Refine
Has Fulltext
- yes (16) (remove)
Document Type
- Doctoral Thesis (10)
- Postprint (4)
- Monograph/Edited Volume (1)
- Master's Thesis (1)
Language
- English (16)
Is part of the Bibliography
- yes (16)
Keywords
- deep learning (16) (remove)
Institute
- Hasso-Plattner-Institut für Digital Engineering GmbH (4)
- Institut für Geowissenschaften (3)
- Institut für Umweltwissenschaften und Geographie (3)
- Department Linguistik (2)
- Fachgruppe Betriebswirtschaftslehre (1)
- Institut für Physik und Astronomie (1)
- Strukturbereich Kognitionswissenschaften (1)
- Wirtschaftswissenschaften (1)
As an emerging sub-field of music information retrieval (MIR), music imagery information retrieval (MIIR) aims to retrieve information from brain activity recorded during music cognition–such as listening to or imagining music pieces. This is a highly inter-disciplinary endeavor that requires expertise in MIR as well as cognitive neuroscience and psychology. The OpenMIIR initiative strives to foster collaborations between these fields to advance the state of the art in MIIR. As a first step, electroencephalography (EEG) recordings of music perception and imagination have been made publicly available, enabling MIR researchers to easily test and adapt their existing approaches for music analysis like fingerprinting, beat tracking or tempo estimation on this new kind of data. This paper reports on first results of MIIR experiments using these OpenMIIR datasets and points out how these findings could drive new research in cognitive neuroscience.
In this era of high-speed informatization and globalization, online education is no longer an exquisite concept in the ivory tower, but a rapidly developing industry closely relevant to people's daily lives. Numerous lectures are recorded in form of multimedia data, uploaded to the Internet and made publicly accessible from anywhere in this world. These lectures are generally addressed as e-lectures. In recent year, a new popular form of e-lectures, the Massive Open Online Courses (MOOCs), boosts the growth of online education industry and somehow turns "learning online" into a fashion.
As an e-learning provider, besides to keep improving the quality of e-lecture content, to provide better learning environment for online learners is also a highly important task. This task can be preceded in various ways, and one of them is to enhance and upgrade the learning materials provided: e-lectures could be more than videos. Moreover, this process of enhancement or upgrading should be done automatically, without giving extra burdens to the lecturers or teaching teams, and this is the aim of this thesis.
The first part of this thesis is an integrated framework of multi-lingual subtitles production, which can help online learners penetrate the language barrier. The framework consists of Automatic Speech Recognition (ASR), Sentence Boundary Detection (SBD) and Machine Translation (MT), among which the proposed SBD solution is major technical contribution, building on Deep Neural Network (DNN) and Word Vector (WV) and achieving state-of-the-art performance. Besides, a quantitative evaluation with dozens of volunteers is also introduced to measure how these auto-generated subtitles could actually help in context of e-lectures.
Secondly, a technical solution "TOG" (Tree-Structure Outline Generation) is proposed to extract textual content from the displaying slides recorded in video and re-organize them into a hierarchical lecture outline, which may serve in multiple functions, such like preview, navigation and retrieval. TOG runs adaptively and can be roughly divided into intra-slide and inter-slides phases. Table detection and lecture video segmentation can be implemented as sub- or post-application in these two phases respectively. Evaluation on diverse e-lectures shows that all the outlines, tables and segments achieved are trustworthily accurate.
Based on the subtitles and outlines previously created, lecture videos can be further split into sentence units and slide-based segment units. A lecture highlighting process is further applied on these units, in order to capture and mark the most important parts within the corresponding lecture, just as what people do with a pen when reading paper books. Sentence-level highlighting depends on the acoustic analysis on the audio track, while segment-level highlighting focuses on exploring clues from the statistical information of related transcripts and slide content. Both objective and subjective evaluations prove that the proposed lecture highlighting solution is with decent precision and welcomed by users.
All above enhanced e-lecture materials have been already implemented in actual use or made available for implementation by convenient interfaces.
Moving beyond ERP components
(2018)
Relationships between neuroimaging measures and behavior provide important clues about brain function and cognition in healthy and clinical populations. While electroencephalography (EEG) provides a portable, low cost measure of brain dynamics, it has been somewhat underrepresented in the emerging field of model-based inference. We seek to address this gap in this article by highlighting the utility of linking EEG and behavior, with an emphasis on approaches for EEG analysis that move beyond focusing on peaks or "components" derived from averaging EEG responses across trials and subjects (generating the event-related potential, ERP). First, we review methods for deriving features from EEG in order to enhance the signal within single-trials. These methods include filtering based on user-defined features (i.e., frequency decomposition, time-frequency decomposition), filtering based on data-driven properties (i.e., blind source separation, BSS), and generating more abstract representations of data (e.g., using deep learning). We then review cognitive models which extract latent variables from experimental tasks, including the drift diffusion model (DDM) and reinforcement learning (RL) approaches. Next, we discuss ways to access associations among these measures, including statistical models, data-driven joint models and cognitive joint modeling using hierarchical Bayesian models (HBMs). We think that these methodological tools are likely to contribute to theoretical advancements, and will help inform our understandings of brain dynamics that contribute to moment-to-moment cognitive function.
Medical imaging plays an important role in disease diagnosis, treatment planning, and clinical monitoring. One of the major challenges in medical image analysis is imbalanced training data, in which the class of interest is much rarer than the other classes. Canonical machine learning algorithms suppose that the number of samples from different classes in the training dataset is roughly similar or balance. Training a machine learning model on an imbalanced dataset can introduce unique challenges to the learning problem.
A model learned from imbalanced training data is biased towards the high-frequency samples. The predicted results of such networks have low sensitivity and high precision. In medical applications, the cost of misclassification of the minority class could be more than the cost of misclassification of the majority class. For example, the risk of not detecting a tumor could be much higher than referring to a healthy subject to a doctor. The current Ph.D. thesis introduces several deep learning-based approaches for handling class imbalanced problems for learning multi-task such as disease classification and semantic segmentation.
At the data-level, the objective is to balance the data distribution through re-sampling the data space: we propose novel approaches to correct internal bias towards fewer frequency samples. These approaches include patient-wise batch sampling, complimentary labels, supervised and unsupervised minority oversampling using generative adversarial networks for all.
On the other hand, at algorithm-level, we modify the learning algorithm to alleviate the bias towards majority classes. In this regard, we propose different generative adversarial networks for cost-sensitive learning, ensemble learning, and mutual learning to deal with highly imbalanced imaging data.
We show evidence that the proposed approaches are applicable to different types of medical images of varied sizes on different applications of routine clinical tasks, such as disease classification and semantic segmentation. Our various implemented algorithms have shown outstanding results on different medical imaging challenges.
Earthquake swarms are characterized by large numbers of events occurring in a short period of time within a confined source volume and without significant mainshock aftershock pattern as opposed to tectonic sequences. Intraplate swarms in the absence of active volcanism usually occur in continental rifts as for example in the Eger Rift zone in North West Bohemia, Czech Republic. A common hypothesis links event triggering to pressurized fluids. However, the exact causal chain is often poorly understood since the underlying geotectonic processes are slow compared to tectonic sequences. The high event rate during active periods challenges standard seismological routines as these are often designed for single events and therefore costly in terms of human resources when working with phase picks or computationally costly when exploiting full waveforms.
This methodological thesis develops new approaches to analyze earthquake swarm seismicity as well as the underlying seismogenic volume. It focuses on the region of North West (NW) Bohemia, a well studied, well monitored earthquake swarm region.
In this work I develop and test an innovative approach to detect and locate earthquakes using deep convolutional neural networks. This technology offers great potential as it allows to efficiently process large amounts of data which becomes increasingly important given that seismological data storage grows at increasing pace. The proposed deep neural network trained on NW Bohemian earthquake swarm records is able to locate 1000 events in less than 1 second using full waveforms while approaching precision of double difference relocated catalogs. A further technological novelty is that the trained filters of the deep neural network’s first layer can be repurposed to function as a pattern matching event detector without additional training on noise datasets. For further methodological development and benchmarking, I present a new toolbox to generate realistic earthquake cluster catalogs as well as synthetic full waveforms of those clusters in an automated fashion. The input is parameterized using constraints on source volume geometry, nucleation and frequency-magnitude relations. It harnesses recorded noise to produce highly realistic synthetic data for benchmarking and development. This tool is used to study and assess detection performance in terms of magnitude of completeness Mc of a full waveform detector applied to synthetic data of a hydrofracturing experiment at the Wysin site, Poland.
Finally, I present and demonstrate a novel approach to overcome the masking effects of wave propagation between earthquake and stations and to determine source volume attenuation directly in the source volume where clustered earthquakes occur. The new event couple spectral ratio approach exploits high frequency spectral slopes of two events sharing the greater part of their rays. Synthetic tests based on the toolbox mentioned before show that this method is able to infer seismic wave attenuation within the source volume at high spatial resolution. Furthermore, it is independent from the distance towards a station as well as the complexity of the attenuation and velocity structure outside of the source volume of swarms. The application to recordings of the NW Bohemian earthquake swarm shows increased P phase attenuation within the source volume (Qp < 100) based on results at a station located close to the village Luby (LBC). The recordings of a station located in epicentral proximity, close to Nový Kostel (NKC), show a relatively high complexity indicating that waves arriving at that station experience more scattering than signals recorded at other stations. The high level of complexity destabilizes the inversion. Therefore, the Q estimate at NKC is not reliable and an independent proof of the high attenuation finding given the geometrical and frequency constraints is still to be done. However, a high attenuation in the source volume of NW Bohemian swarms has been postulated before in relation to an expected, highly damaged zone bearing CO 2 at high pressure.
The methods developed in the course of this thesis yield the potential to improve our understanding regarding the role of fluids and gases in intraplate event clustering.
Precipitation forecasting has an important place in everyday life – during the day we may have tens of small talks discussing the likelihood that it will rain this evening or weekend. Should you take an umbrella for a walk? Or should you invite your friends for a barbecue? It will certainly depend on what your weather application shows.
While for years people were guided by the precipitation forecasts issued for a particular region or city several times a day, the widespread availability of weather radars allowed us to obtain forecasts at much higher spatiotemporal resolution of minutes in time and hundreds of meters in space. Hence, radar-based precipitation nowcasting, that is, very-short-range forecasting (typically up to 1–3 h), has become an essential technique, also in various professional application contexts, e.g., early warning, sewage control, or agriculture.
There are two major components comprising a system for precipitation nowcasting: radar-based precipitation estimates, and models to extrapolate that precipitation to the imminent future. While acknowledging the fundamental importance of radar-based precipitation retrieval for precipitation nowcasts, this thesis focuses only on the model development: the establishment of open and competitive benchmark models, the investigation of the potential of deep learning, and the development of procedures for nowcast errors diagnosis and isolation that can guide model development.
The present landscape of computational models for precipitation nowcasting still struggles with the availability of open software implementations that could serve as benchmarks for measuring progress. Focusing on this gap, we have developed and extensively benchmarked a stack of models based on different optical flow algorithms for the tracking step and a set of parsimonious extrapolation procedures based on image warping and advection. We demonstrate that these models provide skillful predictions comparable with or even superior to state-of-the-art operational software. We distribute the corresponding set of models as a software library, rainymotion, which is written in the Python programming language and openly available at GitHub (https://github.com/hydrogo/rainymotion). That way, the library acts as a tool for providing fast, open, and transparent solutions that could serve as a benchmark for further model development and hypothesis testing.
One of the promising directions for model development is to challenge the potential of deep learning – a subfield of machine learning that refers to artificial neural networks with deep architectures, which may consist of many computational layers. Deep learning showed promising results in many fields of computer science, such as image and speech recognition, or natural language processing, where it started to dramatically outperform reference methods.
The high benefit of using "big data" for training is among the main reasons for that. Hence, the emerging interest in deep learning in atmospheric sciences is also caused and concerted with the increasing availability of data – both observational and model-based. The large archives of weather radar data provide a solid basis for investigation of deep learning potential in precipitation nowcasting: one year of national 5-min composites for Germany comprises around 85 billion data points.
To this aim, we present RainNet, a deep convolutional neural network for radar-based precipitation nowcasting. RainNet was trained to predict continuous precipitation intensities at a lead time of 5 min, using several years of quality-controlled weather radar composites provided by the German Weather Service (DWD). That data set covers Germany with a spatial domain of 900 km x 900 km and has a resolution of 1 km in space and 5 min in time. Independent verification experiments were carried out on 11 summer precipitation events from 2016 to 2017. In these experiments, RainNet was applied recursively in order to achieve lead times of up to 1 h. In the verification experiments, trivial Eulerian persistence and a conventional model based on optical flow served as benchmarks. The latter is available in the previously developed rainymotion library.
RainNet significantly outperformed the benchmark models at all lead times up to 60 min for the routine verification metrics mean absolute error (MAE) and critical success index (CSI) at intensity thresholds of 0.125, 1, and 5 mm/h. However, rainymotion turned out to be superior in predicting the exceedance of higher intensity thresholds (here 10 and 15 mm/h). The limited ability of RainNet to predict high rainfall intensities is an undesirable property which we attribute to a high level of spatial smoothing introduced by the model. At a lead time of 5 min, an analysis of power spectral density confirmed a significant loss of spectral power at length scales of 16 km and below.
Obviously, RainNet had learned an optimal level of smoothing to produce a nowcast at 5 min lead time. In that sense, the loss of spectral power at small scales is informative, too, as it reflects the limits of predictability as a function of spatial scale. Beyond the lead time of 5 min, however, the increasing level of smoothing is a mere artifact – an analogue to numerical diffusion – that is not a property of RainNet itself but of its recursive application. In the context of early warning, the smoothing is particularly unfavorable since pronounced features of intense precipitation tend to get lost over longer lead times. Hence, we propose several options to address this issue in prospective research on model development for precipitation nowcasting, including an adjustment of the loss function for model training, model training for longer lead times, and the prediction of threshold exceedance.
The model development together with the verification experiments for both conventional and deep learning model predictions also revealed the need to better understand the source of forecast errors. Understanding the dominant sources of error in specific situations should help in guiding further model improvement. The total error of a precipitation nowcast consists of an error in the predicted location of a precipitation feature and an error in the change of precipitation intensity over lead time. So far, verification measures did not allow to isolate the location error, making it difficult to specifically improve nowcast models with regard to location prediction.
To fill this gap, we introduced a framework to directly quantify the location error. To that end, we detect and track scale-invariant precipitation features (corners) in radar images. We then consider these observed tracks as the true reference in order to evaluate the performance (or, inversely, the error) of any model that aims to predict the future location of a precipitation feature. Hence, the location error of a forecast at any lead time ahead of the forecast time corresponds to the Euclidean distance between the observed and the predicted feature location at the corresponding lead time.
Based on this framework, we carried out a benchmarking case study using one year worth of weather radar composites of the DWD. We evaluated the performance of four extrapolation models, two of which are based on the linear extrapolation of corner motion; and the remaining two are based on the Dense Inverse Search (DIS) method: motion vectors obtained from DIS are used to predict feature locations by linear and Semi-Lagrangian extrapolation.
For all competing models, the mean location error exceeds a distance of 5 km after 60 min, and 10 km after 110 min. At least 25% of all forecasts exceed an error of 5 km after 50 min, and of 10 km after 90 min. Even for the best models in our experiment, at least 5 percent of the forecasts will have a location error of more than 10 km after 45 min. When we relate such errors to application scenarios that are typically suggested for precipitation nowcasting, e.g., early warning, it becomes obvious that location errors matter: the order of magnitude of these errors is about the same as the typical extent of a convective cell. Hence, the uncertainty of precipitation nowcasts at such length scales – just as a result of locational errors – can be substantial already at lead times of less than 1 h. Being able to quantify the location error should hence guide any model development that is targeted towards its minimization. To that aim, we also consider the high potential of using deep learning architectures specific to the assimilation of sequential (track) data.
Last but not least, the thesis demonstrates the benefits of a general movement towards open science for model development in the field of precipitation nowcasting. All the presented models and frameworks are distributed as open repositories, thus enhancing transparency and reproducibility of the methodological approach. Furthermore, they are readily available to be used for further research studies, as well as for practical applications.
DeepGeoMap
(2021)
In recent years, deep learning improved the way remote sensing data is processed. The classification of hyperspectral data is no exception. 2D or 3D convolutional neural networks have outperformed classical algorithms on hyperspectral image classification in many cases. However, geological hyperspectral image classification includes several challenges, often including spatially more complex objects than found in other disciplines of hyperspectral imaging that have more spatially similar objects (e.g., as in industrial applications, aerial urban- or farming land cover types). In geological hyperspectral image classification, classical algorithms that focus on the spectral domain still often show higher accuracy, more sensible results, or flexibility due to spatial information independence. In the framework of this thesis, inspired by classical machine learning algorithms that focus on the spectral domain like the binary feature fitting- (BFF) and the EnGeoMap algorithm, the author of this thesis proposes, develops, tests, and discusses a novel, spectrally focused, spatial information independent, deep multi-layer convolutional neural network, named 'DeepGeoMap’, for hyperspectral geological data classification. More specifically, the architecture of DeepGeoMap uses a sequential series of different 1D convolutional neural networks layers and fully connected dense layers and utilizes rectified linear unit and softmax activation, 1D max and 1D global average pooling layers, additional dropout to prevent overfitting, and a categorical cross-entropy loss function with Adam gradient descent optimization. DeepGeoMap was realized using Python 3.7 and the machine and deep learning interface TensorFlow with graphical processing unit (GPU) acceleration. This 1D spectrally focused architecture allows DeepGeoMap models to be trained with hyperspectral laboratory image data of geochemically validated samples (e.g., ground truth samples for aerial or mine face images) and then use this laboratory trained model to classify other or larger scenes, similar to classical algorithms that use a spectral library of validated samples for image classification. The classification capabilities of DeepGeoMap have been tested using two geological hyperspectral image data sets. Both are geochemically validated hyperspectral data sets one based on iron ore and the other based on copper ore samples. The copper ore laboratory data set was used to train a DeepGeoMap model for the classification and analysis of a larger mine face scene within the Republic of Cyprus, where the samples originated from. Additionally, a benchmark satellite-based dataset, the Indian Pines data set, was used for training and testing. The classification accuracy of DeepGeoMap was compared to classical algorithms and other convolutional neural networks. It was shown that DeepGeoMap could achieve higher accuracies and outperform these classical algorithms and other neural networks in the geological hyperspectral image classification test cases. The spectral focus of DeepGeoMap was found to be the most considerable advantage compared to spectral-spatial classifiers like 2D or 3D neural networks. This enables DeepGeoMap models to train data independently of different spatial entities, shapes, and/or resolutions.
Generative adversarial networks (GANs) have been broadly applied to a wide range of application domains since their proposal. In this thesis, we propose several methods that aim to tackle different existing problems in GANs. Particularly, even though GANs are generally able to generate high-quality samples, the diversity of the generated set is often sub-optimal. Moreover, the common increase of the number of models in the original GANs framework, as well as their architectural sizes, introduces additional costs. Additionally, even though challenging, the proper evaluation of a generated set is an important direction to ultimately improve the generation process in GANs. We start by introducing two diversification methods that extend the original GANs framework to multiple adversaries to stimulate sample diversity in a generated set. Then, we introduce a new post-training compression method based on Monte Carlo methods and importance sampling to quantize and prune the weights and activations of pre-trained neural networks without any additional training. The previous method may be used to reduce the memory and computational costs introduced by increasing the number of models in the original GANs framework. Moreover, we use a similar procedure to quantize and prune gradients during training, which also reduces the communication costs between different workers in a distributed training setting. We introduce several topology-based evaluation methods to assess data generation in different settings, namely image generation and language generation. Our methods retrieve both single-valued and double-valued metrics, which, given a real set, may be used to broadly assess a generated set or separately evaluate sample quality and sample diversity, respectively. Moreover, two of our metrics use locality-sensitive hashing to accurately assess the generated sets of highly compressed GANs. The analysis of the compression effects in GANs paves the way for their efficient employment in real-world applications. Given their general applicability, the methods proposed in this thesis may be extended beyond the context of GANs. Hence, they may be generally applied to enhance existing neural networks and, in particular, generative frameworks.
Neural conversation models aim to predict appropriate contributions to a (given) conversation by using neural networks trained on dialogue data. A specific strand focuses on non-goal driven dialogues, first proposed by Ritter et al. (2011): They investigated the task of transforming an utterance into an appropriate reply. Then, this strand evolved into dialogue system approaches using long dialogue histories and additional background context. Contributing meaningful and appropriate to a conversation is a complex task, and therefore research in this area has been very diverse: Serban et al. (2016), for example, looked into utilizing variable length dialogue histories, Zhang et al. (2018) added additional context to the dialogue history, Wolf et al. (2019) proposed a model based on pre-trained Self-Attention neural networks (Vasvani et al., 2017), and Dinan et al. (2021) investigated safety issues of these approaches. This trend can be seen as a transformation from trying to somehow carry on a conversation to generating appropriate replies in a controlled and reliable way.
In this thesis, we first elaborate the meaning of appropriateness in the context of neural conversation models by drawing inspiration from the Cooperative Principle (Grice, 1975). We first define what an appropriate contribution has to be by operationalizing these maxims as demands on conversation models: being fluent, informative, consistent towards given context, coherent and following a social norm. Then, we identify different targets (or intervention points) to achieve the conversational appropriateness by investigating recent research in that field.
In this thesis, we investigate the aspect of consistency towards context in greater detail, being one aspect of our interpretation of appropriateness.
During the research, we developed a new context-based dialogue dataset (KOMODIS) that combines factual and opinionated context to dialogues. The KOMODIS
dataset is publicly available and we use the data in this thesis to gather new insights in context-augmented dialogue generation.
We further introduced a new way of encoding context within Self-Attention based neural networks. For that, we elaborate the issue of space complexity from knowledge graphs,
and propose a concise encoding strategy for structured context inspired from graph neural networks (Gilmer et al., 2017) to reduce the space complexity of the additional context. We discuss limitations of context-augmentation for neural conversation models, explore the characteristics of knowledge graphs, and explain how we create and augment knowledge graphs for our experiments.
Lastly, we analyzed the potential of reinforcement and transfer learning to improve context-consistency for neural conversation models. We find that current reward functions need to be more precise to enable the potential of reinforcement learning, and that sequential transfer learning can improve the subjective quality of generated dialogues.
Identifying urban pluvial flood-prone areas is necessary but the application of two-dimensional hydrodynamic models is limited to small areas. Data-driven models have been showing their ability to map flood susceptibility but their application in urban pluvial flooding is still rare. A flood inventory (4333 flooded locations) and 11 factors which potentially indicate an increased hazard for pluvial flooding were used to implement convolutional neural network (CNN), artificial neural network (ANN), random forest (RF) and support vector machine (SVM) to: (1) Map flood susceptibility in Berlin at 30, 10, 5, and 2 m spatial resolutions. (2) Evaluate the trained models' transferability in space. (3) Estimate the most useful factors for flood susceptibility mapping. The models' performance was validated using the Kappa, and the area under the receiver operating characteristic curve (AUC). The results indicated that all models perform very well (minimum AUC = 0.87 for the testing dataset). The RF models outperformed all other models at all spatial resolutions and the RF model at 2 m spatial resolution was superior for the present flood inventory and predictor variables. The majority of the models had a moderate performance for predictions outside the training area based on Kappa evaluation (minimum AUC = 0.8). Aspect and altitude were the most influencing factors on the image-based and point-based models respectively. Data-driven models can be a reliable tool for urban pluvial flood susceptibility mapping wherever a reliable flood inventory is available.