119092 research outputs found
Sort by
Multi-modal body part segmentation of infants using deep learning
No full textBackground: Monitoring the body temperature of premature infants is vital, as it allows optimal temperature control and may provide early warning signs for severe diseases such as sepsis. Thermography may be a non-contact and wireless alternative to state-of-the-art, cable-based methods. For monitoring use in clinical practice, automatic segmentation of the different body regions is necessary due to the movement of the infant.
Methods: This work presents and evaluates algorithms for automatic segmentation of infant body parts using deep learning methods. Based on a U-Net architecture, three neural networks were developed and compared. While the first two only used one imaging modality (visible light or thermography), the third applied a feature fusion of both. For training and evaluation, a dataset containing 600 visible light and 600 thermography images from 20 recordings of infants was created and manually labeled. In addition, we used transfer learning on publicly available datasets of adults in combination with data augmentation to improve the segmentation results.
Results: Individual optimization of the three deep learning models revealed that transfer learning and data augmentation improved segmentation regardless of the imaging modality. The fusion model achieved the best results during the final evaluation with a mean Intersection-over-Union (mIoU) of 0.85, closely followed by the RGB model. Only the thermography model achieved a lower accuracy (mIoU of 0.75). The results of the individual classes showed that all body parts were well-segmented, only the accuracy on the torso is inferior since the models struggle when only small areas of the skin are visible.
Conclusion: The presented multi-modal neural networks represent a new approach to the problem of infant body segmentation with limited available data. Robust results were obtained by applying feature fusion, cross-modality transfer learning and classical augmentation strategies
A Coarse-grained Reconfigurable Array for High-Performance Computing
No full textHigh-performance computing (HPC) comprises applications which run compute-intensive tasks on vast and at times scattered data and which need to be run on multi-node server systems communicating over high-speed networks to compute results as fast as possible. These systems are hardware-software-codesigned to guarantee a good trade-off between flexibility for different application mappings and high hardware utilization. Yet, the HPC applications are ultimately memory-bound, on the one hand, or communication-bound, on the other hand: The access latencies of memories historically have not decreased as fast as the compute performance of processors has increased, and the latencies of growing networks are inherently bottlenecking due to the physical distance of the nodes.
Designing HPC systems and tuning application mappings towards the best performance comes at the cost of a high power consumption: A huge portion of the total power goes into cooling the central processing units (CPUs). It is crucial to reduce the power consumption of processors to implicitly perform computations more power-efficiently. Hardware accelerators like coarse-grained reconfigurable arrays (CGRAs) are promising in this respect, since they have shown to have a better energy efficiency than traditional processors in the past. CGRAs feature multiple processing elements (PEs) which can be reconfigured at runtime which opens up more freedom for application mappings and opportunities for efficient execution of tasks.
In this work, the High-Performance Reconfigurable Processor (HiPReP) is proposed, a CGRA designed for HPC. Further, a methodology is presented on how to parametrize this CGRA such that it can harness the available bandwidth when running an application, e.g. a dense matrix-matrix multiplication (MMM), and on how to gauge trade-offs of scaling the number of PEs versus scaling the number of HiPReP instances in a system. Each HiPReP instance accesses the memory through a hardware component which we call frontend (FE) and which is connected to a regular level one cache (L1). Since loading and storing data as fast as possible is paramount to reduce idle times, the FE plays an important role for the achieved performance. We show that for the chosen compute system CGRAs with 5x5 PEs are sufficient to utilize the L1's bandwidth when running MMMs, and that beyond that size scaling up the number of HiPRePs scales better. For other applications which will be mapped in the future, 5x5 CGRAs are a minimal setup to explore the design space
Development of a compact hydrogen engine for commercial applications
No full textAs the individual and commercial vehicle industries seek sustainable alternatives to conventional internal combustion engines (ICEs), hydrogen-fueled rotary engines are emerging as a promising solution for several applications. This paper presents an innovative approach for the development of a hydrogen rotary engine that is integrated within a hybrid system. By exploiting the unique characteristics of rotary engines, such as compact size and high power-to-weight ratio, the electric machine, the battery and the rotary engine can be accommodated in the installation space of a conventional ICE with comparable power, despite the reduced power density of hydrogen as a fuel in ICEs.
As a first step, the hydrogen engine is naturally aspirated and equipped with direct injection. To develop a suitable calibration for the engine’s application, the influence of calibration parameters such as ignition and injection are investigated. The study examines the influence of these on operating behavior, fuel efficiency and emissions. This is supported by comprehensive measurement systems including cylinder pressure indication and emission analysis, which allows deep insights into the combustion process. First results show, that the ignition timing has only a minor influence on efficiency, but the formation of nitrogen oxides (NOx) varies by a factor of more than five as a result of an adjustment. Instead, the efficiency is primarily dependent on the mixture formation, which is significantly influenced by the injection pressure and timing.
This research highlights the potential of hydrogen rotary engines as a viable route to near-zero emission mobility. Further research and testing are underway to fully realize the benefits of this powertrain configuration in the transition to a more sustainable transport ecosystem
Degradation-aware charging strategies for bidirectional electric vehicles using deep reinforcement learning
No full textThis paper demonstrates the potential application of reinforcement learning approaches to generate optimized charging strategies for bidirectional electric vehicles (EVs). Simulative modeling of charging infrastructure and electric vehicles is used in combination with a Proximal Policy Optimization algorithm to train an agent to control the charging parameters. While ensuring the sufficient charging of the EV, our approach aims to decrease power peaks within a production environment, by taking the energy demand of the complete industrial plant into account. Apart from an uncontrolled charging strategy, two agents are benchmarked against each other: a degradation-aware agent that accounts for battery degradation during training and one that does not. The results indicate that both RL agents can significantly reduce power peaks in all investigated situations, while uncontrolled charging frequently leads to higher peaks. Furthermore, by taking battery degradation into account during training, the trained agent can even decrease battery degradation compared to uncontrolled charging, while still maintaining peak shaving performance
How trait-based interactions shape the stability and structure of mutualistic networks
No full textMutualistic interactions including pollination are essential for the functioning of ecosystems, but increasingly endangered due to anthropogenic disturbances. Understanding factors that affect their stability is therefore essential. We use a population dynamics model in combination with a trait-based network architecture to investigate the stability and structure of bipartite mutualistic networks. We find that pollinator specialisation increases robustness, i.e. the proportion of surviving species, when survival is limited by pollinator competition, by low interaction strengths, or by high mortality rates. In the opposite situations, networks with more generalists are more robust. When plant-pollinator interactions depend in an asymmetric way on their trait difference, nestedness is increased compared to a model with symmetric interactions, and specialists and generalists show a different distribution in trait space. Increased nestedness is only in some situations correlated with a higher robustness, while in other situations, the correlation is negative or neutral
Jump Table Instrumentation for Performance Analysis
No full textTo better understand the runtime behavior of a program and improve its performance, measurement techniques such as instrumentation are used. Instrumentation is the process of inserting measuring probes at certain points in the program. Since the measurements generate an overhead that can lead to drastically increased runtimes and inaccurate results, it is necessary to limit the instrumentation to particularly interesting regions. However, finding a suitable selection of regions is not trivial and usually requires several attempts. The problem with many instrumentation tools is that any change in the configuration requires the program to be recompiled, which can be very time-consuming.
In this thesis, a new dynamic compiler instrumentation technique called Jump Table Instrumentation is proposed, which allows to change its configuration without recompilation. At compile time, all functions of a given program are duplicated and instrumented in different ways (or not at all). In addition, all internal function calls are redirected via a jump table that forwards to one of the available variants of the function. By changing the entries in the jump table, the instrumentation can then be adapted dynamically. Based on a number of benchmarks, it is shown that the overhead caused by using the jump table is negligible in most cases. In return, the compile time and the binary file size increase with the number of created variants
DeepGaze III: Modeling free-viewing human scanpaths with deep learning
No full textHumans typically move their eyes in “scanpaths” of fixations linked by saccades. Here we present DeepGaze III, a new model that predicts the spatial location of consecutive fixations in a free-viewing scanpath over static images. DeepGaze III is a deep learning–based model that combines image information with information about the previous fixation history to predict where a participant might fixate next. As a high-capacity and flexible model, DeepGaze III captures many relevant patterns in the human scanpath data, setting a new state of the art in the MIT300 dataset and thereby providing insight into how much information in scanpaths across observers exists in the first place. We use this insight to assess the importance of mechanisms implemented in simpler, interpretable models for fixation selection. Due to its architecture, DeepGaze III allows us to disentangle several factors that play an important role in fixation selection, such as the interplay of scene content and scanpath history. The modular nature of DeepGaze III allows us to conduct ablation studies, which show that scene content has a stronger effect on fixation selection than previous scanpath history in our main dataset. In addition, we can use the model to identify scenes for which the relative importance of these sources of information differs most. These data-driven insights would be difficult to accomplish with simpler models that do not have the computational capacity to capture such patterns, demonstrating an example of how deep learning advances can be used to contribute to scientific understanding
Semantic object-scene inconsistencies affect eye movements, but not in the way predicted by contextualized meaning maps
No full textSemantic information is important in eye movement control. An important semantic influence on gaze guidance relates to object-scene relationships: objects that are semantically inconsistent with the scene attract more fixations than consistent objects. One interpretation of this effect is that fixations are driven toward inconsistent objects because they are semantically more informative. We tested this explanation using contextualized meaning maps, a method that is based on crowd-sourced ratings to quantify the spatial distribution of context-sensitive “meaning” in images. In Experiment 1, we compared gaze data and contextualized meaning maps for images, in which objects-scene consistency was manipulated. Observers fixated more on inconsistent versus consistent objects. However, contextualized meaning maps did not assign higher meaning to image regions that contained semantic inconsistencies. In Experiment 2, a large number of raters evaluated image-regions, which were deliberately selected for their content and expected meaningfulness. The results suggest that the same scene locations were experienced as slightly less meaningful when they contained inconsistent compared to consistent objects. In summary, we demonstrated that — in the context of our rating task — semantically inconsistent objects are experienced as less meaningful than their consistent counterparts and that contextualized meaning maps do not capture prototypical influences of image meaning on gaze guidance
Electric dipole polarizability constraints on neutron skin and symmetry energy
No full textWe review the experimental knowledge on the dipole polarizability (DP) of nuclei and its relation to the neutron skin thickness and properties of the neutron-rich matter equation of state (EOS). The discussion focuses on recent experiments using relativistic Coulomb excitation in inelastic proton scattering at extreme forward angles covering a mass range from ⁴⁰Ca to ²⁰⁸Pb. Constraints on the neutron skins and the density dependence of the symmetry energy are derived from a systematic comparison to calculations based on density functional theory (DFT) and ab initio methods utilizing interactions derived from chiral effective field theory (χEFT). The results consistently favor a soft EOS around or slightly below the saturation point. An outlook is provided on possible improvements in the precision achievable in stable nuclei and studies of exotic neutron-rich unstable nuclei with upcoming experimental facilities
Decentralized Control Mechanism for Adaptive Locomotion in Centipedes: Transition Between Walking and Peristalsis
No full textCentipedes adapt to various environmental changes by using different locomotor patterns. For example, when walking on land at high speed, they enhance propulsion by coordinating leg movements with body undulation. When climbing a rod, they combine leg movements with peristaltic body motion. In water, they swim solely by body undulation. Interestingly, when transitioning between different environments, such as water and land, they exhibit a mixture of different locomotor patterns.Understanding the underlying control principles of such flexible behaviors in centipedes will contribute to developing multi-legged robots capable of adapting to unpredictable terrain. Some studies were conducted focusing on the adaptive locomotion of centipedes. However, these have been limited to understanding behaviors in open spaces. In confined spaces, where leg movement may be restricted, centipedes are likely to rely on other body parts for locomotion. Understanding locomotion in confined spaces, as well as transitions between confined and open environments, could provide insights into control mechanisms that enable flexible adaptation to complex and diverse environments. To address this issue, we observed the locomotion of centipedes as they passed through a narrow aisle. Additionally, to understand the control principles behind this behavior, we developed a mathematical model. We found that centipedes perform peristaltic movement in the narrow aisle, followed by a transition to walking after passing through the aisle into an open space. Based on this finding, we extended the control rules proposed in our previous walking model and could replicate similar behaviors. Specifically, we added local sensory feedback based on the leg's bending angle, enabling adaptive transitions between walking and peristalsis in response to environmental constraints