Search

We present a new mixed precision algorithm to compute low-rank matrix and tensor approximations, a fundamental task in numerous applications in scientific computing and data analysis. Our algorithm is reminiscent of the iterative refinement framework for linear systems: we first compute a low-rank approximation in low precision and then refine its accuracy by iteratively updating it. We carry out an error analysis of our algorithm which proves that we can reach a high accuracy while performing most of the operations in low precision. We measure the computational cost of the algorithm, which depends on the numerical rank of the input (matrix or tensor) as well as the speed ratio between low and high precision arithmetic. We identify two situations where our method has a strong potential : when the hardware provides fast low precision matrix multiplyaccumulate units, and when the numerical rank of the input is small at low accuracy levels. We confirm experimentally the potential of our algorithm for computing various low-rank matrix and tensor decompositions such as SVD, QR, Tucker, hierarchical Tucker, and tensor-train.

We propose a novel method called Adjacency Propagation Algorithm (APAL) which considers the notion that the adjacent vertices are the best candidates for detecting overlapping communities in an undirected, unweighted, nontrivial graph. This is a compact algorithm with a single threshold parameter used to filter the detected communities according to their intraconnectivity property . In this study, APAL was tested rigorously using synthetic generators, such as the widely accepted LFR benchmark, as well as real data sets of yeast and human protein interactions networks . It was compared against the foremost algorithms in the field; the Clique Percolation Method (CPM), Community Overlap Propagation Algorithm (COPRA) and Neighbourhood-Inflated Seed Expansion (NISE). The results show that APAL outperforms its competitors for networks with increases in the number of memberships of the overlapping vertices. Such conditions are often found in biological networks, where a particular protein subunit may form part of several complexes. We believe that this shows the value of the implementation of APAL for protein interaction and other biological networks.

Tartarus is a commonly used benchmark problem for genetic programming. However, it has never been fully explored for its difficulty tuning property. Using the data from a previous study in which we have executed millions of Tartarus instances, we contribute to the literature with an equation to estimate their difficulty. Our approach uses four metrics that are embedded into the equation. These metrics are related to the number of clusters and clusters sizes, the distances of boxes to the edges of the board grid, the number of boxes around the agent, and the minimum number of actions for the agent to reach the largest cluster. The coefficients of these metrics have been fit to the data using the general linear model and a mean residual error of ~0.1 has been achieved. This is the first study that can estimate the difficulty of a Tartarus board without modifying the problem in any way., Tartarus genetik programlamada sıkça kullanılan bir kıyaslama problemidir. Fakat zorluk ayarı özelliği henüz tam olarak araştırılmamıştır. Literatüre milyonlarca Tartarus örneği çalıştırdığımız önceki bir çalışmanın verilerini kullanarak zorluklarını tahmin edebilen bir denklemle katkıda bulunuyoruz. Yaklaşımımız denklemin içinde yer alan dört yeni metrik kullanıyor. Bu metrikler küme sayıları ve büyüklüklerine, kutuların kenarlardan uzaklığına, yazılım etmeninin etrafındaki kutuların sayısına ve etmenin en büyük kümeye varması için gereken hareket sayısına bağlıdır. Metriklerin katsayıları veriye genel doğrusal model ile uyarlanmış ve ortalama ~0.1 kadar bir hata başarısına ulaşılmıştır. Bu çalışma Tartarus probleminde bir değişiklik yapmadan problemin zorluğunu tahmin edebilen ilk çalışmadır.

Procedural content generation (PCG) methods are commonly employed in computer games, simulations, and other related industries. While these methods are used for levels, terrains, stories and missions, their usage for procedural city generation is relatively rare because cities are heterogeneous structures with different components such as roads, layouts and buildings that depend on and affect each other. Additionally, ancient cities grew organically to areas that are safe and to those that provide food and water. This resulted in cities that do not have apparent regular patterns, such as rectangular building blocks. We propose an approach that uses cellular automata (CA) that generates clusters of areas. The CA is repeated for each cluster to hierarchically create different levels of the city. This procedure creates an organic city layout with fractal properties. The layout specifies the building blocks, main roads, and foliage. We also present a set of methods that can transform this layout into a three-dimensional model of the city. The results are promising; cities can be created in under a minute with minimal required input, and the resulting virtual city looks organic, rather than an algorithmic layout that has repeating patterns.

The Tartarus Problem is one of the candidate benchmark problems in evolutionary algorithms. We take advantage of the graphical processing unit (GPU) to improve the results of the software agents that use finite state machines (FSMs) for this benchmark. While doing so we also contribute to the study of the problem on several grounds. Similar to existing studies we use genetic algorithms to evolve FSMs, but unlike most of them we use adaptive operators for controlling the parameters of the algorithm. We show that the actual number of valid boards is not 297,040, but 74,760, because the agent is indifferent to the rotations of the board. We also show that the agent can only come across 383 different combinations, rather than 6561 that is used in the current literature. A final contribution is that we report the first true scores for the agents by testing them with all available 74,760 boards. Our best solution has a mean score of 8.5379 on all boards.

Speech is the most natural way of expressing ourselves as humans. It is only natural then to extend this communication medium to computer applications. We define speech emotion recognition (SER) systems as a collection of methodologies that process and classify speech signals to detect the embedded emotions. SER is not a new field, it has been around for over two decades, and has regained attention thanks to the recent advancements. These novel studies make use of the advances in all fields of computing and technology, making it necessary to have an update on the current methodologies and techniques that make SER possible. We have identified and discussed distinct areas of SER, provided a detailed survey of current literature of each, and also listed the current challenges.

New research fields and applications on human computer interaction will emerge based on the recognition of emotions on faces. With such aim, our study evaluates the features extracted from faces to recognize emotions. To increase the success rate of these features, we have run several tests to demonstrate how age and gender affect the results. The artificial neural networks were trained by the apparent regions on the face such as eyes, eyebrows, nose, mouth, and jawline and then the networks are tested with different age and gender groups. According to the results, faces of older people have a lower performance rate of emotion recognition. Then, age and gender based groups are created manually, and we show that performance rates of facial emotion recognition have increased for the networks that are trained using these particular groups.

WOS: 000436416800034 #REF! Aim: To examine theeffects on the brain of 2-month treatment withamethylphenidate extended-release formulation (OROS-MPH) using [Tc-(99m)] TRODAT-1SPECT in a sample of treatment-naive adolescents with Attention Deficit/Hyperactivity Disorder (ADHD). In addition, to assess whether risk alleles (homozygosity for 10-repeat allele at the DAT1 gene were associated with alterations in striatal DAT availability. Methods: Twenty adolescents with ADHD underwent brain single-photon emission computed tomography (SPECT) scans with [Tc-(99m)] TRODAT-1 at baseline and two months after starting OROS-MPH treatment with dosages up to 1 mg/kg/day. Severity of illness was estimated using the Clinical Global Impression Scale (CGI-S) and DuPaul ADHD Rating Scale-Clinician version (ARS) before treatment,1 month and 2 months after initiating OROS-MPH treatment. Results: Decreased DAT availability was found in both the right caudate (pretreatment DAT binding: 224.76 +/- 33.77, post-treatment DAT binding: 208.86 +/- 28.75, p = 0.02) and right putamen (pre-treatment DAT binding: 314.41 +/- 55.24, post-treatment DAT binding: 285.66 +/- 39.20, p = 0.05) in adolescents with ADHD receiving OROS-MPH treatment. Adolescents with ADHD who showed a robust response to OROS-MPH (n = 7) had significantly greater reduction of DAT density in the right putamen than adolescents who showed less robust response to OROS-MPH (n = 13) (p = 0.02). However, between-group differences by treatment responses were not related with DAT density in the right caudate. Risk alleles (homozygosity for the 10-repeat allele of DAT1 gene) in the DAT1 gene were not associated with alterations in striatal DAT availability. Conclusion: Two months of OROS-MPH treatment decreased DAT availability in both the right caudate and putamen. Adolescents with ADHD who showed a robust response to OROS-MPH had greater reduction of DAT density in the right putamen. However,our findings did not support an association between homozygosity for a 10-repeat allele in the DAT1 gene and DAT density, assessedusing[Tc-(99m)] TRODAT-1SPECT. Dokuz Eylul University Research CommitteeDokuz Eylul University [2008132]; Fogarty MH/DD program; Institute of Nuclear Energy Research (INER-Taipei, Taiwan) This study was partially supported by Dokuz Eylul University Research Committee (2008132), Fogarty MH/DD program and the Institute of Nuclear Energy Research (INER-Taipei, Taiwan).

We propose two contributions with novel approaches to fMRI activation analysis. The first is to apply confidence intervals to locate activations in real-time, and second is a new metric based on robust regression of fMRI signals. These contributions are implemented in our four proposed methods; Instantaneous Activation Method (TAM), Instantaneous Activation Method with Past Blocks (TAMP) for real-time analysis, Task Robust Regression Distance Method (TRRD) for the new metric with robust regression and Instantaneous Robust Regression Distance Method (IRRD) for both contributions. For comparison, a statistical offline method called Task Activation Method (TAM) and a correlation analysis method are also implemented. The methods are initially evaluated with synthetic data generated using two different approaches; first using varying hemodynamic response function signals to simulate a wide range of stimuli responses, along with a Gaussian white noise, and second using no activity state data of a real fMRI experiment, which removes the need to generate noise. The methods are also tested with real fMRI experiments and compared with the results obtained by the widely used SPM tool. The results show that instantaneous methods reveal activations that are lost statistically in an offline analysis. They also reveal further improvements by robust fitting application, which minimizes the outlier effect. TRRD has an area under the ROC curve of 0,7127 for very noisy synthetic images, is reaching up to 0,9608 as the noise decreases, while the instantaneous score is in the range of 0,6124 to 0,8019 in the same noise levels. (C) 2017 Elsevier B.V. All rights reserved.

Genetic algorithms can be used to evolve finite state machines for problems that require a large number of states and transitions. Tartarus problem is such a problem in which the purpose is to push the boxes towards the walls of a six by six grid using a bulldozer that can only sense its 8-neighbourhood. The bulldozer can rotate left, right, or move forward, each taking a single move out of its initial 80 moves. The result is scored by the number of boxes that are against a wall when the bulldozer is out of moves. Several approaches have been proposed, with genetic algorithms being the most common. We are proposing a representation of the problem using varying number of states and adaptive modification of the mutation parameter to decrease the probability of the population getting stuck at a local minima. Our results show improvement over the application of the genetic algorithm without parameter modification and dependency on the number states and the size of the population.

Functional magnetic resonance imaging aims to correlate brain functions with the regions of the brain. To find out the region responsible for a brain function, the brain should be exposed to stimuli related to the function, and the changes in time should be monitored. To acquire images that change over time makes it more complex and difficult to extract deductions from functional images than structural images. To get the design right, it is necessary to ask the right questions, to give the correct stimuli at the right times, and to analyze the obtained data according to the design of the experiment. As a result, the design of the experiment affects the whole process. This paper presents a wide view of the complete process in the design, execution and the analysis of fMRI experiments, along with the related software and hardware requirements, as well as mentioning the challenges involved.

Three Dimensional magnetic resonance imaging (3D-MRI) has been utilized to classify patients with neuroanatomical abnormalities apart from healthy controls (HCs). The studies on the diagnosis of Bipolar Disorder (BD) focuses also on the unaffected relatives of BD patients in order to examine the heritable resistance factors associated with the disorder. Hence, the comparison of Healthy Siblings of Bipolar Disorder patients (HSBDs) and HCs is also required owing to the high heritability of BD. In this paper, the classification of 27HSBDs from 38 HCs has been studied by using 3D-MRI and Computer-Aided Detection (CAD). The pre-processing of 3D-MRI data is performed by taking advantage of Voxel-Based Morphometry (VBM) and the structural deformations in the Gray Matter (GM) and White Matter (WM) are obtained by using a general linear model. The model is configured by using a two sample t-test technique and Total Intracranial Volume (TIV) as a covariate. The altered voxels between data groups are considered as Voxel of Interests (VOIs) and the 3D masks are generated for GM and WM tissue probability maps. The Relief-F algorithm is utilized to rank the features and a Fisher Criterion (FC) method is considered to determine the number of top-ranked discriminative features. The performances of Support Vector Machines (SVM) and the Naive Bayes (NB) algorithms are compared on the classification of HSBD and HC. The experiments are performed for GM-only, WM-only, and their combinations. The experimental results indicate that the changes between the brain regions of HSBD and HC might provide information on the heritable factors associated with the BD. Additionally, it is concluded that using the combination of GM and WM tissue probability map provides better results than considering them, separately. Finally, it is obtained that the classification accuracy of SVM on HSBD and HC comparison is better than that of NB.

International audience; It is common to accelerate the boundary element method by compression methods (FMM, H-Matrix / ACA) that enable a more accurate solution or a solution in higher frequency. In this work, we present a compression method based on a transformation of the linear system into Tensor-Train format by the quantization technique. The method is applied to a scattering problem by a canonical object with a regular mesh and improves the performance obtained by the previous methods.

Three-Dimensional Magnetic Resonance Imaging (3D-MRI) and Computer-Aided Detection (CAD) have been widely studied in the detection and diagnosis of neuroanatomical abnormalities, including bipolar disorder (BD). Pre-processing of 3D-MRI scans plays an important role in post-processing. In this study, Voxel-Based Morphometry (VBM) is used to compare the morphological differences at the grey matter (GM) and white matter (WM) of BD subjects versus healthy controls (HCs). The effects of using different covariates (i.e. total intracranial volume (TIV), age, sex, and their combinations) on classification of BDs from HCs have been investigated for GM-only, WM-only, and their combination. 3D masks for GM and WM are generated separately by using local differences between BPs and HCs and the two sample t-test method. Principle component analysis based dimensionality reduction and support vector machine with Gaussian kernel are employed for classification of 26 BDs and 38 HCs obtained from Ege University, School of Medicine, Department of Psychiatry. The results indicate that using only TIV as a covariate provides more robust results for BD classification compared to other covariate combinations. Furthermore, the combination of GM and WM improves classification performance. The highest classification accuracies obtained for GM, WM, and their combination are 70.30%, 79.70% and 82.80% respectively.

EgeUn###, Detection of Bipolar Disease (BD), one of the most common neuroanatomical abnormalities, using machine learning algorithms together with Magnetic Resonance Imaging (MRI) data has been widely studied. BD is a highly heritable disease, yet not all siblings tend to have it despite they might have similar genetic and environmental risk factors. In this paper, the classifications of two self-acquired data groups, namely 26BD patients and 38 unrelated Healthy Controls (HCs) as well as 27 Healthy Siblings of BD (BDHSs) and 38HCs are examined. Voxel-Based Morphometry (VBM) is utilized to segment and pre-process the MRI data. In order to obtain the morphological alterations in the Gray Matter (GM) and White Matter (WM) of data groups separately, a general linear model is configured and a two sample t-test based statistical method is used. The obtained differentiated voxels are considered as Voxel of Interests (VOIs) and using VOIs reduces the dimension of the original data into the number of VOIs. The effects of using different covariates (i.e. total intracranial volume (TIV), age, and sex) on classification of the two data groups have been studied for GM-only, WM-only, and their combination. Principle Component Analysis (PCA) is used to reduce the dimension of the extracted VOIs data and Support Vector Machine (SVM) with Gaussian kernel is taken into account as a classifier. The experimental results indicate that among three covariates, TIV provides better results for both data groups and the classification accuracies of the combination of GM and WM maps is higher than that of GM-only and WM- only for both groups. In BD and HC comparison, the highest classification accuracies of 70.3% for GM, 79.7% for WM, and 82.8% for fusion of extracted GM as well as WM are obtained. In BDHS and HC comparison, the highest classification accuracies of 72.3% for GM, 76.9% for WM, and 78.5% for fusion of extracted GM as well as WM are obtained. © 2019 IEEE.

Medical Technologies Congress (TIPTEKNO) -- OCT 03-05, 2019 -- Izmir, TURKEY, Oguz, Kaya/0000-0002-1860-9127, WOS:000516830900117, Three-Dimensional Magnetic Resonance Imaging (3D-MRI) and Computer-Aided Detection (CAD) have been widely studied in the detection of bipolar disorder (BD). in this study, the structural alterations at the grey matter (GM) and white matter (WM) of BD subjects versus healthy controls (HCs) have been compared using Voxel-Based Morphometry (VBM). in order to obtain 3D GM and WM masks, the two sample t-test method and total intracranial volumes of BD and HC as a covariate have been utilized. in addition to analyzing effects of GM and WM tissue maps separately in the detection of BD, impacts of both GM and WM ones are studied by concatenating them in a matrix. The correlation-based feature selection (CFS) feature ranking method is applied to the obtained 3D masks to rank the features, the number of selected top-ranked features are determined using a Fisher criterion (FC) approach, and different classification algorithms are used to classify BD apart from HCs. in this study, 26 BDs and 38 HCs data are used. The experimental results indicate that the classification accuracy of Naive Bayes outperforms the other four classification algorithms used in this study. Additionally, concatenation of GM and WM tissue maps enhances the classification performances of using GM-only and WM-only ones. The classification accuracies obtained for GM, WM, and their concatenation are 72.92%, 78.33%, and 80.00% respectively., Biyomedikal Klinik Muhendisligi Dernegi, Izmir Katip Celebi Univ, Biyomedikal Muhendisligi Bolumu

Non-negative matrix factorization (NMF), the problem of finding two non-negative low-rank factors whose product approximates an input matrix, is a useful tool for many data mining and scientific applications such as topic modeling in text mining and blind source separation in microscopy.In this paper, we focus on scaling algorithms for NMF to very large sparse datasets and massively parallel machines by employing effective algorithms, communication patterns, and partitioning schemes that leverage the sparsity of the input matrix. In the case of machine learning workflow, the computations after SpMM must deal with dense matrices, as Sparse-Dense matrix multiplication will result in a dense matrix. Hence, the partitioning strategy considering only SpMM will result in a huge imbalance in the overall workflow especially on computations after SpMM and in this specific case of NMF on non-negative least squares computations. Towards this, we consider two previous works developed for related problems, one that uses a fine-grained partitioning strategy using a point-to-point communication pattern and on that uses a checkerboard partitioning strategy using a collective-based communication pattern.We show that a combination of the previous approaches balances the demands of the various computations within NMF algorithms and achieves high efficiency and scalability. From the experiments, we could see that our proposed algorithm communicates atleast 4x less than the collective and achieves upto 100x speed up over the baseline FAUN on real world datasets. Our algorithm was experimented in two different super computing platforms and we could scale up to 32000 processors on Bluegene/Q.; La factorisation de matrice non-négative (NMF), le problème de trouver deux facteurs de rang faible non négatifs dont le produit se rapproche d'une matrice d'entrée, est un outil utile pour de nombreuses applications scientifiques et d'exploration de données telles que la modélisation de textes et la séparation de signaux en microscopie.Dans cet article, nous etudions les algorithmes passant à l'échelle pour NMF à de très grands ensembles de données creuses et des machines massivement parallèles en utilisant des algorithmes efficaces, des modèles de communication et des schémas de partitionnement qui exploitent la structure creuse de la matrice.Dans le cadre de cet algorithme, les calculs après SpMM doivent traiter des matrices denses, car la multiplication SpMM produira une matrice dense.Par conséquent, la stratégie de partitionnement ne prenant en compte que SpMM entraînera un déséquilibre énorme dans l'algorithme global, en particulier sur les calculs après SpMM et dans ce cas spécifique de NMF sur les calculs de moindres carrés non négatifs.À cet égard, nous considérons deux travaux antérieurs développés pour des problèmes connexes, l'un utilisant une stratégie de partitionnement de granularité ffine utilisant un modèle de communication ``point-to-point'' et utilisant une stratégie de partitionnement en damier utilisant un modèle de communication collectif.Nous montrons qu'une combinaison des approches précédentes permet d'équilibrer les exigences des divers calculs au sein des algorithmes NMF et permet d'obtenir une efficacité et une évolutivité élevées. À partir des expériences, nous avons constaté que notre algorithme proposé communique au moins4x moins que le collectif et atteint jusqu'à 100 fois la vitesse de base sur les jeux de données réels. Notre algorithme a été expérimenté sur deux plates-formes superinformatiques différentes et nous avons pu passer à 32 000 processeurs sur Bluegene / Q.

26th IEEE Signal Processing and Communications Applications Conference (SIU) -- MAY 02-05, 2018 -- Izmir, TURKEY, Oguz, Kaya/0000-0002-1860-9127, WOS:000511448500126, Change point analysis is an efficient method for understanding the unexpected behavior of the data used in many different disciplines including medical imaging. It is important to find the instances the activations occur as much as finding the activation areas in the analysis of functional magnetic resonance imaging (fMRI). Change point detection algorithms can be used to find the activation instances. in this study, a regression based point detection method is proposed to find the activation instances in fMRI experiments. The proposed method is applied to a fMRI experiment which includes a motor task. A linear based evaluation method is also proposed. The analyses show that the activations are in accordance with the established methods in the literature., IEEE, Huawei, Aselsan, NETAS, IEEE Turkey Sect, IEEE Signal Proc Soc, IEEE Commun Soc, ViSRATEK, Adresgezgini, Rohde & Schwarz, Integrated Syst & Syst Design, Atilim Univ, Havelsan, Izmir Katip Celebi Univ

Dementia affects the lives of millions of people. Digital recognition and evaluation of the Clock Drawing Test that is used for diagnosis will remove subjective manual evaluation and will increase the awareness and early diagnosis. The proposed system evaluates the clocks drawn to the paper and loaded to it by extracting their graphical features. Besides image processing methods, artificial neural networks are used to detect the numbers. The results of the system are in accordance with the manual evaluation for 35 clocks drawn by healthy subjects.

2017 European Conference on Electrical Engineering and Computer Science, EECS 2017 -- 17 November 2017 through 19 November 2017 -- -- 138047, 2-s2.0-85050962752, Change point analysis is an efficient method for understanding the unexpected behaviour of the data used in many different disciplines. Although the literature contains a variety of change point analysis methods, there are relatively fewer studies that focus on the performance of parameter selection and outlier removal that are applied on real data sets. In this study two methods based on regression and statistical properties are proposed and compared with a method using Bayesian approach to evaluate their performance on the selection of parameters and removal of outliers. The methods are executed using different parameters on the well-log data set with and without outliers that are removed either manually or automatically. The results show that different data sets require different parameters to locate their change points. The proposed methods have intuitive parameters to control the algorithm, run faster, and do not require any assumptions to be made such as maximum number of change points. These properties also make them good candidates for online change point analysis. © 2017 IEEE., Ege Üniversitesi: :17-UBE-001, This study was supported by Ege University Scientific Research Projects Coordination Unit (Project Number:17-UBE-001).

Tensor factorization has been increasingly used to analyze high-dimensional low-rank data ofmassive scale in numerous application domains, including recommender systems, graphanalytics, health-care data analysis, signal processing, chemometrics, and many others.In these applications, efficient computation of tensor decompositions is crucial to be able tohandle such datasets of high volume. The main focus of this thesis is on efficient decompositionof high dimensional sparse tensors, with hundreds of millions to billions of nonzero entries,which arise in many emerging big data applications. We achieve this through three majorapproaches.In the first approach, we provide distributed memory parallel algorithms with efficientpoint-to-point communication scheme for reducing the communication cost. These algorithmsare agnostic to the partitioning of tensor elements and low rank decomposition matrices, whichallow us to investigate effective partitioning strategies for minimizing communication cost whileestablishing computational load balance. We use hypergraph-based techniques to analyze computational and communication requirements in these algorithms, and employ hypergraphpartitioning tools to find suitable partitions that provide much better scalability.Second, we investigate effective shared memory parallelizations of these algorithms. Here, we carefully determine unit computational tasks and their dependencies, and express them using aproper data structure that exposes the parallelism underneath.Third, we introduce a tree-based computational scheme that carries out expensive operations(involving the multiplication of the tensor with a set of vectors or matrices, found at the core ofthese algorithms) faster by factoring out and storing common partial results and effectivelyre-using them. With this computational scheme, we asymptotically reduce the number oftensor-vector and -matrix multiplications for high dimensional tensors, and thereby rendercomputing tensor decompositions significantly cheaper both for sequential and parallelalgorithms.Finally, we diversify this main course of research with two extensions on similar themes.The first extension involves applying the tree-based computational framework to computingdense tensor decompositions, with an in-depth analysis of computational complexity andmethods to find optimal tree structures minimizing the computational cost. The second workfocuses on adapting effective communication and partitioning schemes of our parallel sparsetensor decomposition algorithms to the widely used non-negative matrix factorization problem,through which we obtain significantly better parallel scalability over the state of the artimplementations.We point out that all theoretical results in the thesis are nicely corroborated by parallelexperiments on both shared-memory and distributed-memory platforms. With these fastalgorithms as well as their tuned implementations for modern HPC architectures, we rendertensor and matrix decomposition algorithms amenable to use for analyzing massive scaledatasets.; La factorisation des tenseurs est au coeur des méthodes d'analyse des données massives multidimensionnelles dans de nombreux domaines, dont les systèmes de recommandation, les graphes, les données médicales, le traitement du signal, la chimiométrie, et bien d'autres.Pour toutes ces applications, l'obtention rapide de la décomposition des tenseurs est cruciale pour pouvoir traiter manipuler efficacement les énormes volumes de données en jeu.L'objectif principal de cette thèse est la conception d'algorithmes pour la décomposition de tenseurs multidimensionnels creux, possédant de plusieurs centaines de millions à quelques milliards de coefficients non-nuls. De tels tenseurs sont omniprésents dans les applications citées plus haut.Nous poursuivons cet objectif via trois approches.En premier lieu, nous proposons des algorithmes parallèles à mémoire distribuée, comprenant des schémas de communication point-à-point optimisés, afin de réduire les coûts de communication. Ces algorithmes sont indépendants du partitionnement des éléments du tenseur et des matrices de faible rang. Cette propriété nous permet de proposer des stratégies de partitionnement visant à minimiser le coût de communication tout en préservant l'équilibrage de charge entre les ressources. Nous utilisons des techniques d'hypergraphes pour analyser les paramètres de calcul et de communication de ces algorithmes, ainsi que des outils de partitionnement d'hypergraphe pour déterminer des partitions à même d'offrir un meilleur passage à l'échelle. Deuxièmement, nous étudions la parallélisation sur plate-forme à mémoire partagée de ces algorithmes. Dans ce contexte, nous déterminons soigneusement les tâches de calcul et leur dépendances, et nous les exprimons en termes d'une structure de données idoine, et dont la manipulation permet de révéler le parallélisme intrinsèque du problème. Troisièmement, nous présentons un schéma de calcul en forme d'arbre binaire pour représenter les noyaux de calcul les plus coûteux des algorithmes, comme la multiplication du tenseur par un ensemble de vecteurs ou de matrices donnés. L'arbre binaire permet de factoriser certains résultats intermédiaires, et de les ré-utiliser au fil du calcul. Grâce à ce schéma, nous montrons comment réduire significativement le nombre et le coût des multiplications tenseur-vecteur et tenseur-matrice, rendant ainsi la décomposition du tenseur plus rapide à la fois pour la version séquentielle et la version parallèle des algorithmes.Enfin, le reste de la thèse décrit deux extensions sur des thèmes similaires. La première extension consiste à appliquer le schéma d'arbre binaire à la décomposition des tenseurs denses, avec une analyse précise de la complexité du problème et des méthodes pour trouver la structure arborescente qui minimise le coût total. La seconde extension consiste à adapter les techniques de partitionnement utilisées pour la décomposition des tenseurs creux à la factorisation des matrices non-négatives, problème largement étudié et pour lequel nous obtenons des algorithmes parallèles plus efficaces que les meilleurs actuellement connus.Tous les résultats théoriques de cette thèse sont accompagnés d'implémentations parallèles,aussi bien en mémoire partagée que distribuée. Tous les algorithmes proposés, avec leur réalisation sur plate-forme HPC, contribuent ainsi à faire de la décomposition de tenseurs un outil prometteur pour le traitement des masses de données actuelles et à venir.

Dense tensor decompositions have been widely used in many signal processing problems including analyzing speech signals, identifying the localization of signal sources, and many other communication applications.Computing these decompositions poses major computational challenges for big datasets emerging in these domains.CANDECOMP/PARAFAC~(CP) and Tucker formulations are the prominent tensor decomposition schemes heavily used in these fields, and the algorithms for computing them involve applying two core operations, namely tensor-times-matrix~(TTM) and -vector~(TTV) multiplication, which are executed repetitively within an iterative framework.In the recent past, efficient computational schemes using a data structure called dimension tree are employed to significantly reduce the cost of these two operations through storing and reusing partial results that are commonly used across different iterations of these algorithms.This framework has been introduced for sparse CP and Tucker decompositions in the literature, and a recent work investigates using an optimal binary dimension tree structure in computing dense Tucker decompositions.In this paper, we investigate finding an optimal dimension tree for both CP and Tucker decompositions.We show that finding an optimal dimension tree is NP-hard for both decompositions, provide faster exact algorithms for finding an optimal dimension tree in $O(3^N)$ time using $O(2^N)$ space for the Tucker case, and extend the algorithm to the case of CP decomposition with the same time and space complexities.; Les décompositions de tenseurs sont largement utilisées dans plusieurs problèmes de traitement du signal, notamment l’analyse des signaux vocaux, l’identification de la localisation des sources des signaux et d’autres applications de communication. Le calcul de ces décompositions pose un énorme défi calculatoire, particulièrement pour les grands jeux de données apparaissant dans ces domaines. Les formulations du CANDECOMP/PARAFAC (CP) et Tucker sont les deux décompositions les plus intensivement utilisées. Les algorithmes pour calculer ces décompositions incluent deux operations principales, tenseur-fois-matrice (TTM) et -vector (TTV), qui se répètent dans une méthode iterative. Dans ce rapport, on introduit un nouveau schema de calcul pour éffectuer ces deux operations efficacement en calculant les decompositions CP et Tucker. Pour ce faire, on adopte une structure d’arbre qui permet de stocker et ré-utiliser les résultats partiels pour réduire le coût de calcul de manière significative. Ensuite, on fournit un algorithme glouton, s’exécutant en temps linéaire, pour trouver l’ordre optimal minimisant le coût de calcule d’une serie de TTMs. Finalement, on examine le problème de trouver un arbre optimal pour calculer les décompositions CP et Tucker avec le nouveau schema de calcul proposé, et on prouve que ce problème est NP-difficile dans les deux cas.

This study investigated the relationship between childhood trauma, irritability, and emotion recognition, in schizophrenia patients during a psychotic break. Thirty-six schizophrenia inpatients and 36 healthy controls were assessed with the Irritability Questionnaire (IRQ) and two facial emotion recognition tasks, the Emotion Discrimination Test (EDT) and Emotion Identification Test (EIT). Patients were further assessed with the Structured Clinical Interview for DSM III-R Axis II Disorders (SCID-II), the Positive and Negative Symptom Scale (PANSS), and the Childhood Trauma Questionnaire-28 (CTQ-28). EDT and EIT performance was significantly impaired in patients compared to healthy controls. Furthermore, patients tended to misidentify sad, surprised, or angry faces as showing fear, and this misidentification correlated with the patients' irritability. Childhood adversity increased irritability both directly and indirectly through emotion misidentification.

Objective Although emotion dysregulation, one of the core features of depression, has long been thought to be a vulnerability factor for major depressive disorder (MDD), surprisingly few functional magnetic resonance imaging (fMRI) studies have investigated neural correlates of emotion regulation strategies in unaffected high risk individuals. Method Sixteen high risk (RSK) young women and fifteen matched low risk controls (CTL) were scanned using fMRI while performing an emotion regulation task. During this task, participants were instructed to reappraise their negative emotions elicited by International Affective Picture System images (IAPS). In addition, Difficulties in Emotion Regulation Strategies Scale (DERS) was used to assess participants' emotion dysregulation levels. Results Both RSK and CTL individuals show increased amygdala activation in response to negative emotional stimuli, however no difference was found between groups in using cognitive reappraisal strategies and functions of brain regions implicated in cognitive reappraisal. Interestingly, our psychometric test results indicate that high risk individuals are characterised by lower perceived emotional clarity (EC). Conclusion Results of the current study suggest depression vulnerability may not be linked to the effectiveness of cognitive reappraisal. Alternatively, lower EC may be a vulnerability factor for depression.

Corpus callosum (CC) is an important structure for medical image registration. We propose three novel fully automated for the extraction of CC. Our first algorithm, Valley matching (VM), is based on fixed searched range in histogram processing and uses prior anatomical information for locating CC. The second one, Evolutionary CC Detection (ECD), based on genetic algorithm presents a new fitness function that uses anatomical ratios, instead of fixed prior knowledge without the need for preprocessing. The final one, called Evolutionary Valley Matching (EVM), takes advantages of the strong points of the first and second algorithms. The search space defined for ECD is reduced by VM which uses crowding method to find the peaks in the multi-modal histogram. Another important contribution of this study is that there is no existing method using genetic algorithm for extracting CC. Our proposed algorithms perform with the success rates up to 95.5%.