Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

628.3 <043> --- Academic collection --- #BIBC:T1997 --- Sewage. Surface and foul water. Properties and treatment. Sewage works--Dissertaties --- Theses --- 628.3 <043> Sewage. Surface and foul water. Properties and treatment. Sewage works--Dissertaties

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Academic collection --- Courses - Syllabuses

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Data is everywhere, but it is only useful if information can be extracted from it. To this end, data analysis techniques have been developed to find features underlying the data. One class of data analysis techniques uses tensor decompositions. Tensors are higher-dimensional extensions of vectors and matrices and allow us to represent multiway data in a natural way. One of the key strengths of tensors is that their decompositions are unique under mild conditions without imposing additional constraints, unlike matrix decompositions. This property opens up many interesting applications of tensor decompositions, in particular with respect to blind source separation and independent component analysis.Independent component analysis (ICA) tries to find the statistically independent signals underlying a mixture, which is useful in many fields such as telecommunications, speech separation and biomedical data analysis. These mixtures are often modeled as instantaneous mixtures of source signals and tensor methods that can blindly separate such mixtures are well known. However, in many applications a convolutive mixture model is more appropriate to take delays and reflections of the source signals into account. This is for instance the case for speech signals in a room or telecommunications signals impinging on antennas. When the mixtures are convolutive, additional structure arises in the separation problem. Current tensor-based methods in the literature do not fully exploit this Toeplitz or Hankel structure. This thesis presents new subspace-based methods that take more of this structure into account, which leads to efficient and more accurate results than the current state-of-the-art tensor methods. Additionally, relaxed uniqueness bounds are formulated that exploit the available structure as well. This thesis also presents a method for a related structured tensor decomposition in which all factor matrices have block-circulant structure.Apart from exploiting the structure that appears in convolutive ICA, this thesis also presents how known techniques in instantaneous ICA can be ported to the convolutive case. This includes coupled tensor decompositions to combine second- and fourth-order statistics, and the usage of incomplete tensors to reduce the computational complexity.Another interesting question unrelated to convolutive ICA is whether we can compare underlying tensor factors without having to compute their full decompositions. This is highly relevant for any tensor classification problem but has only received limited attention in research. In this thesis, we present foundational theorems and algorithms that show how tensor factors can be compared in two modes for several underlying tensor decompositions. One of these methods is used further to develop an algorithm that is able to compare the cluster centers in two different datasets, without having to compute the actual clusters.

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Over the years, many algorithms have been designed to decompose a matrix into a product of other matrices. These matrix decompositions can be used to compress data with a minimal loss of information or for extracting meaningful components. More recently, tensor decompositions such as the canonicalpolyadic decomposition (CPD) and the low multilinear rank approximation (LMLRA) have been designed as higher-order generalizations of these matrix decompositions leading to better results in certain applications. While matrix methods offer a natural way to process data involving two variables, theyoften fail to recognize higher-order patterns. Tensors, in contrast, can have any number of modes and can preserve the higher-order structure of the data, which makes them highly useful to analyze multi-variable data. Second, tensor decompositions are often unique in cases where their matrix counterparts arenot. As fewer (artificial) restrictions such as orthogonality or triangle structure have to be imposed on the terms to enforce their uniqueness, this can lead to a decomposition in more meaningful terms.Current tensor decomposition methods commonly make two assumptions about the data supplied by the user. First, they assume that all data is available at once and second, they consider noise on the data to be additive and Gaussian distributed. In many applications, these assumptions are justified and one can decompose the tensor using standard batch methods that minimize the least-squares distance between the tensor and the low-rank model. However, as tensor methods are applied to more and more real-world problems, the number of cases where these assumptions are clearly violated increases as well. Any real-time application, for instance, violates the assumption of having all data available at the start. In such applications, such as process control or patient monitoring, new data arrive at certain time intervals. This data should be included in the tensor model without recomputing the full decomposition from scratch, as such approaches are not time- nor resource-efficient. Similarly, it is not hard to find applications where the assumption of additive Gaussian noise is not satisfied. Audio data, for instance, is generally modelled using a non-Gaussian noise distribution while low-light imaging assumes Poisson distributed pixel intensities. When the real and imaginary parts of a signal are collected independently, as is typically the case in MRI imaging, the noise distribution of the signal magnitude is Rician instead of Gaussian. For these applications, fitting a low-rank model to the tensor with the least-squares distance as cost function will lead to suboptimal solutions while a suitable choice of an alternative cost function can provide a model that corresponds better to the underlying components.In this thesis, we provide methods that can still efficiently compute a suitable tensor decomposition when either of the aforementioned assumptions is not satisfied. First, we propose updating methods for both the CPD and the LMLRA. These methods start from an existing tensor decomposition and efficiently update this decomposition when new data arrives. By exploiting the multilinear structure of the tensor models, these methods are both efficient and accurate in tracking low-rank representations of the data. When desired, one can even use a weighted least-squares approach to weight certain tensor entries more heavily than others, e.g., when the uncertainty on the collected tensor entries varies between sensors. The performance of the CPD method is demonstrated for the monitoring of brain hemodynamics coupling in neonates. A second contribution is that we extend the existing least-squares CPD algorithms to other cost functions. This makes it possible to use a dedicated cost function for the application at hand and to obtain better solutions than when standard least-squares approaches are used. We focus in particular on the class of beta-divergence cost functions, of which the least-squares distance and the Kullback—Leibler and Itakura—Saito divergences are special cases, but we also supply a dedicated method for the Rician cost function. Additionally, any twice-differentiable entry-wise cost function is supported. We demonstrate the improved effectiveness of the method over least-squares approaches by blindly separating a series of fMRI signals. Finally, we propose an improved algebraic algorithm to compute the CPD of a tensor in the least-squares sense. By recursively splitting the generalized eigenspace into successively smaller subspaces instead of determining all generalized eigenvalues at once, the decomposition is more accurate, yet can still be obtained efficiently. The algebraic solution can be used as an initialization for more costly optimization algorithms and significantly reduce the number of iterations and, hence, also the runtime of these methods.