| 1. |
- Broman, David, 1977-
(författare)
-
Meta-Languages and Semantics for Equation-Based Modeling and Simulation
- 2010
-
Doktorsavhandling (övrigt vetenskapligt/konstnärligt)abstract
- Performing computational experiments on mathematical models instead of building and testing physical prototypes can drastically reduce the develop cost for complex systems such as automobiles, aircraft, and powerplants. In the past three decades, a new category of equation-based modeling languages has appeared that is based on acausal and object-oriented modeling principles, enabling good reuse of models. However, the modeling languages within this category have grown to be large and complex, where the specifications of the language's semantics are informally defined, typically described in natural languages. The lack of a formal semantics makes these languages hard to interpret unambiguously and to reason about. This thesis concerns the problem of designing the semantics of such equation-based modeling languages in a way that allows formal reasoning and increased correctness. The work is presented in two parts.In the first part we study the state-of-the-art modeling language Modelica. We analyze the concepts of types in Modelica and conclude that there are two kinds of type concepts: class types and object types. Moreover, a concept called structural constraint delta is proposed, which is used for isolating the faults of an over- or under-determined model.In the second part, we introduce a new research language called the Modeling Kernel Language (MKL). By introducing the concept of higher-order acausal models (HOAMs), we show that it is possible to create expressive modeling libraries in a manner analogous to Modelica, but using a small and simple language concept. In contrast to the current state-of-the-art modeling languages, the semantics of how to use the models, including meta operations on models, are also specified in MKL libraries. This enables extensible formal executable specifications where important language features are expressed through libraries rather than by adding completely new language constructs. MKL is a statically typed language based on a typed lambda calculus. We define the core of the language formally using operational semantics and prove type safety. An MKL interpreter is implemented and verified in comparison with a Modelica environment.
|
|
| 2. |
- Lindsten, Fredrik, 1984-
(författare)
-
Particle filters and Markov chains for learning of dynamical systems
- 2013
-
Doktorsavhandling (övrigt vetenskapligt/konstnärligt)abstract
- Sequential Monte Carlo (SMC) and Markov chain Monte Carlo (MCMC) methods provide computational tools for systematic inference and learning in complex dynamical systems, such as nonlinear and non-Gaussian state-space models. This thesis builds upon several methodological advances within these classes of Monte Carlo methods.Particular emphasis is placed on the combination of SMC and MCMC in so called particle MCMC algorithms. These algorithms rely on SMC for generating samples from the often highly autocorrelated state-trajectory. A specific particle MCMC algorithm, referred to as particle Gibbs with ancestor sampling (PGAS), is suggested. By making use of backward sampling ideas, albeit implemented in a forward-only fashion, PGAS enjoys good mixing even when using seemingly few particles in the underlying SMC sampler. This results in a computationally competitive particle MCMC algorithm. As illustrated in this thesis, PGAS is a useful tool for both Bayesian and frequentistic parameter inference as well as for state smoothing. The PGAS sampler is successfully applied to the classical problem of Wiener system identification, and it is also used for inference in the challenging class of non-Markovian latent variable models.Many nonlinear models encountered in practice contain some tractable substructure. As a second problem considered in this thesis, we develop Monte Carlo methods capable of exploiting such substructures to obtain more accurate estimators than what is provided otherwise. For the filtering problem, this can be done by using the well known Rao-Blackwellized particle filter (RBPF). The RBPF is analysed in terms of asymptotic variance, resulting in an expression for the performance gain offered by Rao-Blackwellization. Furthermore, a Rao-Blackwellized particle smoother is derived, capable of addressing the smoothing problem in so called mixed linear/nonlinear state-space models. The idea of Rao-Blackwellization is also used to develop an online algorithm for Bayesian parameter inference in nonlinear state-space models with affine parameter dependencies.
|
|
| 3. |
- Wilroth, Johanna, 1994-
(författare)
-
Exploring Auditory Attention Using EEG
- 2024
-
Licentiatavhandling (övrigt vetenskapligt/konstnärligt)abstract
- Listeners with normal-hearing often overlook their ability to comprehend speech in noisy environments effortlessly. Our brain’s adeptness at identifying and amplifying attended voices while suppressing unwanted background noise, known as the cocktail party problem, has been extensively researched for decades. Yet, many aspects of this complex puzzle remain unsolved and listeners with hearing-impairment still struggle to focus on a specific speaker in noisy environments. While recent intelligent hearing aids have improved noise suppression, the problem of deciding which speaker to enhance remains unsolved, leading to discomfort for many hearing aid users in noisy environments.In this thesis, we explore the complexities of the human brain in challenging auditory environments. Two datasets are investigated where participants were tasked to selectively attend to one of two competing voices, replicating a cocktail-party scenario. The auditory stimuli trigger neurons to generate electrical signals that propagate in all directions. When a substantial number of neurons fire simultaneously, their collective electrical signal becomes detectable by small electrodes placed on the head. This method of measuring brain activity, known as electroencephalography (EEG), holds potential to provide feedback to the hearing aids, enabling adjustments to enhance attended voice(s).EEG data is often noisy, incorporating neural responses with artifacts such as muscle movements, eye blinks and heartbeats. In the first contribution of this thesis, we focus on comparing different manual and automatic artifact-rejection techniques and assessing their impact on auditory attention decoding (AAD).While EEG measurements offer high temporal accuracy, spatial resolution is inferior compared to alternative tools like magnetoencephalography (MEG). This difference poses a considerable challenge for source localization with EEG data. In the second contribution of this thesis, we demonstrate anticipated activity in the auditory cortex using EEG data from a single listener, employing Neuro-Current Response Functions (NCRFs). This method, previously evaluated only with MEG data, holds significant promise in hearing aid development.EEG data may involve both linear and nonlinear components due to the propagation of the electrical signals through brain tissue, skull, and scalp with varying conductivities. In the third contribution, we aim to enhance source localization by introducing a binning-based nonlinear detection and compensation method. The results suggest that compensating for some nonlinear components produces more precise and synchronized source localization compared to original EEG data.In the fourth contribution, we present a novel domain adaptation framework that improves AAD performances for listeners with initially low classification accuracy. This framework focuses on classifying the direction (left or right) of attended speech and shows a significant accuracy improvement when transporting poor data from one listener to the domain of good data from different listeners.Taken together, the contributions of this thesis hold promise for improving the lives of hearing-impaired individuals by closing the loop between the brain and hearing aids.
|
|
| 4. |
- Sjanic, Zoran, 1975-
(författare)
-
Navigation and SAR Auto-focusing in a Sensor Fusion Framework
- 2011
-
Licentiatavhandling (övrigt vetenskapligt/konstnärligt)abstract
- Since its discovery, in the 1940's, radar (Radio Detection and Ranging) has become an important ranging sensor in many areas of technology and science. Most of the military and many civilian applications are unimaginable today without radar. With technology development, radar application areas have become larger and more available. One of these applications is Synthetic Aperture Radar (SAR), where an airborne radar is used to create high resolution images of the imaged scene. Although known since the 1950's, the SAR methods have been continuously developed and improved and new algorithms enabling real-time applications have emerged lately. Together with making the hardware components smaller and lighter, SAR has become an interesting sensor to be mounted on smaller unmanned aerial vehicles (UAV's). One important thing needed in the SAR algorithms is the estimate of the platform's motion, like position and velocity. Since this estimate is always corrupted with errors, particularly if lower grade navigation system, common in UAV applications, is used, the SAR images will be distorted. One of the most frequently appearing distortions caused by the unknown platform's motion is the image defocus. The process of correcting the image focus is called auto-focusing in SAR terminology. Traditionally, this problem was solved by methods that discard the platform's motion information, mostly due to the off-line processing approach, i.e. the images were created after the flight. Since the image (de)focus and the motion of the platform are related to each other, it is possible to utilise the information from the SAR images as a sensor and improve the estimate of the platform's motion. The auto-focusing problem can be cast as a sensor fusion problem. Sensor fusion is the process of fusing information from different sensors, in order to obtain best possible estimate of the states. Here, the information from sensors measuring platform's motion, mainly accelerometers, will be fused together with the information from the SAR images to estimate the motion of the flying platform. Two different methods based on this approach are tested on the simulated SAR data and the results are evaluated. One method is based on an optimisation based formulation of the sensor fusion problem, leading to batch processing, while the other method is based on the sequential processing of the radar data, leading to a filtering approach. The obtained results are promising for both methods and the obtained performance is comparable with the performance of a high precision navigation aid, such as Global Positioning System (GPS).
|
|
| 5. |
- Skoglund, Martin, 1981-
(författare)
-
Inertial Navigation and Mapping for Autonomous Vehicles
- 2014
-
Doktorsavhandling (övrigt vetenskapligt/konstnärligt)abstract
- Navigation and mapping in unknown environments is an important building block for increased autonomy of unmanned vehicles, since external positioning systems can be susceptible to interference or simply being inaccessible. Navigation and mapping require signal processing of vehicle sensor data to estimate motion relative to the surrounding environment and to simultaneously estimate various properties of the surrounding environment. Physical models of sensors, vehicle motion and external influences are used in conjunction with statistically motivated methods to solve these problems. This thesis mainly addresses three navigation and mapping problems which are described below.We study how a vessel with known magnetic signature and a sensor network with magnetometers can be used to determine the sensor positions and simultaneously determine the vessel's route in an extended Kalman filter (EKF). This is a so-called simultaneous localisation and mapping (SLAM) problem with a reversed measurement relationship.Previously determined hydrodynamic models for a remotely operated vehicle (ROV) are used together with the vessel's sensors to improve the navigation performance using an EKF. Data from sea trials is used to evaluate the system and the results show that especially the linear velocity relative to the water can be accurately determined.The third problem addressed is SLAM with inertial sensors, accelerometers and gyroscopes, and an optical camera contained in a single sensor unit. This problem spans over three publications.We study how a SLAM estimate, consisting of a point cloud map, the sensor unit's three dimensional trajectory and speed as well as its orientation, can be improved by solving a nonlinear least-squares (NLS) problem. NLS minimisation of the predicted motion error and the predicted point cloud coordinates given all camera measurements is initialised using EKF-SLAM.We show how NLS-SLAM can be initialised as a sequence of almost uncoupled problems with simple and often linear solutions. It also scales much better to larger data sets than EKF-SLAM. The results obtained using NLS-SLAM are significantly better using the proposed initialisation method than if started from arbitrary points. A SLAM formulation using the expectation maximisation (EM) algorithm is proposed. EM splits the original problem into two simpler problems and solves them iteratively. Here the platform motion is one problem and the landmark map is the other. The first problem is solved using an extended Rauch-Tung-Striebel smoother while the second problem is solved with a quasi-Newton method. The results using EM-SLAM are better than NLS-SLAM both in terms of accuracy and complexity.
|
|
| 6. |
- Abate, Ebba
(författare)
-
The impact of helminth infection in patients with active tuberculosis
- 2013
-
Doktorsavhandling (övrigt vetenskapligt/konstnärligt)abstract
- The geographic distribution of helminth infection and tuberculosis (TB) overlap substantially. Experimental animal models and limited data from humans have shown that intestinal helminths could subvert the host immune response towards a T-helper 2 (Th2)-type immune response and an increased regulatory T-cell activity (Tregs). This in turn affects the host's ability to mount an effective Th1 immune-mediated protection against Mycobacterium tuberculosis. However, evidence for this hypothesis in the human setting from helminth infected TB patients is limited. This thesis primarily focuses on the immunological and clinical impact of helminth infection on pulmonary TB. The kinetics of the Quantiferon-Gold (QFN) assay, which measures IFN-³ response to TB-specific antigens in whole blood was assessed and showed a modest decline during TB treatment to the level observed for healthy blood donors. We further assessed another clinical monitoring tool, the-TB-score, composed of clinical signs and symptoms of TB, and found an early decline two weeks after initiation of TB- treatment where a failure of decline correlated with increased mortality. Overall, the helminth co-infection rate was significantly higher in TB patients compared to healthy controls. Helminth co-infection was associated to a significantly higher rate of eosinophilia and IgE-levels in healthy controls and patients with tuberculosis. During the first weeks of anti-TB treatment, a marked decrease in the rate of helminth infection was observed in HIV co-infected compared to HIV-negative TB patients. However, helminth co-infection was more common in HIV negative than HIV positive TB patients. There was no detectable impact of helminth infection on the clinical presentation of pulmonary tuberculosis. At baseline, helminth co-infected TB patients showed an increased frequency of Tregs compared to helminth negative TB patients and healthy controls. This was accompanied by an increased rate of PPD stimulated IL-5 and spontaneous production of IL-10 by peripheral blood mononuclear cells among helminth co-infected TB patients. A placebo controlled randomized trial was conducted in order to test the hypothesis that albendazole treatment of helminth positive TB patients may improve the clinical response of TB by reducing the immunmodulatory effect of helminthes on TB immunity. A total of 140 helminth co-infected TB patients were randomized to albendazole (400 mg per os for three consecutive days) or placebo. No significant difference was observed between the albendazole and placebo group in terms of the primary outcome (TB score change between baseline and week 8). Among the secondary outcomes, a significant decline of peripheral eosinophil cells was observed in the albendazole treated group, but no effect on other outcome variables (changes in chest x-ray findings, IgE level and sputum smear conversion). Regarding the immunological assessment no significant difference was observed for changes in Tregs, and PPD-induced production of IFN- ³ or IL-5 although a non-significant trend of a decrease in IL-10 expressing PBMCs were observed in the albendazole group. Taken together, the burden of helminth infection was higher in TB patients than in a healthy control group. Helminth co-infection during pulmonary TB in the human setting induces an immune response characterized by increased IgE production, eosinophilia as well as increased levels of Tregs and spontaneous IL-10 production. Thus, the immunological impact of helminth infection on the outcome and risk for developing TB merits further investigation.
|
|
| 7. |
- Andersson, Carl
(författare)
-
Deep probabilistic models for sequential and hierarchical data
- 2022
-
Doktorsavhandling (övrigt vetenskapligt/konstnärligt)abstract
- Consider the problem where we want a computer program capable of recognizing a pedestrian on the road. This could be employed in a car to automatically apply the brakes to avoid an accident. Writing such a program is immensely difficult but what if we could instead use examples and let the program learn what characterizes a pedestrian from the examples. Machine learning can be described as the process of teaching a model (computer program) to predict something (the presence of a pedestrian) with help of data (examples) instead of through explicit programming.This thesis focuses on a specific method in machine learning, called deep learning. This method can arguably be seen as sole responsible for the recent upswing of machine learning in academia as well as in society at large. However, deep learning requires, in human standards, a huge amount of data to perform well which can be a limiting factor. In this thesis we describe different approaches to reduce the amount of data that is needed by encoding some of our prior knowledge about the problem into the model. To this end we focus on sequential and hierarchical data, such as speech and written language.Representing sequential output is in general difficult due to the complexity of the output space. Here, we make use of a probabilistic approach focusing on sequential models in combination with a deep learning structure called the variational autoencoder. This is applied to a range of different problem settings, from system identification to speech modeling.The results come in three parts. The first contribution focus on applications of deep learning to typical system identification problems, the intersection between the two areas and how they can benefit from each other. The second contribution is on hierarchical data where we promote a multiscale variational autoencoder inspired by image modeling. The final contribution is on verification of probabilistic models, in particular how to evaluate the validity of a probabilistic output, also known as calibration.
|
|
| 8. |
- Andersson Naesseth, Christian, 1986-
(författare)
-
Machine learning using approximate inference : Variational and sequential Monte Carlo methods
- 2018
-
Doktorsavhandling (övrigt vetenskapligt/konstnärligt)abstract
- Automatic decision making and pattern recognition under uncertainty are difficult tasks that are ubiquitous in our everyday life. The systems we design, and technology we develop, requires us to coherently represent and work with uncertainty in data. Probabilistic models and probabilistic inference gives us a powerful framework for solving this problem. Using this framework, while enticing, results in difficult-to-compute integrals and probabilities when conditioning on the observed data. This means we have a need for approximate inference, methods that solves the problem approximately using a systematic approach. In this thesis we develop new methods for efficient approximate inference in probabilistic models.There are generally two approaches to approximate inference, variational methods and Monte Carlo methods. In Monte Carlo methods we use a large number of random samples to approximate the integral of interest. With variational methods, on the other hand, we turn the integration problem into that of an optimization problem. We develop algorithms of both types and bridge the gap between them.First, we present a self-contained tutorial to the popular sequential Monte Carlo (SMC) class of methods. Next, we propose new algorithms and applications based on SMC for approximate inference in probabilistic graphical models. We derive nested sequential Monte Carlo, a new algorithm particularly well suited for inference in a large class of high-dimensional probabilistic models. Then, inspired by similar ideas we derive interacting particle Markov chain Monte Carlo to make use of parallelization to speed up approximate inference for universal probabilistic programming languages. After that, we show how we can make use of the rejection sampling process when generating gamma distributed random variables to speed up variational inference. Finally, we bridge the gap between SMC and variational methods by developing variational sequential Monte Carlo, a new flexible family of variational approximations.
|
|
| 9. |
- Gustafsson, Fredrik K., 1993-
(författare)
-
Towards Accurate and Reliable Deep Regression Models
- 2023
-
Doktorsavhandling (övrigt vetenskapligt/konstnärligt)abstract
- Regression is a fundamental machine learning task with many important applications within computer vision and other domains. In general, it entails predicting continuous targets from given inputs. Deep learning has become the dominant paradigm within machine learning in recent years, and a wide variety of different techniques have been employed to solve regression problems using deep models. There is however no broad consensus on how deep regression models should be constructed for best possible accuracy, or how the uncertainty in their predictions should be represented and estimated. These open questions are studied in this thesis, aiming to help take steps towards an ultimate goal of developing deep regression models which are both accurate and reliable enough for real-world deployment within medical applications and other safety-critical domains.The first main contribution of the thesis is the formulation and development of energy-based probabilistic regression. This is a general and conceptually simple regression framework with a clear probabilistic interpretation, using energy-based models to represent the true conditional target distribution. The framework is applied to a number of regression problems and demonstrates particularly strong performance for 2D bounding box regression, improving the state-of-the-art when applied to the task of visual tracking.The second main contribution is a critical evaluation of various uncertainty estimation methods. A general introduction to the problem of estimating the predictive uncertainty of deep models is first provided, together with an extensive comparison of the two popular methods ensembling and MC-dropout. A number of regression uncertainty estimation methods are then further evaluated, specifically examining their reliability under real-world distribution shifts. This evaluation uncovers important limitations of current methods and serves as a challenge to the research community. It demonstrates that more work is required in order to develop truly reliable uncertainty estimation methods for regression.
|
|
| 10. |
- Gustafsson, Stefan, et al.
(författare)
-
Development and validation of deep learning ECG-based prediction of myocardial infarction in emergency department patients
- 2022
-
Ingår i: Scientific Reports. - : Springer Nature. - 2045-2322. ; 12
-
Tidskriftsartikel (refereegranskat)abstract
- Myocardial infarction diagnosis is a common challenge in the emergency department. In managed settings, deep learning-based models and especially convolutional deep models have shown promise in electrocardiogram (ECG) classification, but there is a lack of high-performing models for the diagnosis of myocardial infarction in real-world scenarios. We aimed to train and validate a deep learning model using ECGs to predict myocardial infarction in real-world emergency department patients. We studied emergency department patients in the Stockholm region between 2007 and 2016 that had an ECG obtained because of their presenting complaint. We developed a deep neural network based on convolutional layers similar to a residual network. Inputs to the model were ECG tracing, age, and sex; and outputs were the probabilities of three mutually exclusive classes: non-ST-elevation myocardial infarction (NSTEMI), ST-elevation myocardial infarction (STEMI), and control status, as registered in the SWEDEHEART and other registries. We used an ensemble of five models. Among 492,226 ECGs in 214,250 patients, 5,416 were recorded with an NSTEMI, 1,818 a STEMI, and 485,207 without a myocardial infarction. In a random test set, our model could discriminate STEMIs/NSTEMIs from controls with a C-statistic of 0.991/0.832 and had a Brier score of 0.001/0.008. The model obtained a similar performance in a temporally separated test set of the study sample, and achieved a C-statistic of 0.985 and a Brier score of 0.002 in discriminating STEMIs from controls in an external test set. We developed and validated a deep learning model with excellent performance in discriminating between control, STEMI, and NSTEMI on the presenting ECG of a real-world sample of the important population of all-comers to the emergency department. Hence, deep learning models for ECG decision support could be valuable in the emergency department.
|
|
