Sparse and Parametric Modeling with Applications to Acoustics and Audio

Helena Peic Tukuljac
EPFL, 2020
Thèse EPFL

Résumé

Recent advances in signal processing, machine learning and deep learning with sparse intrinsic structure of data have paved the path for solving inverse problems in acoustics and audio. The main task of this thesis was to bridge the gap between the powerful mathematical tools and challenging problems in acoustics and audio. This thesis consists out of two main parts.

The first part of the thesis focuses on the questions related to acoustic simulations that comply with the "real world" constraints and the acoustic data acquisition inside of closed spaces. The simulated and measured data is used to solve various types of inverse problems with underlying sparsity. By using the technique of compressed sensing, we estimate the room modes, localize sound sources in a room and also estimate room's geometry. The Finite Rate of Innovation technique is coupled with non-convex optimization for the task of blind deconvolution in the context of echo retrieval. We also invent a new statistical measure for the echo density for the purpose of detecting the type of acoustic environment from its acoustic impulse response, even beyond fully closed spaces. These types solutions can have an application in the blooming domain of virtual, augmented and mixed reality for sound compression and rendering.

The second part of the thesis focuses on the recent trends in machine learning that are centered around deep learning. Large scale data acquisition of acoustic impulse responses is still a challenging and very expensive task. Also, the existing databases tend to be too heterogeneous to be merged, due to the lack of the standardization of the acquisition procedure, and also the available metadata tends to be incomplete. In order to keep up with the recent trends and avoid the difficulties that come from the lack of large scale acoustical data, the last part of research in this thesis has diverged from the rest and is devoted to deep learning applied to classification problems in audio with the focus on speech and environmental sounds. The learning procedure is parametrized, which results in an off-grid learning procedure for audio classification. Learned trends align with perceptual trends, which helps the interpretation of the achieved results.

Official source

https://infoscience.epfl.ch/record/273930?ln=fr

À propos de ce résultat

Cette page est générée automatiquement et peut contenir des informations qui ne sont pas correctes, complètes, à jour ou pertinentes par rapport à votre recherche. Il en va de même pour toutes les autres pages de ce site. Veillez à vérifier les informations auprès des sources officielles de l'EPFL.

Concepts associés

Chargement

Publications associées

Chargement

Helena Peic Tukuljac
EPFL, 2020
Thèse EPFL

Résumé

Official source

https://infoscience.epfl.ch/record/273930?ln=fr

À propos de ce résultat

Concepts associés (25)

Concepts associés

Chargement

Publications associées (31)

Publications associées

Chargement

Acoustic tomography aims at recovering the unknown parameters that describe a field of interest by studying the physical characteristics of sound propagating through the considered field. The tomographic approach is appealing in that it is non-invasive and allows to obtain a significantly larger amount of data compared to the classical one-sensor one-measurement setup. It has, however, two major drawbacks which may limit its applicability in a practical setting: the methods by which the tomographic data are acquired and then converted to the field values are computationally intensive and often ill-conditioned. This thesis specifically addresses these two shortcomings by proposing novel acoustic tomography algorithms for signal acquisition and field reconstruction. The first part of our exposition deals with some theoretical aspects of the tomographic sampling problems and associated reconstruction schemes for scalar and vector tomography. We show that the classical time-of-flight measurements are not sufficient for full vector field reconstruction. As a solution, an additional set of measurements is proposed. The main advantage of the proposed set is that it can be directly computed from acoustic measurements. It thus avoids the need for extra measuring devices. We then describe three novel reconstruction methods that are conceptually quite different. The first one is based on quadratic optimization and does not require any a priori information. The second method builds upon the notion of sparsity in order to increase the reconstruction accuracy when little data is available. The third approach views tomographic reconstruction as a parametric estimation problem and solves it using recent sampling results on non-bandlimited signals. The proposed methods are compared and their respective advantages are outlined. The second part of our work is dedicated to the application of the proposed algorithms to three practical problems: breast cancer detection, thermal therapy monitoring, and temperature monitoring in the atmosphere. We address the problem of breast cancer detection by computing a map of sound speed in breast tissue. A noteworthy contribution of this thesis is the development of a signal processing technique that significantly reduces the artifacts that arise in very inhomogeneous and absorbent tissue. Temperature monitoring during thermal therapies is then considered. We show how some of our algorithms allow for an increased spatial resolution and propose ways to reduce the computational complexity. Finally, we demonstrate the feasibility of tomographic temperature monitoring in the atmosphere using a custom-built laboratory-scale experiment. In particular, we discuss various practical aspects of time-of-flight measurement using cheap, off-the-shelf sensing devices.

EPFL2008

Chargement

Robust and Efficient Data Clustering with Signal Processing on Graphs

Lionel Jérémie Martin

Data is pervasive in today's world and has actually been for quite some time. With the increasing volume of data to process, there is a need for faster and at least as accurate techniques than what we already have. In particular, the last decade recorded the effervescence of social networks and ubiquitous sensing (through smartphones and the Internet of Things). These phenomena, including also the progresses in bioinformatics and traffic monitoring, pushed forward the research on graph analysis and called for more efficient techniques. Clustering is an important field of machine learning because it belongs to the unsupervised techniques (i.e., one does not need to possess a ground truth about the data to start learning). With it, one can extract meaningful patterns from large data sources without requiring an expert to annotate a portion of the data, which can be very costly. However, the techniques of clustering designed so far all tend to be computationally demanding and have trouble scaling with the size of today's problems. The emergence of Graph Signal Processing, attempting to apply traditional signal processing techniques on graphs instead of time, provided additional tools for efficient graph analysis. By considering the clustering assignment as a signal lying on the nodes of the graph, one may now apply the tools of GSP to the improvement of graph clustering and more generally data clustering at large. In this thesis, we present several techniques using some of the latest developments of GSP in order to improve the scalability of clustering, while aiming for an accuracy resembling that of Spectral Clustering, a famous graph clustering technique that possess a solid mathematical intuition. On the one hand, we explore the benefits of random signal filtering on a practical and theoretical aspect for the determination of the eigenvectors of the graph Laplacian. In practice, this attempt requires the design of polynomial approximations of the step function for which we provided an accelerated heuristic. We used this series of work in order to reduce the complexity of dynamic graphs clustering, the problem of defining a partition to a graph which is evolving in time at each snapshot. We also used them to propose a fast method for the determination of the subspace generated by the first eigenvectors of any symmetrical matrix. This element is useful for clustering as it serves in Spectral Clustering but it goes beyond that since it also serves in graph visualization (with Laplacian Eigenmaps) and data mining (with Principal Components Projection). On the other hand, we were inspired by the latest works on graph filter localization in order to propose an extremely fast clustering technique. We tried to perform clustering by only using graph filtering and combining the results in order to obtain a partition of the nodes. These different contributions are completed by experiments using both synthetic datasets and real-world problems. Since we think that research should be shared in order to progress, all the experiments made in this thesis are publicly available on my personal Github account.

EPFL2018

Chargement

Learning with Structured Sparsity: From Discrete to Convex and Back.

Marwa El Halabi

In modern-data analysis applications, the abundance of data makes extracting meaningful information from it challenging, in terms of computation, storage, and interpretability. In this setting, exploiting sparsity in data has been essential to the development of scalable methods to problems in machine learning, statistics and signal processing. However, in various applications, the input variables exhibit structure beyond simple sparsity. This motivated the introduction of structured sparsity models, which capture such sophisticated structures, leading to a significant performance gains and better interpretability. Structured sparse approaches have been successfully applied in a variety of domains including computer vision, text processing, medical imaging, and bioinformatics. The goal of this thesis is to improve on these methods and expand their success to a wider range of applications. We thus develop novel methods to incorporate general structure a priori in learning problems, which balance computational and statistical efficiency trade-offs. To achieve this, our results bring together tools from the rich areas of discrete and convex optimization. Applying structured sparsity approaches in general is challenging because structures encountered in practice are naturally combinatorial. An effective approach to circumvent this computational challenge is to employ continuous convex relaxations. We thus start by introducing a new class of structured sparsity models, able to capture a large range of structures, which admit tight convex relaxations amenable to efficient optimization. We then present an in-depth study of the geometric and statistical properties of convex relaxations of general combinatorial structures. In particular, we characterize which structure is lost by imposing convexity and which is preserved. We then focus on the optimization of the convex composite problems that result from the convex relaxations of structured sparsity models. We develop efficient algorithmic tools to solve these problems in a non-Euclidean setting, leading to faster convergence in some cases. Finally, to handle structures that do not admit meaningful convex relaxations, we propose to use, as a heuristic, a non-convex proximal gradient method, efficient for several classes of structured sparsity models. We further extend this method to address a probabilistic structured sparsity model, we introduce to model approximately sparse signals.

EPFL2018

Chargement

Learning with Structured Sparsity: From Discrete to Convex and Back.

Marwa El Halabi

EPFL2018

EPFL2008

Robust and Efficient Data Clustering with Signal Processing on Graphs

Lionel Jérémie Martin

EPFL2018