Weiß, Christian (2017)
Estimating Sparse Representations from Dictionaries With Uncertainty.
Technische Universität Darmstadt
Dissertation, Erstveröffentlichung
Kurzbeschreibung (Abstract)
In the last two decades, sparse representations have gained increasing attention in a variety of engineering applications. A sparse representation of a signal requires a dictionary of basic elements that describe salient and discriminant features of that signal. When the dictionary is created from a mathematical model, its expressiveness depends on the quality of this model. In this dissertation, the problem of estimating sparse representations in the presence of errors and uncertainty in the dictionary is addressed. In the first part, a statistical framework for sparse regularization is introduced. The second part is concerned with the development of methodologies for estimating sparse representations from highly redundant dictionaries along with unknown dictionary parameters. The presented methods are illustrated using applications in direction finding and fiber-optic sensing. They serve as illustrative examples for investigating the abstract problems in the theory of sparse representations. Estimating a sparse representation often involves the solution of a regularized optimization problem. The presented regularization framework offers a systematic procedure for the determination of a regularization parameter that accounts for the joint effects of model errors and measurement noise. It is determined as an upper bound of the mean-squared error between the corrupted data and the ideal model. Despite proper regularization, the quality and accuracy of the obtained sparse representation remains affected by model errors and is indeed sensitive to changes in the regularization parameter. To alleviate this problem, dictionary calibration is performed. The framework is applied to the problem of direction finding. Redundancy enables the dictionary to describe a broader class of observations but also increases the similarity between different entries, which leads to ambiguous representations. To address the problem of redundancy and additional uncertainty in the dictionary parameters, two strategies are pursued. Firstly, an alternating estimation method for iteratively determining the underlying sparse representation and the dictionary parameters is presented. Also, theoretical bounds for the estimation errors are derived. Secondly, a Bayesian framework for estimating sparse representations and dictionary learning is developed. A hierarchical structure is considered to account for uncertainty in prior assumptions. The considered model for the coefficients of the sparse representation is particularly designed to handle high redundancy in the dictionary. Approximate inference is accomplished using a hybrid Markov Chain Monte Carlo algorithm. The performance and practical applicability of both methodologies is evaluated for a problem in fiber-optic sensing, where a mathematical model for the sensor signal is compiled. This model is used to generate a suitable parametric dictionary.
Typ des Eintrags: | Dissertation | ||||
---|---|---|---|---|---|
Erschienen: | 2017 | ||||
Autor(en): | Weiß, Christian | ||||
Art des Eintrags: | Erstveröffentlichung | ||||
Titel: | Estimating Sparse Representations from Dictionaries With Uncertainty | ||||
Sprache: | Englisch | ||||
Referenten: | Zoubir, Prof. Dr. Abdelhak M. ; Gini, Prof. Dr. Fulvio | ||||
Publikationsjahr: | April 2017 | ||||
Ort: | Darmstadt | ||||
Datum der mündlichen Prüfung: | 24 März 2017 | ||||
URL / URN: | http://tuprints.ulb.tu-darmstadt.de/6152 | ||||
Kurzbeschreibung (Abstract): | In the last two decades, sparse representations have gained increasing attention in a variety of engineering applications. A sparse representation of a signal requires a dictionary of basic elements that describe salient and discriminant features of that signal. When the dictionary is created from a mathematical model, its expressiveness depends on the quality of this model. In this dissertation, the problem of estimating sparse representations in the presence of errors and uncertainty in the dictionary is addressed. In the first part, a statistical framework for sparse regularization is introduced. The second part is concerned with the development of methodologies for estimating sparse representations from highly redundant dictionaries along with unknown dictionary parameters. The presented methods are illustrated using applications in direction finding and fiber-optic sensing. They serve as illustrative examples for investigating the abstract problems in the theory of sparse representations. Estimating a sparse representation often involves the solution of a regularized optimization problem. The presented regularization framework offers a systematic procedure for the determination of a regularization parameter that accounts for the joint effects of model errors and measurement noise. It is determined as an upper bound of the mean-squared error between the corrupted data and the ideal model. Despite proper regularization, the quality and accuracy of the obtained sparse representation remains affected by model errors and is indeed sensitive to changes in the regularization parameter. To alleviate this problem, dictionary calibration is performed. The framework is applied to the problem of direction finding. Redundancy enables the dictionary to describe a broader class of observations but also increases the similarity between different entries, which leads to ambiguous representations. To address the problem of redundancy and additional uncertainty in the dictionary parameters, two strategies are pursued. Firstly, an alternating estimation method for iteratively determining the underlying sparse representation and the dictionary parameters is presented. Also, theoretical bounds for the estimation errors are derived. Secondly, a Bayesian framework for estimating sparse representations and dictionary learning is developed. A hierarchical structure is considered to account for uncertainty in prior assumptions. The considered model for the coefficients of the sparse representation is particularly designed to handle high redundancy in the dictionary. Approximate inference is accomplished using a hybrid Markov Chain Monte Carlo algorithm. The performance and practical applicability of both methodologies is evaluated for a problem in fiber-optic sensing, where a mathematical model for the sensor signal is compiled. This model is used to generate a suitable parametric dictionary. |
||||
Alternatives oder übersetztes Abstract: |
|
||||
URN: | urn:nbn:de:tuda-tuprints-61526 | ||||
Sachgruppe der Dewey Dezimalklassifikatin (DDC): | 600 Technik, Medizin, angewandte Wissenschaften > 620 Ingenieurwissenschaften und Maschinenbau | ||||
Fachbereich(e)/-gebiet(e): | 18 Fachbereich Elektrotechnik und Informationstechnik > Institut für Nachrichtentechnik 18 Fachbereich Elektrotechnik und Informationstechnik > Institut für Nachrichtentechnik > Signalverarbeitung Exzellenzinitiative > Graduiertenschulen > Graduate School of Computational Engineering (CE) Exzellenzinitiative > Graduiertenschulen 18 Fachbereich Elektrotechnik und Informationstechnik Exzellenzinitiative |
||||
Hinterlegungsdatum: | 23 Apr 2017 19:55 | ||||
Letzte Änderung: | 23 Apr 2017 19:55 | ||||
PPN: | |||||
Referenten: | Zoubir, Prof. Dr. Abdelhak M. ; Gini, Prof. Dr. Fulvio | ||||
Datum der mündlichen Prüfung / Verteidigung / mdl. Prüfung: | 24 März 2017 | ||||
Export: | |||||
Suche nach Titel in: | TUfind oder in Google |
Frage zum Eintrag |
Optionen (nur für Redakteure)
Redaktionelle Details anzeigen |