrix and data matrix with outliers, two-factor vs. three-factor, and orthogonal constraint to factor matrices. / 博士(文化情報学) / Doctor of Culture and Information Science / 同志社大学 / Doshisha University

This work entails efforts to understand effects of solvent choice on paracetamol crystallization. Various techniques have been developed and implemented to study aforementioned. A clear-cut, direct evidence of two-step nucleation mechanism is demonstrated using a bench top Raman spectrometer and a novel method named as OSANO. / Polymorphismus ist die Eigenschaft vieler anorganischer und insbesondere organischer Moleküle, in mehr als einer Struktur zu kristallisieren. Es ist wichtig, die Faktoren zu verstehen, die den Polymorphismus beeinflussen, da er viele physikochemische Eigenschaften wie Stabilität und Löslichkeit beeinflusst. Nahezu 80 % der vermarkteten Medikamente weisen Polymorphismus auf. In dieser Arbeit wurde der Einfluss der Wahl des organischen Lösungsmittels auf den Polymorphismus von Paracetamol untersucht und verschiedene Methoden entwickelt und angewandt, um den Einfluss genauer zu verstehen. Es wurde festgestellt, dass Ethanol viel stärker auf Paracetamol-Kristallisation als Methanol wirkt. Nichtgleichgewichts-Molekulardynamiksimulationen mit periodischer, simulierter Abkühlung (Simulated Annealing) wurden verwendet, um Vorläufer der metastabilen Zwischenprodukte im Kristallisationsprozess zu untersuchen. Es wurde festgestellt, dass die Strukturen der Bausteine der Paracetamol-Kristalle durch geometrische Wechselwirkungen zwischen Lösungsmittel und Paracetamol bestimmt werden. Die statistisch häufigsten Bausteine in der Selbstassemblierung definieren die finale Kristallstruktur. Ein speziell angefertigter akustischer Levitator hat die Proben zuverlässig gehalten, wodurch die Untersuchung des Einflusses von Lösungsmitteln ermöglicht, heterogene Keimbildung abgeschwächt und andere Umgebungsfaktoren stabilisiert wurden. Die Kristallisation wurde in diesem Aufbau mit zeitaufgelöster In-situ-Raman-Spektroskopie verfolgt und mit einer neuen Zielfunktion basierenden Methode der nichtnegativen Matrixfaktorisierung (NMF) analysiert. Orthogonale Zeitrafferfotografie wurde in Verbindung mit NMF verwendet, um eindeutige und genaue Faktoren zu erhalten, die sich auf die Spektren und Konzentrationen verschiedener Anteile der Paracetamol-Kristallisation beziehen, die als latente Komponenten in den unbehandelten Daten vorhanden sind. / Polymorphism is the property exhibited by many inorganic and organic molecules to crystallize in more than one crystal structure. There is a strong need for understanding the influencing factors on polymorphism, as it is responsible for differences in many physicochemical properties such as stability and solubility. Nearly 80 % of marketed drugs exhibit polymorphism. In this work, we took the model system of paracetamol to investigate the influence of solvent choice on its polymorphism. Different methods were developed and employed to understand the influence of small organic solvents on the crystallization of paracetamol. Non-equilibrium molecular dynamics simulations with periodic simulated annealing were used as a tool to probe the nature of precursors of the metastable intermediates occurring in the crystallization process. Using this method, it was found that the structures of the building blocks of crystals of paracetamol is governed by solvent-solute interactions. In situ Raman spectroscopy was used with a custom-made acoustic levitator to follow crystallization. This set-up is a reliable method for investigating solvent influence, attenuating heterogeneous nucleation and stabilizing other environmental factors. It was established that as a solvent, ethanol is much stronger than methanol in its effect of driving paracetamol solutions to their crystal form. The time-resolved Raman spectroscopy crystallization data was processed using a newly developed objective function based non-negative matrix factorization method (NMF). An orthogonal time-lapse photography was used in conjunction with NMF to get unique and accurate factors that pertain to the spectra and concentrations of different moieties of paracetamol crystallization existing as latent components in the untreated data.

Die Dissertation untersucht das Potenzial von Machine-Learning-Methoden zur Verwaltung von Schwanzrisiken in nicht-stationären und hochdimensionalen Umgebungen. Dazu vergleichen wir auf robuste Weise datenabhängige Ansätze aus parametrischer oder nicht-parametrischer Statistik mit datenadaptiven Methoden. Da datengetriebene Methoden reproduzierbar sein müssen, um Vertrauen und Transparenz zu gewährleisten, schlagen wir zunächst eine neue Plattform namens Quantinar vor, die einen neuen Standard für wissenschaftliche Veröffentlichungen setzen soll. Im zweiten Kapitel werden parametrische, lokale parametrische und nicht-parametrische Methoden verglichen, um eine dynamische Handelsstrategie für den Schutz vor Schwanzrisiken in Bitcoin zu entwickeln. Das dritte Kapitel präsentiert die Portfolio-Allokationsmethode NMFRB, die durch eine Dimensionsreduktionstechnik hohe Dimensionen bewältigt. Im Vergleich zu klassischen Machine-Learning-Methoden zeigt NMFRB in zwei Universen überlegene risikobereinigte Renditen. Das letzte Kapitel kombiniert bisherige Ansätze zu einer Schwanzrisikoschutzstrategie für Portfolios. Die erweiterte NMFRB berücksichtigt Schwanzrisikomaße, behandelt nicht-lineare Beziehungen zwischen Vermögenswerten während Schwanzereignissen und entwickelt eine dynamische Schwanzrisikoschutzstrategie unter Berücksichtigung der Nicht-Stationarität der Vermögensrenditen. Die vorgestellte Strategie reduziert erfolgreich große Drawdowns und übertrifft andere moderne Schwanzrisikoschutzstrategien wie die Value-at-Risk-Spread-Strategie. Die Ergebnisse werden durch verschiedene Data-Snooping-Tests überprüft. / This dissertation shows the potential of machine learning methods for managing tail risk in a non-stationary and high-dimensional setting. For this, we compare in a robust manner data-dependent approaches from parametric or non-parametric statistics with data-adaptive methods. As these methods need to be reproducible to ensure trust and transparency, we start by proposing a new platform called Quantinar, which aims to set a new standard for academic publications. In the second chapter, we dive into the core subject of this thesis which compares various parametric, local parametric, and non-parametric methods to create a dynamic trading strategy that protects against tail risk in Bitcoin cryptocurrency. In the third chapter, we propose a new portfolio allocation method, called NMFRB, that deals with high dimensions thanks to a dimension reduction technique, convex Non-negative Matrix Factorization. This technique allows us to find latent interpretable portfolios that are diversified out-of-sample. We show in two universes that the proposed method outperforms other classical machine learning-based methods such as Hierarchical Risk Parity (HRP) concerning risk-adjusted returns. We also test the robustness of our results via Monte Carlo simulation. Finally, the last chapter combines our previous approaches to develop a tail-risk protection strategy for portfolios: we extend the NMFRB to tail-risk measures, we address the non-linear relationships between assets during tail events by developing a specific non-linear latent factor model, finally, we develop a dynamic tail risk protection strategy that deals with the non-stationarity of asset returns using classical econometrics models. We show that our strategy is successful at reducing large drawdowns and outperforms other modern tail-risk protection strategies such as the Value-at-Risk-spread strategy. We verify our findings by performing various data snooping tests.

A substantial number of works have aimed at modeling the receptive field properties of the primary visual cortex (V1). Their evaluation criterion is usually the similarity of the model response properties to the recorded responses from biological organisms. However, as several algorithms were able to demonstrate some degree of similarity to biological data based on the existing criteria, we focus on the robustness against loss of information in the form of occlusions as an additional constraint for better understanding the algorithmic level of early vision in the brain. We try to investigate the influence of competition mechanisms on the robustness. Therefore, we compared four methods employing different competition mechanisms, namely, independent component analysis, non-negative matrix factorization with sparseness constraint, predictive coding/biased competition, and a Hebbian neural network with lateral inhibitory connections. Each of those methods is known to be capable of developing receptive fields comparable to those of V1 simple-cells. Since measuring the robustness of methods having simple-cell like receptive fields against occlusion is difficult, we measure the robustness using the classification accuracy on the MNIST hand written digit dataset. For this we trained all methods on the training set of the MNIST hand written digits dataset and tested them on a MNIST test set with different levels of occlusions. We observe that methods which employ competitive mechanisms have higher robustness against loss of information. Also the kind of the competition mechanisms plays an important role in robustness. Global feedback inhibition as employed in predictive coding/biased competition has an advantage compared to local lateral inhibition learned by an anti-Hebb rule.