Institut für Informatik und Computational Science
Refine
Year of publication
- 2024 (4)
- 2023 (8)
- 2022 (21)
- 2021 (28)
- 2020 (27)
- 2019 (21)
- 2018 (36)
- 2017 (17)
- 2016 (25)
- 2015 (30)
- 2014 (45)
- 2013 (35)
- 2012 (43)
- 2011 (54)
- 2010 (43)
- 2009 (32)
- 2008 (24)
- 2007 (43)
- 2006 (60)
- 2005 (41)
- 2004 (52)
- 2003 (32)
- 2002 (21)
- 2001 (30)
- 2000 (42)
- 1999 (43)
- 1998 (36)
- 1997 (28)
- 1996 (36)
- 1995 (20)
- 1994 (19)
- 1993 (1)
Document Type
- Article (576)
- Doctoral Thesis (203)
- Monograph/Edited Volume (135)
- Other (28)
- Conference Proceeding (17)
- Part of a Book (12)
- Master's Thesis (10)
- Postprint (10)
- Preprint (4)
- Bachelor Thesis (1)
Is part of the Bibliography
- yes (998) (remove)
Keywords
- answer set programming (13)
- Answer Set Programming (10)
- Answer set programming (10)
- Machine Learning (7)
- Maschinelles Lernen (7)
- Antwortmengenprogrammierung (6)
- E-Learning (6)
- Informatik (6)
- Modellierung (5)
- Informatikdidaktik (4)
Institute
- Institut für Informatik und Computational Science (998) (remove)
Student teachers often struggle to keep track of everything that is happening in the classroom, and particularly to notice and respond when students cause disruptions. The complexity of the classroom environment is a potential contributing factor that has not been empirically tested. In this experimental study, we utilized a virtual reality (VR) classroom to examine whether classroom complexity affects the likelihood of student teachers noticing disruptions and how they react after noticing. Classroom complexity was operationalized as the number of disruptions and the existence of overlapping disruptions (multidimensionality) as well as the existence of parallel teaching tasks (simultaneity). Results showed that student teachers (n = 50) were less likely to notice the scripted disruptions, and also less likely to respond to the disruptions in a comprehensive and effortful manner when facing greater complexity. These results may have implications for both teacher training and the design of VR for training or research purpose. This study contributes to the field from two aspects: 1) it revealed how features of the classroom environment can affect student teachers' noticing of and reaction to disruptions; and 2) it extends the functionality of the VR environment-from a teacher training tool to a testbed of fundamental classroom processes that are difficult to manipulate in real-life.
Background:
Pruritus often accompanies chronic skin diseases, exerting considerable burden on many areas of patient functioning; this burden and the features of pruritus remain insufficiently characterized.
Objective:
To investigate characteristics, including localization patterns, and burden of pruritus in patients with chronic dermatoses.
Methods:
We recruited 800 patients with active chronic skin diseases. We assessed pruritus intensity, localization, and further characteristics. We used validated questionnaires to assess quality of life, work productivity and activity impairment, anxiety, depression, and sleep quality.
Results:
Nine out of every 10 patients had experienced pruritus throughout their disease and 73% in the last 7 days. Pruritus often affected the entire body and was not restricted to skin lesions. Patients with moderate to severe pruritus reported significantly more impairment to their sleep quality and work productivity, and they were more depressed and anxious than control individuals and patients with mild or no pruritus. Suicidal ideations were highly prevalent in patients with chronic pruritus (18.5%) and atopic dermatitis (11.8%).
Conclusions:
Pruritus prevalence and intensity are very high across all dermatoses studied; intensity is linked to impairment in many areas of daily functioning. Effective treatment strategies are urgently required to treat pruritus and the underlying skin disease. ( J Am Acad Dermatol 2021;84:691-700.)
Tropical cyclones range among the costliest disasters on Earth. Their economic repercussions along the supply and trade network also affect remote economies that are not directly affected. We here simulate possible global repercussions on consumption for the example case of Hurricane Sandy in the US (2012) using the shock-propagation model Acclimate. The modeled shock yields a global three-phase ripple: an initial production demand reduction and associated consumption price decrease, followed by a supply shortage with increasing prices, and finally a recovery phase. Regions with strong trade relations to the US experience strong magnitudes of the ripple. A dominating demand reduction or supply shortage leads to overall consumption gains or losses of a region, respectively. While finding these repercussions in historic data is challenging due to strong volatility of economic interactions, numerical models like ours can help to identify them by approaching the problem from an exploratory angle, isolating the effect of interest. For this, our model simulates the economic interactions of over 7000 regional economic sectors, interlinked through about 1.8 million trade relations. Under global warming, the wave-like structures of the economic response to major hurricanes like the one simulated here are likely to intensify and potentially overlap with other weather extremes.
Due to climate change the frequency and character of precipitation are changing as the hydrological cycle intensifies. With regards to snowfall, global warming has two opposing influences; increasing humidity enables intense snowfall, whereas higher temperatures decrease the likelihood of snowfall. Here we show an intensification of extreme snowfall across large areas of the Northern Hemisphere under future warming. This is robust across an ensemble of global climate models when they are bias-corrected with observational data. While mean daily snowfall decreases, both the 99th and the 99.9th percentiles of daily snowfall increase in many regions in the next decades, especially for Northern America and Asia. Additionally, the average intensity of snowfall events exceeding these percentiles as experienced historically increases in many regions. This is likely to pose a challenge to municipalities in mid to high latitudes. Overall, extreme snowfall events are likely to become an increasingly important impact of climate change in the next decades, even if they will become rarer, but not necessarily less intense, in the second half of the century.
Discriminative Models for Biometric Identification using Micro- and Macro-Movements of the Eyes
(2021)
Human visual perception is an active process. Eye movements either alternate between fixations and saccades or follow a smooth pursuit movement in case of moving targets. Besides these macroscopic gaze patterns, the eyes perform involuntary micro-movements during fixations which are commonly categorized into micro-saccades, drift and tremor. Eye movements are frequently studied in cognitive psychology, because they reflect a complex interplay of perception, attention and oculomotor control.
A common insight of psychological research is that macro-movements are highly individual. Inspired by this finding, there has been a considerable amount of prior research on oculomotoric biometric identification. However, the accuracy of known approaches is too low and the time needed for identification is too long for any practical application. This thesis explores discriminative models for the task of biometric identification.
Discriminative models optimize a quality measure of the predictions and are usually superior to generative approaches in discriminative tasks. However, using discriminative models requires to select a suitable form of data representation for sequential eye gaze data; i.e., by engineering features or constructing a sequence kernel and the performance of the classification model strongly depends on the data representation. We study two fundamentally different ways of representing eye gaze within a discriminative framework. In the first part of this thesis, we explore the integration of data and psychological background knowledge in the form of generative models to construct representations. To this end, we first develop generative statistical models of gaze behavior during reading and scene viewing that account for viewer-specific distributional properties of gaze patterns. In a second step, we develop a discriminative identification model by deriving Fisher kernel functions from these and several baseline models. We find that an SVM with Fisher kernel is able to reliably identify users based on their eye gaze during reading and scene viewing. However, since the generative models are constrained to use low-frequency macro-movements, they discard a significant amount of information contained in the raw eye tracking signal at a high cost: identification requires about one minute of input recording, which makes it inapplicable for real world biometric systems. In the second part of this thesis, we study a purely data-driven modeling approach. Here, we aim at automatically discovering the individual pattern hidden in the raw eye tracking signal. To this end, we develop a deep convolutional neural network DeepEyedentification that processes yaw and pitch gaze velocities and learns a representation end-to-end. Compared to prior work, this model increases the identification accuracy by one order of magnitude and the time to identification decreases to only seconds. The DeepEyedentificationLive model further improves upon the identification performance by processing binocular input and it also detects presentation-attacks.
We find that by learning a representation, the performance of oculomotoric identification and presentation-attack detection can be driven close to practical relevance for biometric applications. Eye tracking devices with high sampling frequency and precision are expensive and the applicability of eye movement as a biometric feature heavily depends on cost of recording devices.
In the last part of this thesis, we therefore study the requirements on data quality by evaluating the performance of the DeepEyedentificationLive network under reduced spatial and temporal resolution. We find that the method still attains a high identification accuracy at a temporal resolution of only 250 Hz and a precision of 0.03 degrees. Reducing both does not have an additive deteriorating effect.
Stripe rust (Pst) is a major disease of wheat crops leading untreated to severe yield losses. The use of fungicides is often essential to control Pst when sudden outbreaks are imminent. Sensors capable of detecting Pst in wheat crops could optimize the use of fungicides and improve disease monitoring in high-throughput field phenotyping. Now, deep learning provides new tools for image recognition and may pave the way for new camera based sensors that can identify symptoms in early stages of a disease outbreak within the field. The aim of this study was to teach an image classifier to detect Pst symptoms in winter wheat canopies based on a deep residual neural network (ResNet). For this purpose, a large annotation database was created from images taken by a standard RGB camera that was mounted on a platform at a height of 2 m. Images were acquired while the platform was moved over a randomized field experiment with Pst-inoculated and Pst-free plots of winter wheat. The image classifier was trained with 224 x 224 px patches tiled from the original, unprocessed camera images. The image classifier was tested on different stages of the disease outbreak. At patch level the image classifier reached a total accuracy of 90%. To test the image classifier on image level, the image classifier was evaluated with a sliding window using a large striding length of 224 px allowing for fast test performance. At image level, the image classifier reached a total accuracy of 77%. Even in a stage with very low disease spreading (0.5%) at the very beginning of the Pst outbreak, a detection accuracy of 57% was obtained. Still in the initial phase of the Pst outbreak with 2 to 4% of Pst disease spreading, detection accuracy with 76% could be attained. With further optimizations, the image classifier could be implemented in embedded systems and deployed on drones, vehicles or scanning systems for fast mapping of Pst outbreaks.
Machine learning for improvement of thermal conditions inside a hybrid ventilated animal building
(2021)
In buildings with hybrid ventilation, natural ventilation opening positions (windows), mechanical ventilation rates, heating, and cooling are manipulated to maintain desired thermal conditions. The indoor temperature is regulated solely by ventilation (natural and mechanical) when the external conditions are favorable to save external heating and cooling energy. The ventilation parameters are determined by a rule-based control scheme, which is not optimal. This study proposes a methodology to enable real-time optimum control of ventilation parameters. We developed offline prediction models to estimate future thermal conditions from the data collected from building in operation. The developed offline model is then used to find the optimal controllable ventilation parameters in real-time to minimize the setpoint deviation in the building. With the proposed methodology, the experimental building's setpoint deviation improved for 87% of time, on average, by 0.53 degrees C compared to the current deviations.
Optimized deep learning model as a basis for fast UAV mapping of weed species in winter wheat crops
(2021)
Weed maps should be available quickly, reliably, and with high detail to be useful for site-specific management in crop protection and to promote more sustainable agriculture by reducing pesticide use. Here, the optimization of a deep residual convolutional neural network (ResNet-18) for the classification of weed and crop plants in UAV imagery is proposed. The target was to reach sufficient performance on an embedded system by maintaining the same features of the ResNet-18 model as a basis for fast UAV mapping. This would enable online recognition and subsequent mapping of weeds during UAV flying operation. Optimization was achieved mainly by avoiding redundant computations that arise when a classification model is applied on overlapping tiles in a larger input image. The model was trained and tested with imagery obtained from a UAV flight campaign at low altitude over a winter wheat field, and classification was performed on species level with the weed species Matricaria chamomilla L., Papaver rhoeas L., Veronica hederifolia L., and Viola arvensis ssp. arvensis observed in that field. The ResNet-18 model with the optimized image-level prediction pipeline reached a performance of 2.2 frames per second with an NVIDIA Jetson AGX Xavier on the full resolution UAV image, which would amount to about 1.78 ha h(-1) area output for continuous field mapping. The overall accuracy for determining crop, soil, and weed species was 94%. There were some limitations in the detection of species unknown to the model. When shifting from 16-bit to 32-bit model precision, no improvement in classification accuracy was observed, but a strong decline in speed performance, especially when a higher number of filters was used in the ResNet-18 model. Future work should be directed towards the integration of the mapping process on UAV platforms, guiding UAVs autonomously for mapping purpose, and ensuring the transferability of the models to other crop fields.
In control theory, to solve a finite-horizon sequential decision problem (SDP) commonly means to find a list of decision rules that result in an optimal expected total reward (or cost) when taking a given number of decision steps. SDPs are routinely solved using Bellman's backward induction. Textbook authors (e.g. Bertsekas or Puterman) typically give more or less formal proofs to show that the backward induction algorithm is correct as solution method for deterministic and stochastic SDPs. Botta, Jansson and Ionescu propose a generic framework for finite horizon, monadic SDPs together with a monadic version of backward induction for solving such SDPs. In monadic SDPs, the monad captures a generic notion of uncertainty, while a generic measure function aggregates rewards. In the present paper, we define a notion of correctness for monadic SDPs and identify three conditions that allow us to prove a correctness result for monadic backward induction that is comparable to textbook correctness proofs for ordinary backward induction. The conditions that we impose are fairly general and can be cast in category-theoretical terms using the notion of Eilenberg-Moore algebra. They hold in familiar settings like those of deterministic or stochastic SDPs, but we also give examples in which they fail. Our results show that backward induction can safely be employed for a broader class of SDPs than usually treated in textbooks. However, they also rule out certain instances that were considered admissible in the context of Botta et al. 's generic framework. Our development is formalised in Idris as an extension of the Botta et al. framework and the sources are available as supplementary material.
This study investigates the use of pulse stretching (skew-sized) inverters for monitoring the variation of count rate and linear energy transfer (LET) of energetic particles. The basic particle detector is a cascade of two pulse stretching inverters, and the required sensing area is obtained by connecting up to 12 two-inverter cells in parallel and employing the required number of parallel arrays. The incident particles are detected as single-event transients (SETs), whereby the SET count rate denotes the particle count rate, while the SET pulsewidth distribution depicts the LET variations. The advantage of the proposed solution is the possibility to sense the LET variations using fully digital processing logic. SPICE simulations conducted on IHP's 130-nm CMOS technology have shown that the SET pulsewidth varies by approximately 550 ps over the LET range from 1 to 100 MeV center dot cm(2) center dot mg(-1). The proposed detector is intended for triggering the fault-tolerant mechanisms within a self-adaptive multiprocessing system employed in space. It can be implemented as a standalone detector or integrated in the same chip with the target system.
With the downscaling of CMOS technologies, the radiation-induced Single Event Transient (SET) effects in combinational logic have become a critical reliability issue for modern integrated circuits (ICs) intended for operation under harsh radiation conditions. The SET pulses generated in combinational logic may propagate through the circuit and eventually result in soft errors. It has thus become an imperative to address the SET effects in the early phases of the radiation-hard IC design. In general, the soft error mitigation solutions should accommodate both static and dynamic measures to ensure the optimal utilization of available resources. An efficient soft-error-aware design should address synergistically three main aspects: (i) characterization and modeling of soft errors, (ii) multi-level soft error mitigation, and (iii) online soft error monitoring. Although significant results have been achieved, the effectiveness of SET characterization methods, accuracy of predictive SET models, and efficiency of SET mitigation measures are still critical issues. Therefore, this work addresses the following topics: (i) Characterization and modeling of SET effects in standard combinational cells, (ii) Static mitigation of SET effects in standard combinational cells, and (iii) Online particle detection, as a support for dynamic soft error mitigation.
Since the standard digital libraries are widely used in the design of radiation-hard ICs, the characterization of SET effects in standard cells and the availability of accurate SET models for the Soft Error Rate (SER) evaluation are the main prerequisites for efficient radiation-hard design. This work introduces an approach for the SPICE-based standard cell characterization with the reduced number of simulations, improved SET models and optimized SET sensitivity database. It has been shown that the inherent similarities in the SET response of logic cells for different input levels can be utilized to reduce the number of required simulations. Based on characterization results, the fitting models for the SET sensitivity metrics (critical charge, generated SET pulse width and propagated SET pulse width) have been developed. The proposed models are based on the principle of superposition, and they express explicitly the dependence of the SET sensitivity of individual combinational cells on design, operating and irradiation parameters. In contrast to the state-of-the-art characterization methodologies which employ extensive look-up tables (LUTs) for storing the simulation results, this work proposes the use of LUTs for storing the fitting coefficients of the SET sensitivity models derived from the characterization results. In that way the amount of characterization data in the SET sensitivity database is reduced significantly.
The initial step in enhancing the robustness of combinational logic is the application of gate-level mitigation techniques. As a result, significant improvement of the overall SER can be achieved with minimum area, delay and power overheads. For the SET mitigation in standard cells, it is essential to employ the techniques that do not require modifying the cell structure. This work introduces the use of decoupling cells for improving the robustness of standard combinational cells. By insertion of two decoupling cells at the output of a target cell, the critical charge of the cell’s output node is increased and the attenuation of short SETs is enhanced. In comparison to the most common gate-level techniques (gate upsizing and gate duplication), the proposed approach provides better SET filtering. However, as there is no single gate-level mitigation technique with optimal performance, a combination of multiple techniques is required. This work introduces a comprehensive characterization of gate-level mitigation techniques aimed to quantify their impact on the SET robustness improvement, as well as introduced area, delay and power overhead per gate. By characterizing the gate-level mitigation techniques together with the standard cells, the required effort in subsequent SER analysis of a target design can be reduced. The characterization database of the hardened standard cells can be utilized as a guideline for selection of the most appropriate mitigation solution for a given design.
As a support for dynamic soft error mitigation techniques, it is important to enable the online detection of energetic particles causing the soft errors. This allows activating the power-greedy fault-tolerant configurations based on N-modular redundancy only at the high radiation levels. To enable such a functionality, it is necessary to monitor both the particle flux and the variation of particle LET, as these two parameters contribute significantly to the system SER. In this work, a particle detection approach based on custom-sized pulse stretching inverters is proposed. Employing the pulse stretching inverters connected in parallel enables to measure the particle flux in terms of the number of detected SETs, while the particle LET variations can be estimated from the distribution of SET pulse widths. This approach requires a purely digital processing logic, in contrast to the standard detectors which require complex mixed-signal processing. Besides the possibility of LET monitoring, additional advantages of the proposed particle detector are low detection latency and power consumption, and immunity to error accumulation.
The results achieved in this thesis can serve as a basis for establishment of an overall soft-error-aware database for a given digital library, and a comprehensive multi-level radiation-hard design flow that can be implemented with the standard IC design tools. The following step will be to evaluate the achieved results with the irradiation experiments.
Use of a standard non-rad-hard digital cell library in the rad-hard design can be a cost-effective solution for space applications. In this paper we demonstrate how a standard non-rad-hard flip-flop, as one of the most vulnerable digital cells, can be converted into a rad-hard flip-flop without modifying its internal structure. We present five variants of a Triple Modular Redundancy (TMR) flip-flop: baseline TMR flip-flop, latch-based TMR flip-flop, True-Single Phase Clock (TSPC) TMR flip-flop, scannable TMR flip-flop and self-correcting TMR flipflop. For all variants, the multi-bit upsets have been addressed by applying special placement constraints, while the Single Event Transient (SET) mitigation was achieved through the usage of customized SET filters and selection of optimal inverter sizes for the clock and reset trees. The proposed flip-flop variants feature differing performance, thus enabling to choose the optimal solution for every sensitive node in the circuit, according to the predefined design constraints. Several flip-flop designs have been validated on IHP's 130nm BiCMOS process, by irradiation of custom-designed shift registers. It has been shown that the proposed TMR flip-flops are robust to soft errors with a threshold Linear Energy Transfer (LET) from (32.4 MeV.cm(2)/mg) to (62.5 MeV.cm(2)/mg), depending on the variant.
In recent years, many efforts have been made to apply image processing techniques for plant leaf identification. However, categorizing leaf images at the cultivar/variety level, because of the very low inter-class variability, is still a challenging task. In this research, we propose an automatic discriminative method based on convolutional neural networks (CNNs) for classifying 12 different cultivars of common beans that belong to three various species. We show that employing advanced loss functions, such as Additive Angular Margin Loss and Large Margin Cosine Loss, instead of the standard softmax loss function for the classification can yield better discrimination between classes and thereby mitigate the problem of low inter-class variability. The method was evaluated by classifying species (level I), cultivars from the same species (level II), and cultivars from different species (level III), based on images from the leaf foreside and backside. The results indicate that the performance of the classification algorithm on the leaf backside image dataset is superior. The maximum mean classification accuracies of 95.86, 91.37 and 86.87% were obtained at the levels I, II and III, respectively. The proposed method outperforms the previous relevant works and provides a reliable approach for plant cultivars identification.
Epistemic logic programs constitute an extension of the stable model semantics to deal with new constructs called subjective literals. Informally speaking, a subjective literal allows checking whether some objective literal is true in all or some stable models. As it can be imagined, the associated semantics has proved to be non-trivial, since the truth of subjective literals may interfere with the set of stable models it is supposed to query. As a consequence, no clear agreement has been reached and different semantic proposals have been made in the literature. Unfortunately, comparison among these proposals has been limited to a study of their effect on individual examples, rather than identifying general properties to be checked. In this paper, we propose an extension of the well-known splitting property for logic programs to the epistemic case. We formally define when an arbitrary semantics satisfies the epistemic splitting property and examine some of the consequences that can be derived from that, including its relation to conformant planning and to epistemic constraints. Interestingly, we prove (through counterexamples) that most of the existing approaches fail to fulfill the epistemic splitting property, except the original semantics proposed by Gelfond 1991 and a recent proposal by the authors, called Founded Autoepistemic Equilibrium Logic.
Large-scale databases that report the inhibitory capacities of many combinations of candidate drug compounds and cultivated cancer cell lines have driven the development of preclinical drug-sensitivity models based on machine learning. However, cultivated cell lines have devolved from human cancer cells over years or even decades under selective pressure in culture conditions. Moreover, models that have been trained on in vitro data cannot account for interactions with other types of cells. Drug-response data that are based on patient-derived cell cultures, xenografts, and organoids, on the other hand, are not available in the quantities that are needed to train high-capacity machine-learning models. We found that pre-training deep neural network models of drug sensitivity on in vitro drug-sensitivity databases before fine-tuning the model parameters on patient-derived data improves the models’ accuracy and improves the biological plausibility of the features, compared to training only on patient-derived data. From our experiments, we can conclude that pre-trained models outperform models that have been trained on the target domains in the vast majority of cases.
Answer Set Programming (ASP) is a paradigm for modeling and solving problems for knowledge representation and reasoning. There are plenty of results dedicated to studying the hardness of (fragments of) ASP. So far, these studies resulted in characterizations in terms of computational complexity as well as in fine-grained insights presented in form of dichotomy-style results, lower bounds when translating to other formalisms like propositional satisfiability (SAT), and even detailed parameterized complexity landscapes. A generic parameter in parameterized complexity originating from graph theory is the socalled treewidth, which in a sense captures structural density of a program. Recently, there was an increase in the number of treewidth-based solvers related to SAT. While there are translations from (normal) ASP to SAT, no reduction that preserves treewidth or at least keeps track of the treewidth increase is known. In this paper we propose a novel reduction from normal ASP to SAT that is aware of the treewidth, and guarantees that a slight increase of treewidth is indeed sufficient. Further, we show a new result establishing that, when considering treewidth, already the fragment of normal ASP is slightly harder than SAT (under reasonable assumptions in computational complexity). This also confirms that our reduction probably cannot be significantly improved and that the slight increase of treewidth is unavoidable. Finally, we present an empirical study of our novel reduction from normal ASP to SAT, where we compare treewidth upper bounds that are obtained via known decomposition heuristics. Overall, our reduction works better with these heuristics than existing translations. (c) 2021 Elsevier B.V. All rights reserved.
The highly structured nature of the educational sector demands effective policy mechanisms close to the needs of the field. That is why evidence-based policy making, endorsed by the European Commission under Erasmus+ Key Action 3, aims to make an alignment between the domains of policy and practice. Against this background, this article addresses two issues: First, that there is a vertical gap in the translation of higher-level policies to local strategies and regulations. Second, that there is a horizontal gap between educational domains regarding the policy awareness of individual players. This was analyzed in quantitative and qualitative studies with domain experts from the fields of virtual mobility and teacher training. From our findings, we argue that the combination of both gaps puts the academic bridge from secondary to tertiary education at risk, including the associated knowledge proficiency levels. We discuss the role of digitalization in the academic bridge by asking the question: which value does the involved stakeholders expect from educational policies? As a theoretical basis, we rely on the model of value co-creation for and by stakeholders. We describe the used instruments along with the obtained results and proposed benefits. Moreover, we reflect on the methodology applied, and we finally derive recommendations for future academic bridge policies.
Accurately solving classification problems nowadays is likely to be the most relevant machine learning task. Binary classification separating two classes only is algorithmically simpler but has fewer potential applications as many real-world problems are multi-class. On the reverse, separating only a subset of classes simplifies the classification task. Even though existing multi-class machine learning algorithms are very flexible regarding the number of classes, they assume that the target set Y is fixed and cannot be restricted once the training is finished. On the other hand, existing state-of-the-art production environments are becoming increasingly interconnected with the advance of Industry 4.0 and related technologies such that additional information can simplify the respective classification problems. In light of this, the main aim of this thesis is to introduce dynamic classification that generalizes multi-class classification such that the target class set can be restricted arbitrarily to a non-empty class subset M of Y at any time between two consecutive predictions.
This task is solved by a combination of two algorithmic approaches. First, classifier calibration, which transforms predictions into posterior probability estimates that are intended to be well calibrated. The analysis provided focuses on monotonic calibration and in particular corrects wrong statements that appeared in the literature. It also reveals that bin-based evaluation metrics, which became popular in recent years, are unjustified and should not be used at all. Next, the validity of Platt scaling, which is the most relevant parametric calibration approach, is analyzed in depth. In particular, its optimality for classifier predictions distributed according to four different families of probability distributions as well its equivalence with Beta calibration up to a sigmoidal preprocessing are proven. For non-monotonic calibration, extended variants on kernel density estimation and the ensemble method EKDE are introduced. Finally, the calibration techniques are evaluated using a simulation study with complete information as well as on a selection of 46 real-world data sets.
Building on this, classifier calibration is applied as part of decomposition-based classification that aims to reduce multi-class problems to simpler (usually binary) prediction tasks. For the involved fusing step performed at prediction time, a new approach based on evidence theory is presented that uses classifier calibration to model mass functions. This allows the analysis of decomposition-based classification against a strictly formal background and to prove closed-form equations for the overall combinations. Furthermore, the same formalism leads to a consistent integration of dynamic class information, yielding a theoretically justified and computationally tractable dynamic classification model. The insights gained from this modeling are combined with pairwise coupling, which is one of the most relevant reduction-based classification approaches, such that all individual predictions are combined with a weight. This not only generalizes existing works on pairwise coupling but also enables the integration of dynamic class information.
Lastly, a thorough empirical study is performed that compares all newly introduced approaches to existing state-of-the-art techniques. For this, evaluation metrics for dynamic classification are introduced that depend on corresponding sampling strategies. Thereafter, these are applied during a three-part evaluation. First, support vector machines and random forests are applied on 26 data sets from the UCI Machine Learning Repository. Second, two state-of-the-art deep neural networks are evaluated on five benchmark data sets from a relatively recent reference work. Here, computationally feasible strategies to apply the presented algorithms in combination with large-scale models are particularly relevant because a naive application is computationally intractable. Finally, reference data from a real-world process allowing the inclusion of dynamic class information are collected and evaluated. The results show that in combination with support vector machines and random forests, pairwise coupling approaches yield the best results, while in combination with deep neural networks, differences between the different approaches are mostly small to negligible. Most importantly, all results empirically confirm that dynamic classification succeeds in improving the respective prediction accuracies. Therefore, it is crucial to pass dynamic class information in respective applications, which requires an appropriate digital infrastructure.
Large-scale databases that report the inhibitory capacities of many combinations of candidate drug compounds and cultivated cancer cell lines have driven the development of preclinical drug-sensitivity models based on machine learning. However, cultivated cell lines have devolved from human cancer cells over years or even decades under selective pressure in culture conditions. Moreover, models that have been trained on in vitro data cannot account for interactions with other types of cells. Drug-response data that are based on patient-derived cell cultures, xenografts, and organoids, on the other hand, are not available in the quantities that are needed to train high-capacity machine-learning models. We found that pre-training deep neural network models of drug sensitivity on in vitro drug-sensitivity databases before fine-tuning the model parameters on patient-derived data improves the models’ accuracy and improves the biological plausibility of the features, compared to training only on patient-derived data. From our experiments, we can conclude that pre-trained models outperform models that have been trained on the target domains in the vast majority of cases.