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.

SPATIAL CLUSTERING ALGORITHMS-AN OVERVIEW.

An Overview of known spatial clustering algorithms The space of interest can be the two-dimensional abstraction of the surface of the earth or a man-made space like the layout of a VLSI design, a volume containing a model of the human brain, or another 3d-space representing the arrangement of chains of protein molecules. The data consists of geometric information and can be either discrete or continuous. The explicit location and extension of spatial objects define implicit relations of spatial neighborhood (such as topological, distance and direction relations) which are used by spatial data mining algorithms. Therefore, spatial data mining algorithms are required for spatial characterization and spatial trend analysis. Spatial data mining or knowledge discovery in spatial databases differs from regular data mining in analogous with the differences between non-spatial data and spatial data. The attributes of a spatial object stored in a database may be affected by the attributes of the spatial neighbors of that object. In addition, spatial location, and implicit information about the location of an object, may be exactly the information that can be extracted through spatial data mining

Detection of Flock Movement in Spatio-Temporal Database using Clustering Techniques - an Experience

In this paper, moving flock patterns are mined from spatio- temporal datasets by incorporating a clustering algorithm. A flock is defined as the set of data that move together for a certain continuous amount of time. Finding out moving flock patterns using clustering algorithms is a potential method to find out frequent patterns of movement in large trajectory datasets. In this approach, SPatial clusteRing algoRithm thrOugh sWarm intelligence (SPARROW) is the clustering algorithm used. The advantage of using SPARROW algorithm is that it can effectively discover clusters of widely varying sizes and shapes from large databases. Variations of the proposed method are addressed and also the experimental results show that the problem of scalability and duplicate pattern formation is addressed. This method also reduces the number of patterns produced

Spectral clustering independent component analysis for tissue classification from brain MRI

A spectral angle based feature extraction method, Spectral Clustering Independent Component Analysis (SC-ICA), is proposed in this work to improve the brain tissue classification from Magnetic Resonance Images (MRI). SC-ICA provides equal priority to global and local features; thereby it tries to resolve the inefficiency of conventional approaches in abnormal tissue extraction. First, input multispectral MRI is divided into different clusters by a spectral distance based clustering. Then, Independent Component Analysis (ICA) is applied on the clustered data, in conjunction with Support Vector Machines (SVM) for brain tissue analysis. Normal and abnormal datasets, consisting of real and synthetic T1-weighted, T2-weighted and proton density/fluid-attenuated inversion recovery images, were used to evaluate the performance of the new method. Comparative analysis with ICA based SVM and other conventional classifiers established the stability and efficiency of SC-ICA based classification, especially in reproduction of small abnormalities. Clinical abnormal case analysis demonstrated it through the highest Tanimoto Index/accuracy values, 0.75/98.8%, observed against ICA based SVM results, 0.17/96.1%, for reproduced lesions. Experimental results recommend the proposed method as a promising approach in clinical and pathological studies of brain diseases

Small World Folksonomies: Clustering in Tri-Partite Hypergraphs

Many recent Web 2.0 resource sharing applications can be subsumed under the "folksonomy" moniker. Regardless of the type of resource shared, all of these share a common structure describing the assignment of tags to resources by users. In this report, we generalize the notions of clustering and characteristic path length which play a major role in the current research on networks, where they are used to describe the small-world effects on many observable network datasets. To that end, we show that the notion of clustering has two facets which are not equivalent in the generalized setting. The new measures are evaluated on two large-scale folksonomy datasets from resource sharing systems on the web.

Protein-Expressions-Analysen zur Etablierung prädiktiver Biomarker bei der Behandlung fortgeschrittener Tumoren der Mamma und der Prostata

Aktuelle Entwicklungen auf dem Gebiet der zielgerichteten Therapie zur Behandlung maligner Erkrankungen erfordern neuartige Verfahren zur Diagnostik und Selektion geeigneter Patienten. So ist das Ziel der vorliegenden Arbeit die Identifizierung neuer Zielmoleküle, die die Vorhersage eines Therapieerfolges mit targeted drugs ermöglichen. Besondere Aufmerksamkeit gilt dem humanisierten monoklonalen Antikörper Trastuzumab (Herceptin), der zur Therapie Her-2 überexprimierender, metastasierter Mammakarzinome eingesetzt wird. Jüngste Erkenntnisse lassen eine Anwendung dieses Medikamentes in der Behandlung des Hormon-unabhängigen Prostatakarzinoms möglich erscheinen. Therapie-beeinflussende Faktoren werden in der dem Rezeptor nachgeschalteten Signaltransduktion oder Veränderungen des Rezeptors selbst vermutet. Mittels Immunhistochemie wurden die Expressions- und Aktivierungsniveaus verschiedener Proteine der Her-2-assoziierten Signaltransduktion ermittelt; insgesamt wurden 37 molekulare Marker untersucht. In Formalin fixierte und in Paraffin eingebettete korrespondierende Normal- und Tumorgewebe von 118 Mammakarzinom-Patientinnen sowie 78 Patienten mit Prostatakarzinom wurden in TMAs zusammengefasst. Die in Zusammenarbeit mit erfahrenen Pathologen ermittelten Ergebnisse dienten u.a. als Grundlage für zweidimensionales, unsupervised hierarchisches clustering. Ergebnis dieser Analysen war für beide untersuchten Tumorentitäten die Möglichkeit einer Subklassifizierung der untersuchten Populationen nach molekularen Eigenschaften. Hierbei zeigten sich jeweils neue Möglichkeiten zur Anwendung zielgerichteter Therapien, deren Effektivität Inhalt weiterführender Studien sein könnte. Zusätzlich wurden an insgesamt 43 Frischgeweben die möglichen Folgen des sog. shedding untersucht. Western Blot-basierte Untersuchungen zeigten hierbei die Möglichkeit der Selektion von Patienten aufgrund falsch-positiver Befunde in der derzeit als Standard geltenden Diagnostik. Zusätzlich konnte durch Vergleich mit einer Herceptin-sensitiven Zelllinie ein möglicher Zusammenhang eines Therapieerfolges mit dem Phosphorylierungs-/ Aktivierungszustand des Rezeptors ermittelt werden. Fehlende klinische Daten zum Verlauf der Erkrankung und Therapie der untersuchten Patienten lassen keine Aussagen über die tatsächliche Relevanz der ermittelten Befunde zu. Dennoch verdeutlichen die erhaltenen Resultate eindrucksvoll die Komplexität der molekularen Vorgänge, die zu einem Krebsgeschehen führen und damit Auswirkungen auf die Wirksamkeit von targeted drugs haben können. Entwicklungen auf dem Gebiet der zielgerichteten Therapie erfordern Verbesserungen auf dem Gebiet der Diagnostik, die die sichere Selektion geeigneter Patienten erlauben. Die Zukunft der personalisierten, zielgerichteten Behandlung von Tumorerkrankungen wird verstärkt von molekularen Markerprofilen hnlich den hier vorgestellten Daten beeinflusst werden.

Content aggregation on knowledge bases using graph clustering

Recently, research projects such as PADLR and SWAP have developed tools like Edutella or Bibster, which are targeted at establishing peer-to-peer knowledge management (P2PKM) systems. In such a system, it is necessary to obtain provide brief semantic descriptions of peers, so that routing algorithms or matchmaking processes can make decisions about which communities peers should belong to, or to which peers a given query should be forwarded. This paper proposes the use of graph clustering techniques on knowledge bases for that purpose. Using this clustering, we can show that our strategy requires up to 58% fewer queries than the baselines to yield full recall in a bibliographic P2PKM scenario.

Capturing Emergent Semantics from Social Annotation Systems

The ongoing growth of the World Wide Web, catalyzed by the increasing possibility of ubiquitous access via a variety of devices, continues to strengthen its role as our prevalent information and commmunication medium. However, although tools like search engines facilitate retrieval, the task of finally making sense of Web content is still often left to human interpretation. The vision of supporting both humans and machines in such knowledge-based activities led to the development of different systems which allow to structure Web resources by metadata annotations. Interestingly, two major approaches which gained a considerable amount of attention are addressing the problem from nearly opposite directions: On the one hand, the idea of the Semantic Web suggests to formalize the knowledge within a particular domain by means of the "top-down" approach of defining ontologies. On the other hand, Social Annotation Systems as part of the so-called Web 2.0 movement implement a "bottom-up" style of categorization using arbitrary keywords. Experience as well as research in the characteristics of both systems has shown that their strengths and weaknesses seem to be inverse: While Social Annotation suffers from problems like, e. g., ambiguity or lack or precision, ontologies were especially designed to eliminate those. On the contrary, the latter suffer from a knowledge acquisition bottleneck, which is successfully overcome by the large user populations of Social Annotation Systems. Instead of being regarded as competing paradigms, the obvious potential synergies from a combination of both motivated approaches to "bridge the gap" between them. These were fostered by the evidence of emergent semantics, i. e., the self-organized evolution of implicit conceptual structures, within Social Annotation data. While several techniques to exploit the emergent patterns were proposed, a systematic analysis - especially regarding paradigms from the field of ontology learning - is still largely missing. This also includes a deeper understanding of the circumstances which affect the evolution processes. This work aims to address this gap by providing an in-depth study of methods and influencing factors to capture emergent semantics from Social Annotation Systems. We focus hereby on the acquisition of lexical semantics from the underlying networks of keywords, users and resources. Structured along different ontology learning tasks, we use a methodology of semantic grounding to characterize and evaluate the semantic relations captured by different methods. In all cases, our studies are based on datasets from several Social Annotation Systems. Specifically, we first analyze semantic relatedness among keywords, and identify measures which detect different notions of relatedness. These constitute the input of concept learning algorithms, which focus then on the discovery of synonymous and ambiguous keywords. Hereby, we assess the usefulness of various clustering techniques. As a prerequisite to induce hierarchical relationships, our next step is to study measures which quantify the level of generality of a particular keyword. We find that comparatively simple measures can approximate the generality information encoded in reference taxonomies. These insights are used to inform the final task, namely the creation of concept hierarchies. For this purpose, generality-based algorithms exhibit advantages compared to clustering approaches. In order to complement the identification of suitable methods to capture semantic structures, we analyze as a next step several factors which influence their emergence. Empirical evidence is provided that the amount of available data plays a crucial role for determining keyword meanings. From a different perspective, we examine pragmatic aspects by considering different annotation patterns among users. Based on a broad distinction between "categorizers" and "describers", we find that the latter produce more accurate results. This suggests a causal link between pragmatic and semantic aspects of keyword annotation. As a special kind of usage pattern, we then have a look at system abuse and spam. While observing a mixed picture, we suggest that an individual decision should be taken instead of disregarding spammers as a matter of principle. Finally, we discuss a set of applications which operationalize the results of our studies for enhancing both Social Annotation and semantic systems. These comprise on the one hand tools which foster the emergence of semantics, and on the one hand applications which exploit the socially induced relations to improve, e. g., searching, browsing, or user profiling facilities. In summary, the contributions of this work highlight viable methods and crucial aspects for designing enhanced knowledge-based services of a Social Semantic Web.

Hierarchical Object Recognition Using Libraries of Parameterized Model Sub-Parts

This thesis describes the development of a model-based vision system that exploits hierarchies of both object structure and object scale. The focus of the research is to use these hierarchies to achieve robust recognition based on effective organization and indexing schemes for model libraries. The goal of the system is to recognize parameterized instances of non-rigid model objects contained in a large knowledge base despite the presence of noise and occlusion. Robustness is achieved by developing a system that can recognize viewed objects that are scaled or mirror-image instances of the known models or that contain components sub-parts with different relative scaling, rotation, or translation than in models. The approach taken in this thesis is to develop an object shape representation that incorporates a component sub-part hierarchy- to allow for efficient and correct indexing into an automatically generated model library as well as for relative parameterization among sub-parts, and a scale hierarchy- to allow for a general to specific recognition procedure. After analysis of the issues and inherent tradeoffs in the recognition process, a system is implemented using a representation based on significant contour curvature changes and a recognition engine based on geometric constraints of feature properties. Examples of the system's performance are given, followed by an analysis of the results. In conclusion, the system's benefits and limitations are presented.

Formal Multilevel Hierarchical Verification of Synchronous MOS Circuits

I have designed and implemented a system for the multilevel verification of synchronous MOS VLSI circuits. The system, called Silica Pithecus, accepts the schematic of an MOS circuit and a specification of the circuit's intended digital behavior. Silica Pithecus determines if the circuit meets its specification. If the circuit fails to meet its specification Silica Pithecus returns to the designer the reason for the failure. Unlike earlier verifiers which modelled primitives (e.g., transistors) as unidirectional digital devices, Silica Pithecus models primitives more realistically. Transistors are modelled as bidirectional devices of varying resistances, and nodes are modelled as capacitors. Silica Pithecus operates hierarchically, interactively, and incrementally. Major contributions of this research include a formal understanding of the relationship between different behavioral descriptions (e.g., signal, boolean, and arithmetic descriptions) of the same device, and a formalization of the relationship between the structure, behavior, and context of device. Given these formal structures my methods find sufficient conditions on the inputs of circuits which guarantee the correct operation of the circuit in the desired descriptive domain. These methods are algorithmic and complete. They also handle complex phenomena such as races and charge sharing. Informal notions such as races and hazards are shown to be derivable from the correctness conditions used by my methods.

A Hierarchical Cache Coherent Protocol

As the number of processors in distributed-memory multiprocessors grows, efficiently supporting a shared-memory programming model becomes difficult. We have designed the Protocol for Hierarchical Directories (PHD) to allow shared-memory support for systems containing massive numbers of processors. PHD eliminates bandwidth problems by using a scalable network, decreases hot-spots by not relying on a single point to distribute blocks, and uses a scalable amount of space for its directories. PHD provides a shared-memory model by synthesizing a global shared memory from the local memories of processors. PHD supports sequentially consistent read, write, and test- and-set operations. This thesis also introduces a method of describing locality for hierarchical protocols and employs this method in the derivation of an abstract model of the protocol behavior. An embedded model, based on the work of Johnson[ISCA19], describes the protocol behavior when mapped to a k-ary n-cube. The thesis uses these two models to study the average height in the hierarchy that operations reach, the longest path messages travel, the number of messages that operations generate, the inter-transaction issue time, and the protocol overhead for different locality parameters, degrees of multithreading, and machine sizes. We determine that multithreading is only useful for approximately two to four threads; any additional interleaving does not decrease the overall latency. For small machines and high locality applications, this limitation is due mainly to the length of the running threads. For large machines with medium to low locality, this limitation is due mainly to the protocol overhead being too large. Our study using the embedded model shows that in situations where the run length between references to shared memory is at least an order of magnitude longer than the time to process a single state transition in the protocol, applications exhibit good performance. If separate controllers for processing protocol requests are included, the protocol scales to 32k processor machines as long as the application exhibits hierarchical locality: at least 22% of the global references must be able to be satisfied locally; at most 35% of the global references are allowed to reach the top level of the hierarchy.

A Detailed Look at Scale and Translation Invariance in a Hierarchical Neural Model of Visual Object Recognition

The HMAX model has recently been proposed by Riesenhuber & Poggio as a hierarchical model of position- and size-invariant object recognition in visual cortex. It has also turned out to model successfully a number of other properties of the ventral visual stream (the visual pathway thought to be crucial for object recognition in cortex), and particularly of (view-tuned) neurons in macaque inferotemporal cortex, the brain area at the top of the ventral stream. The original modeling study only used ``paperclip'' stimuli, as in the corresponding physiology experiment, and did not explore systematically how model units' invariance properties depended on model parameters. In this study, we aimed at a deeper understanding of the inner workings of HMAX and its performance for various parameter settings and ``natural'' stimulus classes. We examined HMAX responses for different stimulus sizes and positions systematically and found a dependence of model units' responses on stimulus position for which a quantitative description is offered. Interestingly, we find that scale invariance properties of hierarchical neural models are not independent of stimulus class, as opposed to translation invariance, even though both are affine transformations within the image plane.

Hierarchical Mixtures of Experts and the EM Algorithm

We present a tree-structured architecture for supervised learning. The statistical model underlying the architecture is a hierarchical mixture model in which both the mixture coefficients and the mixture components are generalized linear models (GLIM's). Learning is treated as a maximum likelihood problem; in particular, we present an Expectation-Maximization (EM) algorithm for adjusting the parameters of the architecture. We also develop an on-line learning algorithm in which the parameters are updated incrementally. Comparative simulation results are presented in the robot dynamics domain.

Integració a la plataforma webDTI de tècniques de clustering de fibres i visualització de clústers

Estudi, disseny i implementació de diferents tècniques d’agrupament de fibres (clustering) per tal d’integrar a la plataforma DTIWeb diferents algorismes de clustering i tècniques de visualització de clústers de fibres de forma que faciliti la interpretació de dades de DTI als especialistes

Global and dominance with concentric hierarchical stimuli and orientation classification task. 'Dominancia global y local con est??mulos conc??ntricos y tareas de clasificaci??n de la orientaci??n'

Resumen tomado de la publicaci??n. Resumen tambi??n en ingl??s