Search

Flow visualization is a crucial part of the aerodynamic development process. Especially in air, deriving three-dimensional flow fields from measurements is not straightforward and requires years of experience from the measurement engineer. Addressing the issue, the work at hand presents a fully autonomous measurement system incorpor- ating multihole pressure probe-based sampling and various agents. The provided data is then used to accelerate the development pro- cess of industries such as aerospace and automotive. A topology of the complete end-to-end system is outlined, including hardware, software, and algorithmic components. Individual data instances are derived from a multihole pressure probe that samples the flow field and is tracked by a motion capture system or forward kinemat- ics of a robotic manipulator. The data model guiding the sampling process has three distinct phases, starting with an initial field ex- ploration without prior knowledge. The second phase of field ex- ploitation utilizes a sparse Gaussian process with a new strategy to include reconstructed turbulence measures in the sampling function. In a final post-processing step, the governing equations are included with a Lattice Boltzmann numerical flow field solver. Besides the domain of interest, the solver only uses data provided by the meas- urement, resulting in a simulated flow field directly controlled by the measured data. An advanced augmented reality visualization feature allows the presented system to incorporate human agents into the active learning measurement process guided by the under- lying data model. This artificial intelligence acts as a substitute for the engineer’s experience and guarantees the measurement’s quality. Further, the head-mounted display also o↵ers in-situ visualizations of the measured data and accelerates engineering decisions. In re- petitive industrial applications, a robotic agent replaces the human, posing a fully autonomous system with the domain of interest as the sole user input. The accuracy of the new measurement system is explored in three di↵erent cases inside a wind tunnel environment. Results highlight the capability of the system to quickly as- sess complex flow fields, even in close proximity to objects. In the case of the robotic agent, a proof of concept for fully autonomous three-dimensional wind tunnel measurements is given. The acceler- ated Lattice Boltzmann data assimilation shows promising results in incorporating the full governing equations into the measurements online and introduces interesting further development paths.

Current trend in neurosciences is to use naturalistic stimuli, such as cinema, class-room biology or video gaming, aiming to understand the brain functions during ecologically valid conditions. Naturalistic stimuli recruit complex and overlapping cognitive, emotional and sensory brain processes. Brain oscillations form underlying mechanisms for such processes, and further, these processes can be modified by expertise. Human cortical functions are often analyzed with linear methods despite brain as a biological system is highly nonlinear. This study applies a relatively robust nonlinear method, Higuchi fractal dimension (HFD), to classify cortical functions of math experts and novices when they solve long and complex math demonstrations in an EEG laboratory. Brain imaging data, which is collected over a long time span during naturalistic stimuli, enables the application of data-driven analyses. Therefore, we also explore the neural signature of math expertise with machine learning algorithms. There is a need for novel methodologies in analyzing naturalistic data because formulation of theories of the brain functions in the real world based on reductionist and simplified study designs is both challenging and questionable. Data-driven intelligent approaches may be helpful in developing and testing new theories on complex brain functions. Our results clarify the different neural signature, analyzed by HFD, of math experts and novices during complex math and suggest machine learning as a promising data-driven approach to understand the brain processes in expertise and mathematical cognition., Scientific Reports, 13, ISSN:2045-2322

In practice, multi-task learning (through learning features shared among tasks) is an essential property of deep neural networks (NNs). While infinite-width limits of NNs can provide good intuition for their generalization behavior, the well-known infinite-width limits of NNs in the literature (e.g., neural tangent kernels) assume specific settings in which wide ReLU-NNs behave like shallow Gaussian Processes with a fixed kernel. Consequently, in such settings, these NNs lose their ability to benefit from multi-task learning in the infinite-width limit. In contrast, we prove that optimizing wide ReLU neural networks with at least one hidden layer using ℓ 2 -regularization on the parameters promotes multitask learning due to representation-learning – also in the limiting regime where the network width tends to infinity. We present an exact quantitative characterization of this infinite width limit in an appropriate function space that neatly describes multi-task learning.

LiDAR-based localization and mapping is one of the core components in many modern robotic systems due to the direct integration of range and geometry, allowing for precise motion estimation and generation of high quality maps in real-time. Yet, as a consequence of insufficient environmental constraints present in the scene, this dependence on geometry can result in localization failure, happening in self-symmetric surroundings such as tunnels. This work addresses precisely this issue by proposing a neural network-based estimation approach for detecting (non-)localizability during robot operation. Special attention is given to the localizability of scan-to-scan registration, as it is a crucial component in many LiDAR odometry estimation pipelines. In contrast to previous, mostly traditional detection approaches, the proposed method enables early detection of failure by estimating the localizability on raw sensor measurements without evaluating the underlying registration optimization. Moreover, previous approaches remain limited in their ability to generalize across environments and sensor types, as heuristic-tuning of degeneracy detection thresholds is required. The proposed approach avoids this problem by learning from a corpus of different environments, allowing the network to function over various scenarios. Furthermore, the network is trained exclusively on simulated data, avoiding arduous data collection in challenging and degenerate, often hard-to-access, environments. The presented method is tested during field experiments conducted across challenging environments and on two different sensor types without any modifications. The observed detection performance is on par with state-of-the-art methods after environment-specific threshold tuning., ISSN:2153-0858, 2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), ISBN:978-1-6654-7927-1, ISBN:978-1-6654-7928-8

3D object detection is a central task for applications such as autonomous driving, in which the system needs to localize and classify surrounding traffic agents, even in the presence of adverse weather. In this paper, we address the problem of LiDAR-based 3D object detection under snowfall. Due to the difficulty of collecting and annotating training data in this setting, we propose a physically based method to simulate the effect of snowfall on real clear weather LiDAR point clouds. Our method samples snow particles in 2D space for each LiDAR line and uses the induced geometry to modify the measurement for each LiDAR beam accordingly. Moreover, as snowfall often causes wetness on the ground, we also simulate ground wetness on LiDAR point clouds. We use our simulation to generate partially synthetic snowy LiDAR data and leverage these data for training 3D object detection models that are robust to snowfall. We conduct an extensive evaluation using several state-of-the-art 3D object detection methods and show that our simulation consistently yields significant performance gains on the real snowy STF dataset compared to clear weather baselines and competing simulation approaches, while not sacrificing performance in clear weather. Our code is available at github.com/SysCV/LiDAR snow sim., 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), ISBN:978-1-6654-6946-3, ISBN:978-1-6654-6947-0

Das Potential digitaler Kollektionen ist enorm: Auf der einen Seite stellen Bibliotheken und andere Gedächtniseinrichtungen in hoher Qualität digitale und digitalisierte Inhalte in grossem Umfang und inhaltlicher Breite zur Verfügung. Auf der anderen Seite gibt es zahlreiche und rasch wachsende Forschungsfelder, die einen grossen Bedarf an solchen Datensätzen haben, um sie für ihre digitale Forschung zu nutzen. An dieser Schnittstelle zwischen Beständen und innovativen Formen der Nachnutzung entstehen Freiräume, in denen das Zusammenspiel von bibliothekarischen Inhalten und Services, Open Science und Co-Creation gestaltet werden kann und in denen die Bibliothek eine aktive Rolle als Kooperationspartner einnimmt. Um dieses Potential zu nutzen, richteten die Sammlungen und Archive der ETH-Bibliothek Mitte 2021 die Fachstelle Digital Scholarship Services ein. Im Zentrum des Beitrags stehen Beweggründe und Einblicke in die Praxis der Aufbauphase der Fachstelle sowie (Pilot-)Projekte aus dem Machine-Learning-Bereich. Welche Forschungsinteressen haben etwa Informatikprofessuren an Kollektionen digitalisierter Alt- oder Archivbestände? Wie können entsprechende Kooperationen initialisiert und verstetigt werden? Welche Anforderungen an Kompetenzen, Daten und Infrastruktur werden seitens der Forschenden gestellt und wo müssen auch innerhalb der Bibliothek neue Formen der Zusammenarbeit erprobt werden? Welche Mehrwerte können zu gegenseitigem Nutzen realisiert werden? Ein erstes Fazit zum Beitrag der Digital Scholarship Services zum Fokus, bestehende Inhalte KI-gestützt anzureichern und damit neue Arten der Nutzung zu ermöglichen, rundet den Beitrag ab., o-bib. Das offene Bibliotheksjournal, 9 (4), ISSN:2363-9814

Simulations of complex systems are essential for applications ranging from weather forecasting to molecular systems and drug design. The veracity of the resulting predictions hinges on their capacity to capture the un- derlying system dynamics. Massively parallel simulations performed in High-Performance Computing (HPC) clusters capture these dynamics by resolving all spatio-temporal scales. The computational cost is often pro- hibitive for experimentation or optimization, while their findings might not allow for generalization. The design of dimensionality reduction methods and fast data-driven reduced-order models or surrogates have been matters of life-long research efforts. In the first part of this thesis, we focus on the design and training of data-driven recurrent neural networks for forecasting the spatio-temporal dynamics of high-dimensional and reduced-order complex systems. We propose architectural advances and training algorithms that alleviate the pitfalls of previously proposed methods, whose application was limited to lower-order systems. The designed algorithms extend the arsenal of predictive models for complex systems and spatio-temporal chaos. In the second part, we present a novel systematic framework that bridges large-scale simulations and reduced-order models to Learn the Effective Dynamics (LED) of complex systems. The framework forms algorithmic alloys between non-linear machine learning algorithms and the Equation- Free approach for modeling complex systems exhibiting spatio-temporal chaos. LED deploys autoencoders to map between fine- and coarse-grained representations and evolves the latent space dynamics using recurrent neural networks. The algorithm is validated on benchmark problems, and we find that it outperforms state-of-the-art reduced-order models in terms of predictability and large-scale simulations in terms of cost. LED is applicable to systems ranging from chemistry to fluid mechanics and reduces the computational effort by up to two orders of magnitude while maintaining the prediction accuracy of the full system dynamics. We argue that LED constitutes a potent novel modality for the accurate prediction of complex systems.

Tripeptides of the Pro-Pro-Xaa type are highly reactive and stereoselective organocatalysts for conjugate addition reactions. This thesis investigated the use of tripeptides to overcome catalyst deactivation, caused by highly reactive nitroolefins (Chapters 3 and 4), and challenged chemoinformatic workflows, to predict and optimize stereoselective tripeptide catalysts (Chapters 5 and 6). In the first part of the thesis, we investigated H-DPro-Pro-Glu-NH2 and related peptides in the conjugate addition reactions of aldehydes with fluorinated nitroolefins and nitroacrylates. While these electrophiles typically deactivate secondary amine catalysts by N-alkylation, mechanistic studies revealed that this deactivation is reversible for Pro-Pro-Glu type peptides, due to the presence of a well-positioned proton donor in the catalyst. Detailed kinetic analysis showed that the N-alkylation proceeds slow compared to the enamine-formation, and that the formation of the enamine-intermediate is the rate-limiting step. We identified H-DPro-αMePro-Glu-NH2 as highly reactive and stereoselective catalyst for the conjugate addition reaction of aldehydes with fluorinated nitroolefins and obtained a broad substrate scope and chiral downstream derivatives. These reactions were carried out with only 0.5 mol% catalyst loading and provided the products in generally high yields and stereoselectivities. Furthermore, we identified H-(4S)Mep-DPro-DGlu-NH2 as even more stereoselective catalyst than H-DPro-Pro-Glu-NH2 for the reaction of aldehydes with nitroacrylates and provided the first comprehensive substrate scope for this reaction, with catalyst loadings of as little as 0.05 mol%. These low catalyst loadings for reactions with highly electron-deficient nitroolefins are remarkable and highlight the excellent reactivity of Pro-Pro-Glu type peptides, and the value of thorough mechanistic and kinetic investigations. In the second part of this thesis, we challenged state-of-the-art machine learning methods to predict and optimize conformationally flexible tripeptide organocatalysts. First, we showed that machine learning algorithms can be trained to predict the same highly selective tripeptides of the Pro-Pro-Glu type that were previously developed by rational design. Next, we constructed a in silico library with >30000 tripeptide catalysts and synthesized a representative subset of 161 peptides. We tested these peptides on a dienamine-mediated model reactions and obtained a broad experimental data set. This first round of experiments also revealed a stereoselective Pro-Pro-Glu type catalyst that, however, could only be marginally optimized over four consecutive rounds of supervised machine learning. Thus, this work highlights the current limitations of machine learning methods and emphasizes the need for more sophisticated chemoinformatic workflows. We also addressed the question whether and how seven electronegative Cγ-substituents at the N-terminal proline of H-Pro-DPro-DGlu-NH2 influence the catalyst reactivity and stereoselectivity. First, proline methyl ester model compounds were computationally modelled by the two independent programs MacroModel and CREST. Comparison of the predicted conformational and electronic properties showed a high linear correlation between both programs and experimental validation with dihedral angles, determined by NMR spectroscopy for hydroxyproline methyl esters in solution, were in good agreement with the computational data. In a second step, we synthesized the Cγ-substituted tripeptides and tested them as catalysts in two model reactions. While we observed a strong influence of the Cγ-substituents on the catalyst performance, we could not see a clear relationship between the computationally predicted properties and the reaction outcome. Thus, this research emphasizes that isolated properties predicted by molecular modeling do not suffice to explain the complex influence of Cγ-substituents on tripeptide catalysis., Tripeptide vom Typ Pro-Pro-Xaa sind hochreaktive und stereoselektive Organokatalysatoren für konjugierte Additionsreaktionen. In dieser Arbeit wurde der Einsatz von Tripeptiden zur Überwindung der durch hochreaktive Nitroolefine verursachten Katalysatordeaktivierung untersucht (Kapitel 3 und 4) und chemoinformatische Arbeitsabläufe zur Vorhersage und Optimierung stereoselektiver Tripeptidkatalysatoren (Kapitel 5 und 6) getestet. Im ersten Teil der Arbeit untersuchten wir H-DPro-Pro-Glu-NH2 und verwandte Peptide in den konjugierten Additionsreaktionen von Aldehyden mit fluorierten Nitroolefinen und Nitroacrylaten. Während diese Elektrophile normalerweise sekundäre Aminkatalysatoren durch N-Alkylierung deaktivieren, zeigten mechanistische Studien, dass diese Deaktivierung für Peptide des Typs Pro-Pro-Glu reversibel ist, was auf das Vorhandensein eines gut positionierten Protonendonors im Katalysator zurückzuführen ist. Detaillierte kinetische Analysen zeigten, dass die N-Alkylierung im Vergleich zur Enaminbildung langsam verläuft und dass die Bildung des Enamin-Zwischenprodukts der geschwindigkeitsbestimmende Schritt ist. Wir identifizierten H-DPro-αMePro-Glu-NH2 als hochreaktiven und stereoselektiven Katalysator für die konjugierte Additionsreaktion von Aldehyden mit fluorierten Nitroolefinen und synthetisierten ein breites Substratspektrum und davon abgeleitete chirale Verbindungen. Diese Reaktionen wurden mit nur 0.5 mol% Katalysatorladung durchgeführt und wir erhielten die Produkte in allgemein hohen Ausbeuten und Stereoselektivitäten. Darüber hinaus identifizierten wir H-(4S)Mep-DPro-DGlu-NH2 als noch stereoselektiveren Katalysator als H-DPro-Pro-Glu-NH2 für die Reaktion von Aldehyden mit Nitroacrylaten und synthetisierten das erste umfassende Substratspektrum für diese Reaktion, mit Katalysatorbeladungen von nur 0.05 mol%. Diese niedrigen Katalysatorbeladungen für Reaktionen mit hoch elektronenarmen Nitroolefinen sind bemerkenswert und unterstreichen die ausgezeichnete Reaktivität von Peptiden des Typs Pro-Pro-Glu, sowie den Wert gründlicher mechanistischer und kinetischer Untersuchungen. Im zweiten Teil dieser Arbeit haben wir aktuelle Methoden des maschinellen Lernens eingesetzt um Tripeptid-Organokatalysatoren, welche eine hohe Flexibilität der Konformation aufweisen, vorherzusagen und zu optimieren. Zunächst haben wir gezeigt, dass Algorithmen des maschinellen Lernens so trainiert werden können, dass sie dieselben hochselektiven Tripeptide vom Typ Pro-Pro-Glu vorhersagen, die zuvor durch rationales Katalysatordesign entwickelt wurden. Anschließend erstellten wir eine in-silico-Bibliothek mit mehr als 30000 Tripeptid-Katalysatoren und synthetisierten eine repräsentative Teilmenge von 161 Peptiden. Wir testeten diese Peptide an einer Dienamin-vermittelten Modellreaktion und erhielten einen breiten experimentellen Datensatz. In dieser ersten Versuchsrunde wurde ein stereoselektiver Katalysator vom Typ Pro-Pro-Glu entdeckt, der in vier aufeinanderfolgenden Runden des überwachten maschinellen Lernens nur geringfügig optimiert werden konnte. Diese Arbeit zeigt die derzeitigen Grenzen maschineller Lernmethoden auf und unterstreicht den Bedarf an weiterentwickelten chemoinformatischen Arbeitsabläufen. Wir gingen auch der Frage nach, ob und wie sieben elektronegative Cγ-Substituenten am N-terminalen Prolin von H-Pro-DPro-DGlu-NH2 die Katalysatorreaktivität und Stereoselektivität beeinflussen. Zunächst wurden Prolin-Methylester-Modellverbindungen mit den beiden unabhängigen Programmen MacroModel und CREST rechnerisch modelliert. Der Vergleich der vorhergesagten konformationellen und elektronischen Eigenschaften zeigte eine hohe lineare Korrelation zwischen beiden Programmen, und die experimentelle Validierung mit den durch NMR-Spektroskopie für Hydroxyprolinmethylester in Lösung bestimmten Flächenwinkeln stimmte gut mit den berechneten Daten überein. In einem zweiten Schritt haben wir die Cγ-substituierten Tripeptide synthetisiert und sie als Katalysatoren in zwei Modellreaktionen getestet. Während wir einen starken Einfluss der Cγ-Substituenten auf die Katalysatorleistung feststellten, konnten wir keine klare Beziehung zwischen den rechnerisch vorhergesagten Eigenschaften und dem Reaktionsergebnis erkennen. Diese Forschungsarbeit unterstreicht somit, dass isolierte, durch molekulare Modellierung vorhergesagte Eigenschaften nicht ausreichen, um den komplexen Einfluss von Cγ-Substituenten auf die Tripeptidkatalyse zu erklären.

Autonomous driving has been one of the most challenging and exciting research topics in the last years inspiring research and development in both universities and companies. The presented work contributes to this active area of research by means of environment perception with a focus on visual cues. The task is to provide the relevant information of the surrounding scene to decision making modules of the intelligent vehicle. Typically, several different sensors such as camera, Light Detection and Ranging (LIDAR) and radar sensors are fused to that end, which enables addressing tasks like object recognition and tracking, detection of drivable space, and many more. This thesis argues that these tasks require access to information close to the raw sensor signal in order to deliver accurate and robust scene interpretations. Accessing this raw data in each task separately, in particular in the case of high resolution visual data, however, induces high computational costs in terms of data transfer and processing. A design that processes this data in each task, furthermore, prevents from tracing errors to its root cause and makes assessing the performance of the vehicle’s perception abilities difficult and data intensive. This holds particularly true in the case of DNN based processing. To overcome this problem, Stixels are introduced which constitute a mid-level representation. This mid-level approach effectively combines advantages of both: first, a compactness close to traditional high-level approaches such as lists of objects as well as second, an accurate and complete scene representation comparable to the raw data itself. To that end, a Stixel semi-CRF is developed that harmonizes several attempts to formalize and simplify the Stixel model and combines complementary semantic and depth information on pixel-level in a sound probabilistic formulation. Furthermore, several generalizations and extensions are introduced which allow for the accurate representation of arbitrary complex scenarios. The combination of the complementary cues from the semantic and depth domain in the joint optimization process which leverages the strength of the individual inputs allows to even surpass the accuracy of the raw inputs. The pixel-level semantic input used in the Stixel representation is obtained using semantic segmentation, a very active area of research in computer vision. The focus for the methods in this area, however, has been on methods using solely the color information delivered by the camera. In this work, it is proposed to apply the same principle, i.e. combining geometric and appearance cues on a low-level of processing, already in this task. To that end, a novel CNN architecture is proposed which brings both inputs together within the network. This turns out to be superior compared to the standard approaches which either concatenate the inputs directly or combine the results of two separate networks at the very end. In addition to the semantic segmentation input, the Stixel representation builds on dense depth maps as input which we obtain using stereo vision naturally requiring a pair of cameras. As alternative variant for the depth estimation, this dissertation proposes to combine a single camera with a LIDAR sensor which delivers sparse but extremely accurate depth estimates. In order to generate dense depth maps using the sparse measurements, a depth completion method is introduced which uses the color image as guidance in order to distribute the depth information throughout the image. This task is addressed by defining an energy function which considers all measurements within the image in order to infer planar elements and incorporates smoothness assumptions in a local neighborhood of every pixel. This neighborhood within the image turns out to be a crucial part within the optimization. This is why a novel semantic and edge aware geodesic distance is introduced which leverages the available appearance information and is thus able to respect the borders of objects. In doing so, the proposed method is able to reconstruct scenes accurately including even fine structures such as poles or side mirrors with no or only few associated LIDAR measurements. Furthermore, an efficient inference algorithm is introduced which allows for real-time depth estimation using even high resolution images. Thorough experiments of all proposed methods on real-data benchmarks demonstrate the benefits of the novel approaches and examine their most important design choices, parameters and properties. Finally, this thesis demonstrates the usage of the proposed Stixel representation in DNN based processing on fusion-level. To that end, Stixels are considered a generalized LIDAR point cloud which enables to apply an adapted variant of the PointNet(C. R. Qi, Su, Mo, et al. 2017) to the data in order to segment object instances. A method is developed to create Stixel-level groundtruth from pixel-level instance segmentation annotations which is used to train the proposed model. Despite the intended compression and quantization effects of the Stixels, the experiments reveal promising results opening the door for future research on Stixel-level fusion.

Robots hold the promise of becoming useful helpers for many dangerous, laborious, or unpleasant tasks. One significant barrier to the ubiquitous use of robotic assistants today is the difficulty of programming sufficiently intelligent behavior that would allow for autonomous operation. The present thesis focuses on the challenges relating to automatic motion planning and control, which is an essential aspect of the autonomous behavior of any mobile robot. We showcase our developed solutions primarily on quadrupedal walking robots, as legged locomotion offers the necessary versatility to accomplish tasks in many natural and human-made environments. The complexity of walking robots calls for an automatic control approach that handles high-dimensional and nonlinearly coupled dynamics with minimal human involvement. In this work, we derive several theoretical insights and subsequently develop practical algorithms for legged locomotion that are applicable to a large variety of systems. Contrary to many state-of-the-art methods that separate the planning and control aspects, our approach takes a holistic view of the locomotion problem in order to use the entire dynamics of a system. We leverage both optimal control and machine learning techniques to efficiently generate whole-body walking motions in real-time without relying on preceding footstep planners. In the first chapters of this thesis, we devise an optimal control approach for generating motion plans that utilize the system's full dynamic capabilities and do not require a pre-specified contact sequence. We demonstrate that a hard contact model can be incorporated into an efficient optimization scheme to form a contact-implicit planning algorithm. The resulting motions are physically feasible and can exploit sliding contacts if it is expedient for the task. Expanding on the idea of creating trajectories under realistic dynamical constraints, we develop a second algorithm based on a stochastic system model. The ensuing method can discover stable walking motions in environments with only sparse footholds by sampling feasible trajectories. Unlike other stochastic optimal control approaches, our method can handle constraints like standard optimal control while also leveraging data-driven methods to improve sampling efficiency. Through conducting hardware experiments we verify the performance of both algorithms under realistic conditions. One of the practical challenges of our (stochastic) optimal control schemes is the computational complexity at runtime. The necessary powerful onboard computers can cause significant energy expenditure and reduce the payload of the system. To address this issue, the last part of this thesis presents a method for extracting the capabilities of an optimal control algorithm into a neural network. We show that optimal control theory naturally suggests a suitable loss function for policy search that can be minimized with standard machine learning tools. The procedure shifts the computational burden of optimal control into an offline training phase while online control is reduced to a mere evaluation of the learned policy. Our experiments show that the extracted policy exhibits similar performance as the original optimal controller.

The increased focus on evidence-based practice in the health sciences led to a plethora of (un)organised and digitised data. In conjunction with the availability of technological advances in the life sciences, this resulted in extraordinary access to biomedical data. Due to efficient measurement devices, the frequency at which data can be obtained is at an unprecedented high, leading to the adage that data, indeed, could be the new gold. Examples of such high-resolution time series data are the continuous monitoring of patient vital parameters or a single electrocardiogram (ECG) itself. The temporal component introduced by time series data is both a chance and a challenge, necessitating the development of appropriate data analysis techniques. A chance, as it allows us to utilise a measurement’s temporal evolution to characterise or classify the object of interest (e.g. patients, cells, or other organisms). A challenge because local and global correlation structures exacerbate obtaining a complete picture of a time-evolving phenomenon. Moreover, many dynamically changing systems exhibit alterations that occur at multiple scales and in multiple channels. This thesis presents a set of novel methods to help characterise and classify time-varying data with the express purpose of answering questions at the intersection of machine learning and healthcare. Recognising that time series arise from different categories, we first separate them into real-valued and object-valued time series and investigate both types separately. For the analysis of the first type, we propose a novel method to mine time series patterns efficiently. Driven by a statistical approach, we will introduce a way to identify temporal biomarkers and illustrate their utility in a data set of intensive care unit patients. For this, we leverage the expressive power of subsequences to obtain a high-dimensional time series representation. This feature representation is subsequently used to develop a kernel method based on optimal transport theory. The developed algorithm is of general applicability for medium-sized data sets and has proven particularly effective in the classification setting. The first part of this thesis ends with the presentation of a collaborative machine learning system to predict myocardial ischaemia from stress test ECGs. We develop a deep learning-based approach to significantly reduce the number of patients that unnecessarily undergo myocardial perfusion imaging. A subsequent interpretability analysis presents a potential path towards explainable and trustworthy artificial intelligence in cardiology. The second part of this thesis describes an effort to improve our understanding of artificial neural networks. By treating the network as a composition of time-varying graphs, we develop a method that characterises the change of its structural complexity over time. Our method captures the benefit of deep-learning best practices and can be used as an earlystopping criterion without the need for a validation data set. We thus manage to improve our understanding of artificial neural networks and shed light on the properties linked to their generalisation capabilities. Throughout this thesis, we demonstrate and highlight that in the analysis of (biomedical) time series, it is crucial to take the end-user into account. Interpretability and statistical analyses are of utter importance to make the otherwise opaque field of machine learning transparent to clinicians, physicians, and biologists. Moreover, we also hold up the mirror to ourselves as machine learning researchers: Comprehending the underlying mechanisms of our algorithms is at least as important as their empirical successes. The present thesis paves the path towards a better understanding of artificial neural networks and sheds light on complex phenotypes such as sepsis and myocardial ischaemia in clinically relevant ways.

This work addresses the challenging task of LiDAR-based 3D object detection in foggy weather. Collecting and annotating data in such a scenario is very time, labor and cost intensive. In this paper, we tackle this problem by simulating physically accurate fog into clear-weather scenes, so that the abundant existing real datasets captured in clear weather can be repurposed for our task. Our contributions are twofold: 1) We develop a physically valid fog simulation method that is applicable to any LiDAR dataset. This unleashes the acquisition of large-scale foggy training data at no extra cost. These partially synthetic data can be used to improve the robustness of several perception methods, such as 3D object detection and tracking or simultaneous localization and mapping, on real foggy data. 2) Through extensive experiments with several state-of-the-art detection approaches, we show that our fog simulation can be leveraged to significantly improve the performance for 3D object detection in the presence of fog. Thus, we are the first to provide strong 3D object detection baselines on the Seeing Through Fog dataset. Our code is available at www. trace. ethz. ch/lidar_fog_simulation., 2021 IEEE/CVF International Conference on Computer Vision (ICCV), ISBN:978-1-6654-2812-5, ISBN:978-1-6654-2813-2

We might be drinking water for an average of 2 hours daily and sleeping for an average of 8 hours daily but we breathe in the air for 24 hours every day. The air surrounding us does not only consist of gases like nitrogen, oxygen, carbon dioxide or water vapor, but also of large quantities of small particles, called aerosols. The atmospheric aerosols play an important role in climate forcing by affecting the radiative balance, cloud formation, and they affect human health. The fate of the aerosols is spatial and temporal variable, depending on the meteorological, natural and anthropogenic conditions. One of the causes for our limited understanding of aerosols is that characterization of aerosols has mostly been implemented via offline methods for decades. These particles are sampled on a filter and these filtrates, i.e. aerosols of different sizes and respective chemical composition are analyzed via heterogeneous extraction/evaporation and later post-treatments with thermalization, ionization and optical excitation (spectroscopy). All these intermediate post-treatment steps alter to a certain extent the chemical composition and morphological information of the true aerosols in our atmosphere. Recent technical advancements for real-time measurements have improved our capabilities towards a more complete chemical characterization of our ambient aerosols. For example, an extractive electrospray ionization coupled to a time-of-flight mass spectrometer (EESI-TOF) can detect most of the organic species in the aerosols with minimal thermal and ionization-induced fragmentation at a time resolution of 1 Hz with a detection limit of 1 ng m-3. Despite its very high time resolution and low detection limit, its low mass resolution affects molecular identification accuracy and it does not have the capability in resolving the structural information of the chemical species in the aerosols. To solve these limitations, this Ph.D. thesis work started with the development of coupling an EESI to an Orbitrap (EESI-Orbitrap) with a mass resolution that is at least 10 times higher than the EESI-TOF with a similar detection limit and time resolution, for higher accuracy of molecular identification. Besides that, its new capability of tandem mass spectrometry was used for the elucidation of molecular structure and possible sources of secondary organic aerosols. In the second part of this Ph.D. thesis, a technique was developed to detect water-soluble metallic particles using an EESI-TOF. This newly developed technique has similar or lower detection limits than the benchmarking technique of inductively coupled plasma mass spectrometry. Its real-time (seconds) and near-molecular detection capabilities are extremely useful for the investigation of aerosol chemistry in complex emissions. Due to the debatable extraction mechanism of EESI, two different methods for the quantification of detected aerosols were thoroughly investigated. For the first method, the extraction mechanism of EESI was investigated using carefully designed experiments for different sizes, mixing states and compositions of the aerosol. Size-dependent sensitivity was shown to be a fundamental character of the coagulation-like mechanism between the charged electrospray droplets and the particles, elucidating the underlying physics for the partial coalescence that was not reported previously. For the second method, the quantification of measured aerosol for EESI was cross-calibrated by a known gas-phase mass spectrometry technique (Vocus-PTR). First, flow tube modeling was deployed to validate the gas and particle-phase measurements using estimated saturation vapor pressure. Since the saturation vapor pressure estimation could be uncertain due to the non-ideality of the carbon-oxygen and molecular structure, exhaustive feature engineering for regression analysis on the cross-calibrations was thoroughly implemented. The resultant regression analyses provide sensitivity parameterizations for single and mixed SOA systems that can be useful for future quantification prediction. After perfecting the EESI technique for molecular identification at 10 times higher accuracy, structural information elucidation, chemical completeness coverage for water-soluble metallic aerosols, extraction mechanism elucidation and quantification parametrization, the final development stage of this Ph.D. thesis was to extend this technique for a combined measurement of gas and particle phase, namely a Dual-Phase-EESI-TOF (Dual-EESI). This final stage of development is to resolve the chemical evolution of gas and particle in our ambient using a unified ionization scheme. After thorough Dual-EESI inlet optimization and characterization, the method was tested in different SOA systems and shows its capability of having higher chemical coverage in comparison to some well-known gas-phase spectrometry techniques without sacrificing linearity and sensitivity. In addition, the quantification response between gas and particle phase for Dual-EESI was demonstrated for the first time. Three new method developments in this Ph.D. thesis based on extractive electrospray ionization have enabled more holistic characterization and accurate near-molecular identification of our ambient gases and aerosols in real-time. The quantification capability of these methods was supported by two unprecedented characterizations of the EESI presented here. The main goal of these five presented works is to allow detection of organic, inorganic and metal species in our ambient, hoping to fill our knowledge gaps in elucidating sources and physicochemical interactions among these species in the gas and particle phase.

El proyecto de grado titulado Aprendizaje de movimientos en un robot humanoide por medio de imitación implementa algoritmos de aprendizaje supervisado, proporcionando a un robot la capacidad de identificar movimientos realizado en las extremidades superiores por una persona y replicarlos. Se evalúa, además, el desempeño cuantitativo de las arquitecturas implementadas. Para desarrollar el aprendizaje de movimientos se selecciona una plataforma robótica llamada Poppy Torso, correspondiente a un proyecto de código abierto. Se construye este robot por medio de impresión 3D de las piezas y su posterior ensamblaje. Dicho desarrollo contiene además la descripción virtual del robot, lo cual permite realizar simulación en herramientas como Rviz. El desarrollo del proyecto se divide en dos etapas: la primera de ellas consiste en lograr la clasificación de los movimientos humanos. Por esta razón, se crea un dataset de videos RGB-D (RGB más profundidad), que contiene movimientos de las extremidades superiores, realizados por un grupo de personas con características diferentes. Empleando un modelo extractor de poses, se extrae la ubicación espacial de las articulaciones, que sirve como entrada a una red neuronal recurrente (RNN-many to one) y una SVM (Support Vector Machine). Se evalúa el desempeño de estos algoritmos por medio de matrices de confusión. La segunda etapa consiste en un problema de regresión que se basa en replicar la ejecución del movimiento en el robot. Debido a esto, se crea un dataset que contiene las trayectorias de ángulos que toma cada motor, para que el robot desplace sus articulaciones de un punto aleatorio a uno definido. Para el aprendizaje supervisado se implementa una red neuronal recurrente (RNN-one to many), capaz de predecir las trayectorias de cada una de las articulaciones del robot para llegar a un punto objetivo. Los algoritmos de clasificación se evalúan por medio de matrices de confusión, donde se obtienen precisiones de 97\% en el caso de la SVM y 92% para la RNN. Por otro lado, el algoritmo de regresión se evalúa con error de posición final, donde se obtienen errores entre 9.32 x 10^(-8) y 4.21x 10^(-3) radianes respecto a la posición objetivo. Por el valor de estas métricas y el comportamiento de los algoritmos al predecir, se puede concluir que el aprendizaje en los algoritmos de clasificación y regresión se desarrolla adecuadamente, cumpliendo con los objetivos propuestos. The degree project titled Learning movements in a humanoid robot through imitation implements supervised learning algorithms, providing a robot with the ability to identify movements made in the upper limbs by a person and replicate them. The quantitative performance of the implemented architectures is also evaluated. To develop movement learning, a robotic platform called Poppy Torso is selected, corresponding to an open source project. This robot is built by means of 3D printing of the parts and their subsequent assembly. This development also contains the virtual description of the robot, which allows simulation in tools such as Rviz. The development of the project is divided into two stages: the first of them consists of achieving the classification of human movements. For this reason, an RGB-D (RGB plus depth) video dataset is created, containing movements of the upper limbs, performed by a group of people with different characteristics. Using a pose extractor model, the spatial location of the joints is extracted, which serves as input to a recurrent neural network (RNN-many to one) and an SVM (Support Vector Machine). The performance of these algorithms is evaluated by means of confusion matrices. The second stage consists of a regression problem that is based on replicating the execution of the movement in the robot. Due to this, a dataset is created that contains the trajectories of angles that each motor takes, so that the robot moves its joints from a random point to a defined one. For supervised learning, a recurrent neural network (RNN-one to many) is implemented, capable of predicting the trajectories of each of the robot's joints to reach a target point. The classification algorithms are evaluated by means of confusion matrices, where accuracies of 97 \% are obtained for the SVM and 92% for the RNN. On the other hand, the regression algorithm is evaluated with final position error, where errors between 9.32 x 10 ^ (- 8) and 4.21x 10 ^ (- 3) radians are obtained with respect to the target position. Due to the value of these metrics and the behavior of the algorithms when predicting, it can be concluded that the learning in the classification and regression algorithms develops adequately, fulfilling the proposed objectives. Ingeniero Electronico http://unidadinvestigacion.usta.edu.co Pregrado