Filter Results
35659 results
  • Cyber-physical systems (CPS) involve the cyber components of computing and communication interacting with and controlling elements in the physical world. Emerging CPS are increasingly distributed and perform coordinated sensing and actuation over large geographical areas. Examples include local-scale industrial robots, city-scale traffic management, and regional/continental-scale smart grids. Hence, a hierarchy of resource constrained embedded sensing/actuation nodes, edge cloudlets and the cloud will be key to enable scalable coordination, while simultaneously hosting the intelligence behind these systems. To meet the low-latency real-time requirements of CPS, these platforms typically harness a variety of computing resources ranging from multi-core processors to hardware accelerators such as general-purpose Graphics-Processing Units (GP-GPUs). In conjunction with low latency, a shared and precise notion of time is key to enabling coordinated action in distributed CPS. Hence, in this dissertation, we introduce abstractions, system-design methodologies and frameworks that enable time-based coordination in geo-distributed cyber-physical systems. While a shared notion of time enables coordination at the distributed scope, to coordinate effectively it is also necessary to simultaneously schedule multiple application components at the scope of each node, such that all deadlines are met, while ensuring that the resource/physical constraints of the system are satisfied. Therefore, this dissertation also introduces energy-, thermaland resource-efficient analyzable real-time scheduling techniques for applications deployed on platforms utilizing both multi-core processors and hardware accelerators. Our proposed solutions are readily applicable to commodity embedded, edge and cloud platforms, and together can enable time-aware and energy-efficient CPS.
    Data Types:
    • Document
  • Technologies such as mobile apps, web browsers, social networking sites, and IoT devices provide sophisticated services to users. At the same time, they are also increasingly collecting privacy-sensitive data about them. In some domains, such as mobile apps, this trend has resulted in an increase in the breadth of privacy settings made available to users. These settings are necessary because not all users feel comfortable having their data collected by some of these technologies. On mobile phones alone, the sheer number of apps users download is staggering. The variety of sensitive data and functionality requested by these apps has led to a demand for much more specific privacy settings. The same is true in other domains as well, such as social networks, browsers, and various IoT technologies. The result of this situation is that users feel overwhelmed by all of the settings available to them, and are thus unable to take advantage of them effectively. This dissertation examines whether machine learning techniques can be utilized to help users manage an increasingly large number of privacy settings. It specifically focuses on mobile app permissions. The research presented herein aims to simplify people’s tasks in regard to managing their large number of app privacy settings. We present the methods we used for developing models of users’ privacy preferences, and describe the interactive assistant we designed based on these models to help users configure their settings using personalized recommendations. The objective of this work is to alleviate the burden placed on users while increasing alignment between a their preferences and the privacy settings on their phones. This dissertation details three different studies. Specifically, in the first study, we used a dataset of mobile app permission settings obtained from over 200K Android users, explored different machine learning models, and analyzed different combinations of features to predict users’ mobile app permission settings. The study includes the development and evaluation of profile-based models as well as individual prediction models. It also includes simulation studies, wherein we explored the viability of different interactive configuration scenarios by testing different ways of combining dialogue inputs from users with recommendations based on machine learning models. The results of these simulations suggest that by selectively prompting users to indicate how they would like to configure a relatively small percentage of their permission settings, it is possible to accurately predict many of their remaining permission settings. Another significant finding of this first study is that a relatively small number of privacy profiles derived from clusters of like-minded users can help predict many of the permission settings that users in a given cluster prefer. The second study was designed to validate these findings in a field study with actual users. We designed an enhanced version of Android’s permission manager and collected rich information on users’ actual app permission settings. While results from this study involve a much smaller number of users, they were obtained using privacy nudges designed to increase user awareness of data being collected about them and as a result also their engagement with their permission settings. Using data collected as part of this study, we were able to generate and analyze privacy profiles built for groups of like-minded users who exhibited similar privacy preferences. Results of this study confirm that a relatively small number of profiles (or clusters of users) can capture s large percentahe of users’ diverse privacy preferences and help predict many of their desired privacy settings. They also indicate that privacy nudges can be very effective in motivating users to engage with their permission settings and in deriving privacy profiles with strong predictive power. In the third study, we evaluated our profile-based preference models by developing a privacy assistant that helps users configure their app permission settings based on the developed profiles from our second study. We report on the results of a pilot study (N=72) conducted with actual Android users who used our privacy assistant on their smartphones while performing their regular daily activities. The results indicate that participants accepted 78.7% of the recommendations made by the privacy assistant and kept 94.9% of these settings on their phones over the following six days, all while receiving daily nudges designed to motivate them to further review their settings. The dissertation also discusses the privacy profiles designed for this research and identifies essential attributes that separate people associated with different profiles (or clusters). A refined version of the Personalized Privacy Assistant was released to the Google Play store and used to collect some additional data. In summary, through a series of three studies, this dissertation shows that using a small number of privacy decisions made by a given smartphone user, it is often possible to predict a large fraction of the mobile app permission settings this user would want to have. The dissertation further shows how we have been able to effectively operationalize this finding in the form of personalized privacy assistants that can help users configure mobile app permission settings on their smartphones.
    Data Types:
    • Document
  • Supervised learning algorithms take as input a training dataset and produce a model to predict unseen data. The algorithms work well when the deployment condition is similar to the training condition in which the training data were generated. A non-stationary condition occurs when the deployment condition differs from the training condition. The change often results in a drop in performance. Non-stationary conditions are frequently encountered in machine learning applications. For instance, the issue arises when we try to train a snore detector, or vocal emotion recognizer, that trains on a fixed group of subjects but is tested on a different group of subjects. Another example is using a learning-based controller to shoot water at a distant target under changing wind conditions. To compensate for a shift in condition, techniques such as importance weighted learning (IWL) and forgetting are used. However, these techniques are not adequate.IWL can only handle covariate shifts but not concept shifts. It is also not designed for online prediction and, thus, fails to address frequent shifts in condition. While forgetting can be used to address concept shifts, it is wasteful in discarding previously learned models. To address these shortfalls, this thesis proposes looking into the three stages of supervised learning: before, during, and after the learning process. With this new perspective, we broaden our choice of strategy and have devised pre-learning, in-learning, and post-learning shift compensation methods. These new methods not only improve the performance in combating non-stationary conditions but also handle more difficult concept-shift problems and situations that require a timely response. Under the proposed unified view, IWL is grouped as an in-learning method, which modifies the learning process to adapt to a condition change. In-learning methods are applicable when a limited amount of test data is available. For example, IWL uses the test data to implicitly select training data that match the test condition during learning. We also develope an alternative to IWL that uses the concept of transfer learning. It uses test data to further train the prediction model pre-trained on general training data to better adapt to the test condition. We showed the effectiveness of the method by applying it to a vocal emotion recognizer. By using test data with an amount equivalent to half of the training data, we boosted the accuracy by 10 percent. In applications that require a timely response, such as inverse kinematics modeling and vocal emotion recognition for human-robot interaction, post-learning methods that modify prediction dynamically are suitable. Based on this concept, we have developed a local learning technique that handles multiple covariate shifts for inverse kinematics prediction. It also improves the prediction accuracy in vocal emotion recognition. In one instance, the results improved from 88.8% to 93.2% when we switched from IWL to the local learning method. Local learning also allows the use of feature augmentation to convert a more difficult conceptshift problem into an easier covariate-shift problem for our application in water shooting control. When data are abundant, we can leverage pre-learning methods such as condition-specific learning, to avoid non-stationary conditions altogether. This technique helped us in developing a semi-automatic snore labeling software that produces good accuracy (0.93 F1-score) and cuts labeling time from hours to minutes. Besides looking at data, we can also use deep learning methods to learn features that are robust to change. In our ablation study, we showed that features extracted from very deep networks and recurrent networks results in more accurate and robust snore classification. Finally, with the advance of computer simulation, unlimited artificial data can be generated to better approximate and cover possible test conditions. We tested this idea in teaching a double-hull welding robot to climb down safely from a high wall through reinforcement learning and achieved a 90% success rate. In conclusion, by looking at a broader picture of supervised learning, we extend our tools of combating non-stationary conditions from in-learning methods to post-learning and pre-learning methods. We proved the usefulness of this new perspective by applying the in-learning, post-learning, and pre-learning concepts to snore detection, vocal emotion recognition, water shooting control, and controlling a double-hull welding robot climbing down from a tall wall. They produce promising results. From these applications, we also distilled a method selection guideline using the three-stage taxonomy, where the selection is based on data availability, time urgency, and type of shift.
    Data Types:
    • Document
  • How can we model user behavior on social media platforms and social networking websites? How can we use such models to characterize behavior on social media and infer about human behavior and preferences at scale? Specifically, how can we describe users that indulge in posting about risk-taking behavior on social media or mobilize against a particular entity in a firestorm event on Twitter? Online social network platforms (e.g. Facebook, Twitter, Snapchat, Yelp) provide means for users to express themselves, by posting content in the form of images and videos. These platforms allow users to not only interact with content (liking, commenting) but also to other users (social connections, chatting) and items (through ratings and reviews), thus providing rich data with huge potential for mining unexplored and useful patterns. The availability of such data opens up unique opportunities to understand and model nuances of how users interact with such socio-technical systems, while also contributing novel algorithms that can predict genuine user behavior and also detect malicious entities at such a large scale. In this dissertation, we focus on two broad topics - (a) understanding user behavior on social media platforms and (b) detecting fraudulent activities on these platforms. For the first part, we concentrate on user behavior in two different settings - (i) individual user behavior, where we analyze behavior of actions taken at individual scale for example modeling how does individual’s expertise in e-commerce systems (such as wine rating, movie rating) evolve over time? and how can that be used to recommend the next product? The second sub-part (ii) focusses on user-based phenomena, where multiple users are analyzed collectively to discover an interesting phenomena, for example what are the characteristics of communication pattern between users participating in a firestorm event. In the second setting, we tackle the problem of detecting fraudulent activities on social media platforms. We solve two related sub-themes in the problem area, in the first sub area, we characterize various fraudulent activities on social media platforms and propose anomaly detection models to identify fraudulent users and activities. For the next sub-area we propose models that are not only confined to social media platforms, but can also be extended to general settings. Overall, this thesis looks at two closely related problems i.e. modeling user behavior on social media platforms, and then using similarly generated models to detect abnormal and potentially fraudulent behavior.
    Data Types:
    • Document
  • The prevalence of obstructive chronic lung disease in the United States is close to 15% and affects 35 million Americans today. Artificial lungs are employed to support patients that are affected most severely, but need to be replaced every one to four weeks primarily due to clot formation on circuit and oxygenator surfaces. A secondary issue that leads to device replacement is biofouling of the oxygenator surface from adhesion of bacteria from infection. Device surface fouling from clot and bacteria can be reduced through use of anticoagulant drugs, antibiotic drugs, and surface coatings. However, none of these current treatments have produced long-term effectiveness without significant side effects and risks. To address this clinical need, two approaches were used. First, surface generated nitric oxide (NO) from a novel material, Cu-PDMS, was tested for its antithrombotic and antimicrobial properties in the context of hollow fiber membrane artificial lungs. Second, the formation of clot inside hollow fiber membrane lungs was studied at the macro- and micro-scale to determine design recommendations to reduce coagulation. In this thesis work, miniature artificial lungs were tested in parallel, one with 10% wt Cu-PDMS hollow fibers and one with polymethylpentene hollow fibers, the clinical standard. To study the longer-term effects of surface generated NO from Cu-PDMS hollow fibers, this study was conducted in a 72-hour veno-venous extracorporeal membrane oxygenation attachment in a sheep model. The Cu-PDMS fibers markedly reduced blood flow resistance, an indicator of clot formation, when compared to PMP fibers and produced the most effect in the 12-36-hour range. This material was then studied for its antimicrobial effect in an environment that simulated artificial lung conditions in vitro. This study leveraged known antimicrobial agents, NO and copper, to prevent bacterial adhesion in a bioreactor system that simulated a blood stream infection in an ECMO circuit. Short-term and long-term effects of these agents were observed on the growth of Gram-negative bacteria, P. aeruginosa, and Gram-positive bacteria, S. aureus. Reduced adhesion of both strains of bacteria was observed after independent 4-hour exposure of surface generated NO, gaseous NO, and copper. However, the antimicrobial effects were short-term, and the combination of NO delivery with copper did not provide an enhanced antimicrobial effect. Lastly, the effect of different fiber bundle parameters on the initiation and progression of clot formation was studied. Current commercial oxygenators vary widely and are difficult to compare. There is no consensus on how artificial lung parameters such as packing density, path length, and frontal area affect clot formation. This study used a standardized platform in which human blood was pumped through a one-directional flow circuit that included 3D printed urethane acrylate flow chambers of various parameters that simulated the flows and conditions in an adult Quadrox oxygenator with 2 L/min flow. Micro computed tomography captured clot formation at the macro-scale, and fluorescence imaging captured clot formation at the micro-scale. These high throughput, easily repeatable studies concluded that a longer path length and small frontal area with a loosely packed fiber bundle can reduce coagulation. Furthermore, these results enable validation of computational clot models for predicting clot in an artificial lung and inform future artificial lung design.
    Data Types:
    • Document
  • The security of a software system relies on the principle of least privilege,which says that each software component must have only the privilege necessary for its execution and nothing else. Current programming languages do not provide adequate control over the privilege of untrusted software modules. To fill this gap, we designed and implemented a capability-based module system that facilitates controlling what resources each software module accesses. Then, we augmented our module system with an effect system that facilitates controlling how resources are used, i.e., authority over resources. Our approach simplifies the process of ensuring that a software system maintains the principle of least privilege. We implemented our solution as part of the Wyvern programming language. In Wyvern, modules representing or using system resources,such as the file system and network,are considered to be security-critical and are designated as resource modules. References to resource modules are capability-protected, i.e., to access a resource module, the accessing module must have the appropriate capability. Using this feature, we designed our module system in such a way that it is obvious at compile time what capabilities a module have from looking at modules’ interfaces and not their code. This property significantly simplifies the task of checking what capabilities a module holds. From a theoretical viewpoint, our capability analysis uses a novel, non-transitive notion of capabilities, which allows estimating the capabilities each module holds more precisely than in previous formal systems. Further,leveraging the fact that effects are a good proxy for operations performed on a resource, we designed Wyvern’s effect system that can account for the effects a module has on each resource. Our effect system is capability-based and allows specifying and enforcing what operations a module can perform on a resource it accesses, i.e., allows controlling the module’s authority. Similarly to our modulesystem design, effect annotations that convey information about module authority are located in modules’ interfaces, thus simplifying the task of checking resource usage. We formalized both Wyvern’s module system and effect system, and proved Wyvern to be capability- and authority-safe. We also assessed the effectiveness of the module system and the effect system that we designed in terms of how they would be used in practice and how they benefit a security-minded software developer writing an application. To do that, we implemented an extensible text-editor application in Wyvern and performed a security analysis on it.
    Data Types:
    • Document
  • This thesis studies several theoretical problems in nonparametric statistics and machine learning, mostly in the areas of nonparametric density functional estimation (estimating an integral functional of the population distribution from which the data are drawn) and nonparametric density estimation (estimating the entire population distribution from which the data are drawn). A consistent theme is that, although nonparametric density estimation is traditionally thought to be intractable in highdimensions, several equally (or more) useful tasks are relatively more tractable, even with similar or weaker assumptions on the distribution. Our work on density functional estimation focuses on several types of integral functionals, such as information theoretic quantities (entropies, mutual informations, and divergences), measures of smoothness, and measures of (dis)similarity between distributions, which play important roles as subroutines elsewhere in statistics, machine learning, and signal processing. For each of these quantities, under a variety of nonparametric models, we provide some combination of (a) new estimators, (b) upper bounds on convergence rates of these new estimators, (c) new upper bounds on the convergence rates of established estimators, (d) concentration bounds or asymptotic distributions for estimators, or (e) lower bounds on the minimax risk of estimation. We briefly discuss some applications of these density functional estimators to hypothesis testing problems such as two-sample (homogeneity) or (conditional) independence testing. For density estimation, whereas the majority of prior work has focused on estimation under L2 or other Lp losses, we consider minimax convergence rates under several new losses, including the whole spectrum of Wasserstein distances and a large class of metrics called integral probability metrics (IPMs) that includes, for example,Lp, total variation, Kolmogorov-Smirnov, earth-mover, Sobolev, Besov, and some RKHS distances. These losses open several new possibilities for nonparametric density estimation in certain cases; some examples include -convergence rates with no or reduced dependence on dimension -density-free distribution estimation, for data lying in general (e.g., non-Euclidean) metric spaces, or for data whose distribution may not be absolutely continuous with respect to Lebesgue measure -convergence rates depending only on intrinsic dimension of data Our main results here are the derivation of minimax convergence rates. However, we also briefly discuss several consequences of our results. For example, we show that IPMs have close connections with generative adversarial networks (GANs), and we leverage our results to prove the first finite-sample guarantees for GANs, in an idealized model of GANs as density estimators. These results may help explain why these tools appear to perform well at problems that are intractable from traditional perspectives of nonparametric statistics. We also briefly discuss consequences for estimation of certain density functionals, Monte Carlo integration of smooth functions, and distributionally robust optimization.
    Data Types:
    • Document
  • Research efforts in discovering and gaining better understanding of various spin-based physical phenomena over the past decades have propelled the innovation and developments of new generations of memory and logic devices. With utilization of non-volatility inherent in magnetism and low-power consumption characteristics, these novel device concepts present new opportunities for future electronics and computers. In recent years, magnetization switching via spin-orbit torques (SOTs) has come out as a promising candidate for advanced memory and computing applications, as it gives the advantages of low power consumption as well as ultrafast writing speed. The underlying mechanisms by which the SOTs induce the magnetization switching, however, turns out to be quite complex. Furthermore, for perpendicularly magnetized systems, i.e., perpendicular MRAM, the SOT driven magnetization switching of the free layer often requires an external in-plane field that significantly hinders the technological viability of commercial implementations. In this research work, we aim to gain a deeper understanding of the SOTs and their roles in inducing the magnetization switching; in turn, by means of material or device engineering, we can control the SOTs to achieve the desired switching outcomes. We particularly focus our study on the perpendicularly magnetized systems because the high perpendicular magnetic anisotropy (PMA) in these systems makes them appealing for practical applications. A major part of this research work emphasizes on the elimination of the need for an external magnetic field in the SOT switching of a perpendicular magnet. One strategy to achieve the field-free perpendicular SOT switching is through creating a magnetic field that’s localized within the device. The origin of such internal field can come from the interlayer exchange coupling. Based on this idea, we demonstrate robust field-free perpendicular magnetization switching by utilizing the spin Hall effect and interlayer exchange coupling of iridium (Ir). This is the first reported clear experimental demonstration that a heavy metal layer, Ir in particular, is capable of serving as both a spin current source and an interlayer exchange coupling layer. An additional important characteristic of Ir is that its interface with either Co or FeCoB facilitates strong perpendicular magnetic anisotropy. These combined properties allow us to achieve the SOT driven magnetization switching of a perpendicular Co layer in absence of an external field. Besides the field-free switching of a single layer, we also show that the switching scheme can be well integrated with the MgO-based magnetic tunnel junction (MTJ). We show that the three-terminal MTJ device with the Ir-enabled switching exhibits reliable writing and reading operations at zero external field, moving a step closer to the practical applications of the SOT-related magnetoresistive devices. In addition to engineering the SOT materials, we provide another solution by altering the device design. The idea is based on the well-known phenomenon that a current carrying wire produces an effective magnetic field around it. Compared to the conventional three-terminal device, our device contains an additional current line orthogonal to the write path, which can generate an in-plane Oersted field during the SOT writing. Facilitated by this Oersted field, reliable SOT switching of the perpendicular MTJs is obtained without applying an external field. The switching characteristic also renders our device unique advantages in terms of preventing the half selecting issue. In the study of the switching dynamics, we find the switching process in our devices often starts with domain nucleation followed by the domain wall motion (DWM) to expand the reversed domains. This inspires us to dig deeper into the SOT driven DWM and explore the ways in manipulating the DWM so as to control the magnetization state of a perpendicular magnet. In this work, we investigate the DWM in a system with two heavy metal underlayers that have the opposite spin Hall angles. By simply varying the relative thicknesses of these two underlayers, we can manipulate the polarity of the SOTs exerting on the DWs, which further allows us to control the direction of DWM. Based on our findings, we propose a wedge DW device where the SOT driven DWM can effectively give rise to the expansion of reversed domains and thereby realize the magnetization switching. Lastly, we show the initial experimental works for developing a novel DW device known as mCell, which can be used as the computing unit in non-volatile logic circuit without the integration with CMOS. We develop a magnetic oxide (FeOx) layer that can serve as the electric-insulating magnetic layer inserted in between the write path and read path of mCell. The FeOx insertion layer not only provides sufficient magnetic coupling between the adjacent magnetic layers, but also significantly enhances the DWM in terms of the DW velocity and power efficiency.
    Data Types:
    • Document
  • The Digital Humanities Literacy Guidebook (DHLG) was first developed in the summer of 2019 as part of a 5-year A.W. Mellon Foundation grant to Carnegie Mellon University for the advancement of digital humanities and technology-enhanced learning on campus. Its goal is to offer its audience, newcomers to DH, a broad sense of the landscape of digital humanities, and a map to chart their course through it. The DHLG is inspired by a series of five week-long summer workshops offered at CMU in 2015, 2016, 2017, 2018, and 2019. Lessons learned in these workshops helped shape the site into its current structure. This deposit includes a locally-viewable version of the site in dhlg.tar.gz, including all the video assets, and this can be used to view the site on your own machine without needing an internet connection. This deposit also includes a WARC snapshot taken of the live site on 2019-11-04.
    Data Types:
    • Other
  • The increasing demand on memory from the next-generation technologies facilitated the pathfinding and development of emerging memories, among which Resistive Random-Access Memory (RRAM) is one of the most competitive options and least understood. Many attempts have been made to understand the resistive switching phenomena in metal-oxide RRAM, most of which suffered artifacts introduced from the testing or characterization processes. In my PhD work, I selected TiN/TaO2/TiN-based RRAM to understand the mechanism of electroformation, resistive switching, filament evolution, and endurance failure of such memory cells using electrical characterization and electron microscopy techniques. The findings of my thesis work indicate that the behavior of TaO2-based RRAM is a mixture of both Valence Change Model (VCM) and Thermochemical Model (TCM). During electroformation, the Ta moves laterally towards the hot spot and down the direction of the electric field, whereas the O moves up the electric field. The motion of both Ta and O results in the TaO0.4 filament core and Ta2O5 gap after forming, corresponding to the high resistance state (HRS). The localized heating during forming also induces temperature activated ionic interdiffusion of O and Ti across the interfaces. The resistive switching is induced by the electric field applied across the device leading to Ta-rich sub-filaments forming and breaking the connection between the filament core and the electrode. Repeated resistive switching provides a virtual annealing in the conductive filament, causing the material in proximity of the filament to phase separate into metallic Ta and Ta2O5. The phase separation continues if provided with longer electrical stressing and will result in SET failure due to the Ta particles isolated by the Ta2O5 matrix. The intrinsic cycle-to-cycle variability of metal-oxide RRAM causing stochastic sub-filament overgrowth can lead the device to RESET failure. Endurance can be potentially improved by reducing interdiffusion and controlling the temperature in the memory cell during programming operations.
    Data Types:
    • Document