教学课程资源库 更多>>

关键词： Artificial intelligence   Computer science  

摘要： The challenge of ensuring that tools for data science and machine learning enforce ethical notions like privacy and fairness is one of the most important facing modern computer scientists. While the last decade has seen a flurry of research in this area, there are still significant challenges to using existing algorithms and definitions in practice. This thesis considers the theoretical questions arising from practical considerations, with an emphasis on machine learning applications. In particular, we make crucial definitions and obtain new results towards answering the following questions: • How can we learn optimally private classifiers subject to a hard accuracy constraint? • How can we leverage heuristic optimization oracles for private learning while still maintaining rigorous privacy guarantees? • How can we extend the coarse fairness protections provided by statistical notions of fairness to richer subgroup classes? • How can we learn subject to an individual fairness notion whose metric is not provided, but is instead learned from a panel of experts? Behavioral subject experiments validate theoretical results.

关键词： Computer engineering   Electrical engineering   Computer science  

摘要： As CMOS begins to reach fundamental limits of scaling, there has been a growing need for novel computing technologies. One such device is the lateral Silicon Germanium heterojunction bipolar which has immense promise to expand the capabilities of CMOS. The lateral bipolar transistor is able to achieve the THz speeds of the vertical bipolar transistor with a much lower power and area footprint. In addition, the FET-like structure of the lateral device allows for integration into contemporary CMOS processes. The goal of this research is to develop a design environment that can enable the creation of digital circuitry for the proposed technology. This deck will be used to demonstrate the proposed device as an enhancement to CMOS. This work explores the design, characteristics, implementation, and utility of Silicon-Germanium Lateral Heterojunction Bipolar Transistors (LHBT). At the device level, the major findings were THz peak fT/fmax performance, high speed operation in saturation, lower power dissipation enabled by advanced scaling, high drive currents, and CMOS-like density. A device is presented with operating frequencies exceeding 1THz. These results were used to develop a compact model, known as a MEXTRAM, which enables exploration of much larger circuits compared to what is achievable in TCAD. The design environment created for this research has a majority of the features available in a contemporary PDK. Verification tools including LVS, DRC, and PEX have been supported and tested. In addition, advanced EDA tools, including automated placement and routing, have been enabled to support VLSI design. Additional algorithms have been created to support synthesis and automated design with common mode logic. Tools have been verified and tested through the creation of a CML standard cell library, which contains 11 logical functions. The cells demonstrate picosecond rise and fall times, low power dissipation, and greatly reduced area when benchmarked against existing CM

关键词： Biomedical engineering   Public health   Information science   Computer science  

摘要： Sudden Cardiac Death (SCD) is a medical problem that is responsible for over 300,000 deaths per year in the United States and millions worldwide. SCD is defined as death occurring from within one hour of the onset of acute symptoms, an unwitnessed death in the absence of pre-existing progressive circulatory failures or other causes of deaths, or death during attempted resuscitation. Sudden death due to cardiac reasons is a leading cause of death among Congestive Heart Failure (CHF) patients. The use of Electronic Medical Records (EMR) systems has made a wealth of medical data available for research and analysis. Supervised machine learning methods have been successfully used for medical diagnosis. Ensemble classifiers are known to achieve better prediction accuracy than its constituent base classifiers. In an effort to understand the factors contributing to SCD, data on 2,521 patients were collected for the Sudden Cardiac Death in Heart Failure Trial (SCD-HeFT). The data included 96 features that were gathered over a period of 5 years. The goal of this dissertation was to develop a model that could accurately predict SCD based on available features. The prediction model used the Cox proportional hazards model as a score and then used the ExtraTreesClassifier algorithm as a boosting mechanism to create the ensemble. We tested the system at prediction points of 180 days and 365 days. Our best results were at 180-days with accuracy of 0.9624, specificity of 0.9915, and F1 score of 0.9607.

关键词： Computer Science   Computer Engineering   Acoustics   Speech denoising   transfer learning   speech recognition   speech enhancement   teacher-student learning   mimic loss  

摘要： Automatic speech recognition has become a part and parcel of everyday life. From the early 2000s, deep neural networks connected to hidden markov models (DNN-HMMs) have single-handedly pushed the performance of clean speech recognition systems to human level. Since then, the simple feedforward architectures have evolved to more sophisticated ones, like convolutional neural networks, which can correlate complex patterns to phones, and recurrent neural networks, which use horizontal connections to better utilize past (and future) context. These modern neural networks have pushed the boundaries of automatic speech recognition, lowering word error rates drastically. However, the improvement in performance comes at the cost of higher train times and decoding speeds because these neural networks are bulky, i.e. they have a large number of trainable parameters compared to feedforward neural networks. They are also intensely data-driven, i.e. high performance accuracy can only be achieved with a large set of training examples. The straightforwardness and simplicity of the feedforward architecture makes it a powerful contender for real-time speech recognition. There has been a growing amount of research which transfers knowledge from a cumbersome, high complexity "teacher" network to a simpler "student" network with lower complexity. The main focus is to make feedforward neural networks imitate the behavior of convolutional or recurrent neural networks. In the course of this dissertation, I am going to walk through some results on knowledge transfer from recurrent and convolutional neural nets to feedforward neural nets in the realm of speech enhancement and multilingual speech recogntion. Spectral mapping is a form of speech denoising which explicitly maps noisy speech to clean speech. In the first part of this dissertation, I describe a plug-and-play spectral mapping system that can be used as a front end feature denoiser for any speech recognition system. The features der

关键词： Computer science  

摘要： The dramatic increase in dataset volumes available to train machine learning models has led to great advances in machine learning, but at the cost of slowing down training. This dissertation develops and examines data reduction methods to speed up training with little or no drop in accuracy performance, and in some instances with an increase in accuracy. Since many studies have focused on feature selection, but did not adequately consider instance selection, our work focuses on developing novel instance reduction methods, and integrate them with feature reduction methods into integrated holistic reduction approaches. We first examine Simple Random Sample Selection without Replacement for instance selection, integrated with different feature selection methods, such as the Information Gain-based Feature Selection method, and compare its performance with the unintegrated instance selection and feature selection individually, applied at various data reduction rates. Our initial results proved that the integration of instance and feature selection performed much better than instance selection or feature selection alone, in terms of both training speed and even accuracy improvement, encouraging us to pursue integrated reduction further. To that end, we introduce a novel transpose-based instance selection approach, integrate it with different feature selection methods, and investigate the effect of those integrated reduction methods on classifier training time and accuracy performance, for a number of popular classifiers. This approach “tricks” feature selection methods to apply to instance selection, while still applying them for actual feature selection as well. Our results show that this new integrated method leads to greater performance improvement than the one mentioned above, speeding up training significantly without impacting classification accuracy, and in fact, for some classifiers like Naïve Bayes, the accuracy goes up considerably. As data reduction relies on t

关键词： Agriculture economics   Computer science  

摘要： Modern agriculture faces some of the most pressing problems in the 21st century: farm profitability, food security, and environmental sustainability. To address these problems, it is essential to raise productivity in a sustainable manner, an overarching goal known as sustainable intensification. In efforts to increase productivity through improved management of agricultural systems, a fundamental challenge is enormous complexity arising from both biological and social aspects of agricultural systems. However, despite the apparent need for research in modeling such socio-ecological systems without trivializing their complexity, most of the existing agricultural research only focuses on basic component sub-processes, making itself largely irrelevant for practical decision making for sustainable intensification. My contribution to the research community is twofold: exemplify practical models for management of agricultural systems and lay the foundation for some specific problems. In particular, two distinct models are constructed to support decision making at different levels of agricultural systems. Both models are predominantly characterized by their computational approaches, which capitalize on the ever increasing data and computational capacity. At an individual level, adaptive experimental designs based on Bayesian optimization techniques help individual farmers to efficiently learn complex management practices through on-farm experiments. In contrast, agent-based models help policy makers to gain insights into complex socio-ecological systems and design effective mitigation policies for insect resistance management at an aggregate landscape level. Although these models are necessarily ad hoc solutions to the specific problems, their modeling techniques (Bayesian optimization and agent-based modeling) are very general and applicable to many other practical problems.

关键词： Computer science  

关键词： Computer science  

摘要： Modern high performance computing (HPC) architectures feature multi-core processors with deep memory hierarchies, complex out-of-order instruction pipelines, powerful single instruction multiple data (SIMD) components, and heterogeneous accelerators. In practice, due to these architectural complexities, performance portability is a serious problem since programs tuned for one architecture usually achieve sub-optimal performance on another, which translates into excessive waste of energy and entails significant performance tuning efforts. Therefore, automatic performance tuning techniques are in high demand at DOE laboratories. Existing tools are limited in different ways. On one side, traditional compiler-based auto-tuning approaches generate many functionally equivalent code variants and evaluate them on a tuning input to identify the best one. To generate a code variant, these approaches usually compile all program source files with the same compilation flags. Furthermore, after identifying the best code variants, the final performance evaluation is usually done on a small number of testing inputs. However, these experimental settings have limitations in two ways. First, different source modules may need specialized flags to achieve the best performance. Second, a program may have severe input sensitivity so that the tuned executable yields sub-optimal performance on many other inputs in practice. Another problem is posed by multi-threaded HPC program correctness issues, such as deadlocks and data races, due to ad hoc synchronizations introduced by developers for performance purposes. There is also a need for novel tools to support bug detection and semantics validation for ad hoc synchronization constructs, e.g., ad hoc barriers. The state-of-art ad hoc synchronization analysis tools can only detect simple happen-before relationships between different program points and cannot detect complex synchronization constructs, such as ad hoc barriers, neither can they en

关键词： Design   Computer science  

摘要： This thesis showcases the development of a narrative Virtual Reality (VR) experience told in the form of an immersive graphic novel. Virtual reality has the potential to be one of the more intimate forms of storytelling since it puts the user in the center of the story or scene. Similarly the graphic novel is an excellent storytelling device since it relies more on visuals rather than writing, enabling readers to absorb information faster without losing focus. Together they become a visually evocative experience for the audience by enabling them to not just look through the visuals but also jump into each panel and experience the story firsthand. Drawing inspiration from artists and writers like Alan Moore, Hokusai and Shinichiro Watanabe, along with elements from Indian mythology, I have created a graphic narrative based on a poem and present it as an immersive VR experience using the VR 3D painting and animation tool Oculus Quill.

关键词： Artificial intelligence   Computer science  

摘要： Over the last two decades, the evolution of mobile technologies has led to an unprecedented adoption of cellphones in mainstream society. Consumers have moved from having the ability to make telephone calls from anywhere at any time using the traditional cellphones to having an all-access pass to cyberspace using today’s smartphones. In the first phase, we developed an application that has a location-based feature, which stores location data of the user with preferences for services such as WIFI, Internet Data and Bluetooth. When using this feature, a user will be able to save power in the saved locations as all saved preferences will be enforced and will work automatically, thereby eliminating the need to make changes manually. The combination of schedulers which utilizes both saved time-based tasks and location-based features is also novel, one that no other application has implemented until now. These features save power on Android phones, which show significant results for power saving applications of 8%, which is 115.27 minutes in a 24-hour time period. The second phase focuses on power prediction and consumption reduction of mobile phones. This dissertation proposes two models: Model-1 and Model-2. Model-1 utilizes the daily usage of a mobile phone and generates a matching usage pattern using our developed approach based on brute force algorithm. Every user has different usage behavior, so generated patterns will be different for each of them. This pattern will be updated every week. The matched pattern for each week works as the input for Model-2. In Model-2, Linear Regression Analysis [1], Multiple Regression Analysis [2] and Long Short-Term Memory method [3] are applied to our proposed approach to predict power usage. All the applications/services will be given a lower power supply until their next predicted activation time. Our results are significant and promising as it saves 21.7%, which is 312 minutes, from LR, 26.5%, which is 381 minutes, from MLR, and