教学课程资源库 更多>>

关键词： Aviation   Unmanned aerial vehicles   Cold   Software   Aerospace engineering   Computer science   Engineering   Robotics   Transportation  

摘要： Great Basin bristlecone pine (Pinus longaeva D.K. Bailey) is a keystone species of the subalpine forest in the Great Basin and western Colorado Plateau ecoregions in Utah, Nevada, and California. Bristlecone pine is also the world's longest-lived non-clonal organism, with individuals occasionally reaching ages up to 5,000 years old. Because of its longevity, bristlecone pine contains an important proxy record of climate data in its growth rings. Despite its ecological and scientific importance, bristlecone pine's distribution and associated environmental drivers are poorly understood. Geospatial technologies, including unmanned aircraft systems (UAS), remote sensing, geographic information systems (GIS), and spatial modeling techniques can be used to quantify and characterize biotic and abiotic factors that constrain the fundamental and realized niches of bristlecone pine and other subalpine forest species. In Chapter 1, we describe workflows and important technical and logistical considerations for collecting aerial imagery in mountainous areas using small UAS, enabling high-quality remotely sensed datasets to be assembled to study the ecology of subalpine forests. In Chapter 2, we discuss a unique outlier population of bristlecone pine found in the Stansbury Mountains, Utah. We used GIS to delineate boundaries for five small stands of bristlecone pine and examined two competing hypotheses that could explain the species' presence in the range: 1) that the current population is a relict from the Pleistocene, or 2) that long-distance dispersal mechanisms led to bristlecone pine's migration from other mountain ranges during or after the warming period of the Pleistocene/Holocene transition. Potential migration routes and barriers to migration were considered in our effort to understand the dynamics behind the presence of this unique disjunct population of bristlecone pine. Chapter 3 describes a comprehensive mapping effort for bristlecone pine across its entire distri

关键词： Computer science   Artificial intelligence  

摘要： In response to the conflict between the essential assumption in machine learning algorithm that data points are all drawn from the same distribution and the data heterogeneity in real-world tasks, we introduced a new approach called Calibrated Ensemble targeting to improve the performance of classification model with diverse data sources. Instead of using the whole dataset to train individual models in regular ensemble methods, CE trains baseline learners by their calibrated subgroups. Clinically, this can be interpreted as several experts and each expert is proficient in diagnosing patients with some specific manifestations. Our model utilized and amplified the strength of various machine learning algorithms by moving data points to the subgroup given the best prediction performance in the training process. We implemented our model in predicting SLE flares using rSFI as our index by 1541 clinic encounters from MSH. The result from CE exceeded the performance of either individual algorithms or regular ensembles across seven performance evaluation metrics. Several technologies aimed to address imbalance and parameters chosen are applied to improve the models were applied. In addition, we analyze the instances in every cluster and found out the characteristic features in each cluster by ANOVA and Tukey HSD post-hoc test.

关键词： Computer science   Computer engineering   Remote sensing   Civil engineering  

摘要： Sensor networks that collect indoor environmental (IEQ) data are frequently used to drive building systems and to inform research and building standards. The research that connects the trends in IEQ data with human factors is well established and remains an active area of research. Conventional sensor systems often require specialized building infrastructure and are cost prohibitive, making them available primarily to large-scale indoor settings. In this thesis, we describe an alternative sensor network which utilizes a mobile sensor to collect IEQ data in traditionally under-served buildings. The sensor network is composed of inexpensive components and minimally relies on specialized infrastructure. We demonstrate that the sensor network maintains a uniquely fine spatial resolution, illustrating a unique utility beyond static sensor networks in conventional settings.

关键词： Civil engineering   Climate change   Engineering   Soil sciences   Geological engineering   Computer science  

摘要： The Florida Everglades is considered as a vulnerable wetland composed primary of organic rich peat soils, experiencing saltwater intrusion. Impact of increasing salinity on the strength and deformation properties of peat is unknown. A laboratory study was undertaken to evaluate how the growing salinity level due to sea level rise may alter the compressibility behavior of the Everglades soils. Sixteen 1-dimensional oedometer tests were conducted on undisturbed Everglades peat soils in two phases. Phase I included samples from Site 1 (saltwater) and Site 3 (freshwater) without any salinity addition. Phase II consisted of soil from Site 3 (freshwater) saturated in six different levels of salinity artificially added to the samples. Compressibility properties investigated in this study include compression index (Cc), coefficient of consolidation (Cv), hydraulic conductivity (K), and the Cα/Cc ratio. In general, it was observed that the increase in salinity beyond a threshold value tends to increase the soil compressibility properties, indicating a possible reduction in soil stability with saltwater intrusion.

关键词： Computer science   Artificial intelligence  

关键词： Computer engineering   Computer science   Artificial intelligence  

摘要： In recent years, Function-as-a-Service (FaaS) platforms have gained popularity as a way to deploy serverless applications in the cloud instigating the rise of serverless computing. Current serverless FaaS platforms generally use simple, classic scheduling algorithms for distributing function invocations while ignoring FaaS characteristics such as rapid changes in resource utilization and the freeze-thaw life cycle. In this thesis, we present FaaSRank, a scheduling service for serverless FaaS platforms based on information monitored from servers and functions. FaaSRank automatically learns scheduling policies through experience using reinforcement learning (RL) and neural networks. To bridge the gap between FaaS scheduling and RL techniques, we develop a novel Score-Rank-Select architecture for FaaSRank. We implement FaaSRank in Apache OpenWhisk, an open source serverless FaaS platform. We evaluate FaaSRank against other baseline schedulers including OpenWhisk's default scheduler on two 13-node OpenWhisk clusters by adapting real-world serverless workload traces provided by Microsoft Azure. Our results show that FaaSRank minimizes the overall average function completion time by 9.25% and 10.10% over the default scheduler. This improvement is realized while reducing the average number of inflight function invocations by 59.62% and 70.43% freeing computational resources compared to the default scheduler for our experiments on two clusters, respectively.

关键词： Computer science   Artificial intelligence  

摘要： The prevalence of machine learning applications in decision-making has sparked abundant interest in the fairness of machine learning. Existing notions of fairness are mainly defined upon predictions like equalized odds. This paper characterized the unfairness in regression from a new perspective by inspecting the prediction errors. In particular, we first defined a new fairness measurement with equalized error, which measures the dependence of prediction error on sensitive attributes. We then propose a regularization approach called Fairness Regularization with Equalized Error (FREE) which can handle more dimensions of fairness. We conducted two extensive experiments on both simulated datasets and real-world datasets to evaluate our approach's effectiveness in terms of mean square error, Hirschfeld-Gebelein-Rényi (HGR) maximal correlation coefficient, and overlapping index. The results show that our approach reduces unfairness in error more effectively, compared with representative methods.

关键词： Biomedical engineering   Bioinformatics   Computer science   Medical imaging   Neurosciences   Genetics  

摘要： Multi-view learning, also known as data fusion or data integration, is an emerging research direction in machine learning that considers the diversity of different views to improve the generalization performance. With the development of technology, the rate of gathering information has reached an unprecedented level and in more and more scenarios, data are obtained from various domains or collected with different feature extractors. Under those circumstances, the analysis of any single view can only provide limited insight. Furthermore, brutal concatenation of different views to adapt to the single-view learning setting is not an ideal solution as it not only ignores the heterogeneity of each view, but also worsens over-fitting problems. This leads to the growing research interests in multi-view learning algorithms, which are proved with well theoretical underpinnings and have shown great practical success in the past few years. Precision medicine has been a heated topic in the past few years, which focuses on prevention and treatment strategies that take individual variability into account. The rapid development of this field benefits from the acquisition of large-scale biologic databases, and this is where multi-view learning algorithms come in handy. By jointly analyzing multiple types of biological data, in-depth knowledge on the pathology can be gained and used to guide personalized diagnosis, prognosis, and treatment. In this thesis, we focus on the integration of multi-omics data aiming to 1) improve the accuracy of early detection of chronic disease for prevention; 2) identify significant biomarkers from imaging and (epi)genetic data; 3) explore cross-omics biomarker-biomarker associations for a comprehensive understanding of underlying biology.

关键词： Condensed matter physics   Nanotechnology   Computer science   Artificial intelligence   Simulation   Spectrum analysis   Experiments   High temperature   Approximation   Energy   Thin films   Lithium   Vibration  

摘要： Silicon-based computing technology is reaching its natural limitation urging a new computing paradigm to emerge. Neuromorphic computing has risen as this candidate for an advanced computing technology with the promise to excelin artificial intelligence applications with dedicated circuitry and general-purpose computing. Brain-inspired devices thus require internal electronic components to be comparable in size and energy consumption. Neuristor, the first neural spiking circuitry based on memristors, struggles to operate at nano-dimensions leading to fabrication scalability issues contrasting with their silicon counterpart devices. Throughout this thesis, a set of x-ray experiments and computational tools investigate and resolve the nature of the mechanism that drives the switching in niobium and vanadium oxide memristors. X-ray absorption spectroscopy along with DFT simulations of electronic and vibrational properties reveal a directional preference of niobium oxides, driven by local geometric deformations within dimer elongation (NbO2) or delithiation (LiNbO2). Raman shows to be a means for finally characterizing the lithium profile formation in LiNbO2. Moreover, x-ray standing wave spectroscopy and DFT solve the long-standing debate about the nature of the VO2 metal-to-insulator transition showing it occurs hand-by-hand with a structural phase transition. The cooperative nature of the VO2 transition tears down the assumption that VO2-based devices, as archetypal Mott materials, avoid the self-destructive feature of structural change transitions. Instead, dimer length variations rather than the degree of Mott physics become a key ingredient for the operation of potential NbO2 and VO2 memristors. Finally, thin films of niobium dioxide, lithium niobite, and vanadium dioxide are promising candidates for memristor scaling and consequently neuristor performance given the geometric and electronic features of the MIT of these materials with important implications into the

关键词： Hydrologic sciences   Operations research   Computer science   Applied mathematics  

摘要： Monitoring freshwater ecosystems allow us to better understand their overall ecohydrological condition within large and diverse watersheds. Due to the significant costs associated with biological monitoring, hydrological modeling is widely used to calculate ecologically relevant hydrologic indices (ERHIs) for stream health characterization in locations with lacking data. However, the reliability and applicability of these models within ecohydrological frameworks are major concerns. Particularly, hydrologic modeling’s ability to predict ERHIs is limited, especially when calibrating models by optimizing a single objective function or selecting a single optimal solution. The goal of this research was to develop model calibration strategies based on multi-objective optimization and Bayesian parameter estimation to improve the predictability of ERHIs and the overall representation of the streamflow regime. The research objectives were to (1) evaluate the predictions of ERHIs using different calibration techniques based on widely used performance metrics, (2) develop performance and signature-based calibration strategies explicitly constraining or targeting ERHIs, and (3) quantify the modeling uncertainty of ERHIs using the results from multi-objective model calibration and Bayesian inference. The developed strategies were tested in an agriculture-dominated watershed in Michigan, US, using the Unified Non-dominated Sorting Algorithm III (U-NSGA-III) for multi-objective calibration and the Soil and Water Assessment Tool (SWAT) for hydrological modeling. Performance-based calibration used objective functions based on metrics calculated on streamflow time series, whereas signature-based calibration used ERHIs values for objective functions’ formulation. For uncertainty quantification purposes, a lumped error model accounting for heteroscedasticity and autocorrelation was considered and the multiple-try Differential Evolution Adaptive Metropolis (ZS) (MT-DREAM(ZS)) algorithm