Search

Due to its potential improvement on network throughput, network coding hasattracted considerable research interests. Random Linear Network Coding (RLNC),as a branch of research on network coding, is proposed as a cipher to protect theconfidentiality of packets possibly on the future Internet due to its features suchas packet mixing in a decentralized approach. In this paper, we propose attacksto compromise RLNC as a cipher for confidentiality protection. The attacks arebased on the blind source separation (BSS) technique, a statistical signal processingtechnique designed to recover original signals based on mixtures of original signals.We also design the scaling step to filter out artifacts generated by BSS algorithms andthe cross-checking steps to increase confidence on packets recovered by the proposedattacks. Our extensive experiments on packets collected from the Internet and acampus network show that the attacks can successfully recover about 1% of originalpackets.

Solving open-ended problems with paper and pencil is an important part of education and constitutes a large share of course workload, especially in science and engineering programs. However, grading this kind of work can be prohibitively expensive in large classes. Because of the complexity of handwritten free-form work, there are no existing computational methods capable of interpreting it. In this dissertation, we aim to work toward complete interpretation and semantic analysis of handwritten free-form solutions to homework and exam problems of the sort assigned in undergraduate engineering and science courses. In our study, students wrote their homework and exam solutions on dot patterned paper with Livescribe smartpens. Our work comprises three innovations. First, we developed methods to locate final answers. Our methods locate answers by identifying the boxes drawn around them. Experiments demonstrated that our methods are both accurate and efficient at recognizing answer boxes. Second, we developed a novel CNN-BLSTM-CRF network for semantic labeling of students' handwritten assignments. Semantic labeling is the task of classifying pen strokes according to the type of content they represent. Our method distinguishes between cross-out strokes, equation strokes, and free body diagram strokes. The input to our network is a set of pen strokes, which are sorted in chronological order, and the output is a sequence of labels for the strokes. Labeling strokes in this manner is an important step in enabling the semantic analysis of the writing. Our labeling approach outperforms existing methods. Finally, we developed a novel GRU-CRF network to locate complete mathematical equations. The network exploits the temporal context of consecutive strokes and simultaneously finds and groups equation strokes. Once our method has located the equations, they can be interpreted using existing techniques. Together, this work provides a significant step toward the automated grading of handwritten free-form course work.

An esteemed 2016 paper set a new lower bound on the clause density required to strongly refute a k-XOR formula in polynomial time. This paper used the sum of squares algorithm in conjunction with an improved bound for the injective tensor norm to reach its result, which is limited to formulas with all clauses the same size. We consider how this technique could be expanded to formulas with mixed clause sizes. We specifically focus our efforts on what we view as the simplest combination of clause sizes: an XOR formula with clauses of sizes k and 2k with k even. While this thesis does not establish a new refutation bound for this mixed size formula, it does give a prospective structure for the proof, and shows how to expand the 2016 paper’s techniques to mixed size problems. Additionally, this thesis gives an overview of the 2016 paper for new researchers.

The emergence and evolution of ideas is one of the most important processes in human society, and has been a topic of great interest for philosophers and historians. Psychologists have also attempted to develop models of how new ideas arise from the recombination of existing ones, and have proposed to model this process as being similar to biological evolution. However, studying the evolution of ideas has been limited by the difficulty in obtaining systematic data. The recent exponential growth in electronic data promises a solution, but several impediments remain, including a systematic process for extracting ideas and methods for analyzing their dynamics over time. While the general problem of identifying ideas in texts is extremely complex, one possible approach is to look at how meaning is distributed in documents, and to study the evolution of this structure across documents and over time. The research in this dissertation develops a framework for doing this in large, longitudinal corpora of documents. This system, calledthe Framework for the Analysis of Semantic Structure Evolution in Text (FASSET),exploits the statistics of changing word usage within the corpus in combination with machine learning techniques, including topic analysis, semantic embedding, adaptive clustering, and dimensionality reduction for visualization. It represents an integrated model for extracting semantic structure within single documents, within text corpora with a single time-stamp, and across a longitudinally extensive corpus. It includes new methods for text segmentation, longitudinal topic identification, and longitudinal semantic clustering. The goal is to provide a system for exploring a simplified "systems biology" of ideas through which the evolution of ideas can be studied at various levels.The FASSET system is applied to two large longitudinal corpora: Speeches in the U.S. Congress over a period of 36 consecutive years, and papers presented at the International Joint Conference on Neural Networks (IJCNN) over 18 consecutive years. The results show what issues and ideas have animated the areas of American politics and neural networks over time, how various topics have evolved in that time, and how new topics have emerged from them.

Kirigami metamaterials are engineered materials inspired by the ancient art of paper cutting. To produce them, partial cuts are placed in a base material so that stretching the material causes it to deploy, or change size and shape. Other exotic mechanical properties can emerge as well. These materials are applicable for many engineering settings where controllable structural change is desired. Existing research has focused on how kirigami patterns based on simple square and triangle grids behave when cut into different materials or used in different settings. In this thesis, we seek to understand the space of kirigami patterns beyond those based on grids. How do we find these patterns, what new features could they offer, and what useful information might lie in a pattern’s underlying structure? First, we find a set of kirigami patterns representing the 17 wallpaper groups, which are symmetry groups characterizing the space of 2D periodic patterns. By studying the graphs representing these initial kirigami patterns’ underlying connectivity, we develop a pattern augmentation method that can 1) produce new kirigami patterns, 2) help us deploy tilings that were previously difficult or impossible to deploy, and 3) produce large and controllable size change in kirigami patterns. We also study how the symmetries of a kirigami pattern change through deployment. In the second half of our work, we develop kirigami patterns based on aperiodic quasicrystal tilings. Since these tilings never repeat, they require developing more general kirigami design principles based on underlying structure. We introduce three such design methods and develop a computer simulation for verifying their effectiveness. We analyze geometric, mechanical, and topological properties of our methods and show that each method has different properties suitable for different applications. Finally, we show that one of our deployment methods preserves aspects of the unique long-range order of quasicrystal tilings.

In this paper, we provide an overview of the key concepts and data structures that underly Fxfs, one of Google’s newest open-source filesystems. Compared to traditional filesystems such as ext4, Fxfs has an unconventional design as it relies heavily on log-structured merge (LSM) trees to store its metadata. Every filesystem needs an allocator to determine where and how to store files on a device. The existing Fxfs allocator takes a partial fit approach and returns the first free gap that it finds on the device. There is a potential opportunity to improve the performance of the filesystem by designing an allocator that reduces fragmentation. The hypothesis is that with less fragmentation, an allocator will issue fewer disk IOs, which could lead to performance benefits. To test this hypothesis, we implement four new allocators and run them all on a sqlite benchmark. The results suggest that the time spent allocating is a significant driver of performance and should be considered as one of the key factors while designing a strong allocator.

Differential privacy (DP) has a strong theoretical foundation, but the path to implementing DP functions on computers is fraught with challenges. A common approach for creating a DP version of some function is to make a two-part algorithm: the algorithm first computes the true result of the function on a dataset, and the algorithm then draws noise from some probability distribution -- with the scale of this distribution based on the behavior of the function which computes the true result -- to perturb this true result and return an answer that both preserves the privacy of individuals in the dataset and provides utility for data analysts. Previous research has investigated the effects of finite-precision arithmetic on the ``noise-generation step'' of a DP algorithm. These efforts have revealed that finite-precision analogs of noise functions can behave far differently from their continuous counterparts, and that these differences can produce exploitable privacy violations. However, there has been little investigation into the effect of finite-precision arithmetic on the ``true answer step'' of a DP algorithm. In this paper, we describe how the true answer step can also behave far differently than expected, and how these differences in behavior can result in exploitable privacy violations. We demonstrate how a malicious data analyst can take advantage of these differences in behavior and use the outputs of supposedly DP functions to perform membership-inference attacks and reconstruction attacks on datasets. (These attacks affect all of the libraries of general-purpose DP functions that we investigated, including Google's DP library, IBM's diffprivlib, and Microsoft's SmartNoise.) We also describe how the functions in these libraries can be modified to offer privacy, and we prove the privacy of these modified functions.

Data scientists work with large-scale sensitive data, which inevitably leads to privacy risks. Differential Privacy (DP) is a mathematical definition of privacy that aims to mitigate privacy risks inherent in data analysis and machine learning. OpenDP, an open-source software DP tool, allows the government, industry, and academic institutions to share sensitive data to researchers or the public while preserving privacy. An active research question in DP literature is how to bound the total privacy loss of a sequence of DP computations. Most known DP theorems and the OpenDP library require the privacy parameters of each computation to be fixed in advance. However, this prevents the design of privacy preserving machine learning algorithms that change the privacy budget on the fly. For the adaptive parameter setting, privacy filters and odometers are two objects designed to model the total privacy loss. In this paper, we extend the programming framework for OpenDP library to handle DP composition under adaptive privacy budgets through Renyi Differential Privacy (RDP). To do so, we construct Renyi filter and odometer and prove its privacy guarantees by generalizing RDP Adaptive Composition. To generalize our Odometer results, we implement a constructor that converts any Odometer to a Filter. Our results allow for the real world DP deployment of ML algorithms and interactive query interfaces that adaptively update the privacy budget.

COVID-19 pushed millions of workers from the office to Zoom. With remote work becoming the new normal, firms and employees alike continue to wonder about the benefits or drawbacks to remote work on overall productivity. This paper presents a model for a firm of information workers, like that of Microsoft or Apple, and an understanding for how the productivity of such a firm reacts to a shift to remote work. Using real world effects from Microsoft’s own company-wide shift to remote work, we model productivity of a spectrum of firms pre- and post-remote work along with various formal and informal firm structures. We find that across firm types, productivity, as measured by a time-based metric of job-completion for an innovation task and a random task, significantly decreased in a post-remote work world. However, we find that instituting a simple manager network can help ensure a firm is robust to future shifts to remote work. Likewise, encouraging employees to regularly meet with individuals of other worker types and creating large employee gatherings are similarly robust methods of insulating a firm from remote work shocks. The creation and understanding of this model of a firm and its interventions opens up new opportunities for understanding the interplay between intra-/inter- group collaboration, formal/informal firm network dynamics, and helping firms and employees alike navigate a new era of remote work.

This dissertation explores how computer science became a discipline in the United States and how this process of discipline formation and institutionalization shaped the content and culture of computer science. Like other fields of knowledge, computer science in the United States emerged out of established disciplines; however, unlike many other scientific or engineering fields, computer science developed simultaneously out of multiple other disciplines and multiple different interdisciplinary arrangements. This dissertation explores how each of these disciplines in specific university contexts contributed to the construction of the character of computer science and the definition of the identity of the computer scientist and how different interdisciplinary relationships provided opportunities for the development of distinct visions of computer science. Drawing on archival research from university and department records as well as the papers of university administrators, computer scientists, mathematicians, electrical engineers, and others, this work reveals how the processes of discipline formation shaped everything from the intellectual foundation and internal culture of computer science to the modern American university’s institutional structure and approach to interdisciplinary knowledge-making.

Reliably diagnosing and evaluating generative models remains an open problem. The most common metric employed to measure the performance of image generating models has been the Fréchet Inception Distance (FID) score. It has been shown to correlate well with human judgement and it is sensitive to mode collapse. However, recent papers have demonstrated issues with the metric, such as poor performance on certain tasks, long computation time, and failure to match human perception. We propose a novel formulation, Class-Aware Latent Distance (CALD), to alleviate these issues. The proposed score can be applied across domains, is quicker to compute, and matches human judgement where FID fails. To demonstrate the capability of this approach, we perform an empirical study on three image generation tasks and compare our metric to FID and Intrinsic Multiscale Distance (IMD). Across these experiments, we show that the proposed metric offers significant speedup over existing methods, is more reliable, and can be used to diagnose and perform early stopping during the training of Generative Adversarial Networks.

With the popularity of daily exercise, researches on different aspects of exercise attract more and more researchers, and training intensity is one of them. The training intensity from different levels can bring people benefits on physical performance. Hence maintaining a stable intensity during exercise can help people increase training efficiency. Heart rate can be considered as an indicator of exercise intensity for people’s daily physical activities. Five heart rate zone theory is commonly adopted by individuals and professional athletes during their exercises and training. These heart rate zones are based upon percentages of people’s maximal heart rate. Obtaining the heart rate status during the training and exercise can measure the training intensity. With the function of heart rate prediction, stable training intensity can be maintained. The aim of the paper is to propose an optimization training system based on dynamic heart rate prediction, which can predict people’s heart rate under three different types of exercises: walking, running and rope jumping. The system can help people optimize their exercise by advising them to adjust the speed or workload to reach their pre-determined training intensity under different activities. Four Long Short-Term Memory (LSTM) neural networks are deployed, one for human activity recognition (HAR) and three for heart rate prediction.

Random numbers are used in a variety of applications including simulation, sampling, and cryptography. Fortunately, there exist many well-established methods of random number generation. An example of a well-known pseudorandom number generator is the Lagged-Fibonacci Generator (LFG). Marsaglia showed that lagged-Fibonacci generator using addition failed some of his DIEHARD statistical tests while it passed all when longer lags were used. This paper presents a scrambler that takes bits from a pseudorandom number generator and outputs (hopefully) improved pseudorandom numbers. The scrambler is based on a modified Feistel function, a method used in the generation of cryptographic random numbers, and multiplication by a chosen multiplier. We show that this scrambler improves the quality of pseudorandom number by applying it to the additive LFG with small lags. The scrambler performs well based on its performance with the Testu01 suite of randomness tests. The scrambling of the LFG is so successful that scrambled LFGs with small lags perform as well as unscrambled LFGs with long lags. This comes at the cost of a doubling of execution time, and provides users with generators with small memory footprints that can provide parallel generators like the LFGs in the SPRNG parallel random number generation package.

Behavior selection is a problem that AI development has faced for years, whether it be in video games or crowd-simulation software. While video game studios use methods like hierarchical task network (HTN) planners to control the behaviors of their non-playable characters (NPCs), different crowd-simulation software implement various forms of crowd model, some even recently proposing using utility-based behavior models. In this paper, I propose a framework for designing complex behavior structures that can be adapted for individual NPCs, groups of NPCs, or crowds of other AI agents that more closely mimic how people would naturally behave in almost any situation. This framework is achieved through combining the ability to compute the possibility of a future event with the ability to decompose abstract tasks into a plan of action through the use of HTN planners.

Near storage computation has increasingly become a focus in improving the performance of big data systems. Technological trends have moved the bottle neck of data intensive workloads to the interconnects used to move data from storage to memory. This has given a rise to the need for moving processing power closer to where the data is stored. The solution presented in this paper aims to provide a developer friendly approach to computational storage that allows multiple computational storage capable devices to be used effectively by enabling data transfer between computational storage devices directly. In this work, we build a model system on Amazon AWS and run a merge sort workload to evaluate the benefits of allowing device to device communication. We identify the scenarios in which device to device communication is effective and propose additional optimizations and improvements to better the overall solution.

This thesis concerns the task of turning silently mouthed words into audible speech. By using sensors that measure electrical signals from muscle movement (electromyography or EMG), it is possible to capture articulatory information from the face and neck that pertains to speech. Using these signals, we aim to train a machine learning model to generate audio in the original speaker's voice that corresponds to words that were silently mouthed. We call this task voicing silent speech.Voicing silent speech has a wide array of potential real-world applications. For example, it could be used to allow phone or video conversations where other people around the person speaking can't hear anything they say, or it could be useful in some clinical applications for people who can't speak normally but still have use of most of their speech articulators.There have been several papers in the past that have looked at the problem of converting EMG signals to speech. However, these prior EMG-to-speech works have focused on the artificial task of recovering audio from EMG that was recorded during normal vocalized speech. In this work, we will instead generate speech from recordings where no actual sound was produced. Models trained only on vocalized speech perform poorly when applied to silent speech due to signal differences between the two modes. Our work is the first to train a model on EMG from silent speech, allowing us to overcome these signal differences.Training with EMG from silent speech is more challenging than with EMG from vocalized speech, because when training on vocalized EMG data we have time-aligned speech targets but when training on silent EMG data there is no simultaneous audio. Our solution is to adopt a target-transfer approach, where audio output targets are transferred from vocalized recordings to silent recordings of the same utterances. To do this cross-modal training, we need to account for the fact that the two recordings are not time-aligned, so a core component of our work concerns finding the best way to align the vocalized speech targets with the silent utterances.To enable development on this task, we collect and release a dataset of nearly twenty hours of EMG speech recordings, nearly ten times larger than previous publicly available datasets. We then demonstrate a method for training a speech synthesis model on silent EMG and propose a range of other modeling improvements to make the synthesized outputs more intelligible. We validate our methods with both human and automatic metrics, demonstrating major improvements in intelligibility of generated outputs.

With the increase in available data, specifically time series data, the importance of different data analysis techniques has increased. One technique used by many data scientists is finding characteristics within subsequences of the data set. Characteristics, or features can be quantified by a process known as feature extraction . This feature extraction step is often computationally expensive, and usually requires the availability of the entire data set. During the data collection stage, data scientists may want to see patterns or characteristics. This can be achieved with Real-Time Feature Extraction, by computing feature sets while data streams in from a source. FPGAs, or Field Programmable Gate Arrays, have access to a plethora of I/O that allows for data to be streamed directly into the computational units making for efficient,real-time feature extraction. In this paper, we provide an FPGA architecture that is able to extract features in real-time, while offering latency, and power optimizations.

In this paper, I look at applications to the INNO Fund ACORN, a biannualprogram that gives fifty thousand dollar grants to start-up companies that show promise. I attempt to answer many of the common questions that entrepreneurs might have about applying to receive grant funding by using many of the advanced software tools that have been developed for natural language processing purposes. This study would be useful to anyone interested in raising money for a company, as well as someone looking to better understand the field of natural language processing, with applications in business. A transformer encoder-decoder model was used to efficiently and effectively summarize applications. This model was necessary to summarize applications through the synthesis of new sentences with novel words and phrases, as opposed to simply extracting sentences containing the most common words in the document. It was found that companies tend to receive funding more often if they have larger teams, that a founding team's tenacity is an important factor in attempting to raise money multiple times if the original attempt fails, and that many founders are serial v entrepreneurs, meaning they often are involved in the creation of multiple start-up companies throughout their career. Finally, Latent Dirichlet Allocation (LDA) was used to automatically classify the topics of all of the applications in the dataset with high confidence.

Multiple disruptions in road networks have the potential for cascading effects that can cause significant degradations of network performance resulting in large increases in travel time. However, it is challenging to identify vulnerable combinations of links in road networks due to the complex interdependency of road segments and a prohibitively large number of possible combinations of links. In this paper, we present a deep reinforcement learning (DRL) framework to identify vulnerable combinations of links in road networks. We let a DRL agent select links to disrupt, and its policy is parameterized with deep neural networks (DNNs). The policy is directly learned from the consequences of disruptions of links (i.e., congestion incurred by the disruptions) in traffic simulations where multiple links are disrupted in a sequence. As a case study, we analyzed vulnerable combinations of links in the road network in the city of Davis in California, and compared the criticality of the disruptions of links selected by the proposed DRL-based method and heuristic-based methods that use betweenness centrality or traffic counts of links. In the results, we observed that disruptions by the DRL agent induced significantly larger congestion and increase in travel time of vehicles than heuristic-based methods. Furthermore, the links selected by the DRL agent reveal that the disrupted links are selected considering both the static properties such as the topology of the road network and the capacity of the links as well as dynamic properties imposed by the traffic demand.

Real-time systems are the canonical class of applications in signal processing. They drive the development of algorithms for approaching theoretical results within demanding practical constraints, and provide opportunities for devising clever ways to take advantage of hardware capabilities. State-of-the-art contributions are presented on three topics in this field.The first contribution is the hardware, firmware, and software co-design of a wearable hearing aids research system. The system is open-source, easy to develop for, and much more powerful than traditional hearing aids. Its audio performance matches that of standard hearing aids and it can run custom DSP algorithms in usermode with only 2.4 ms of latency. The system also includes local web-based user control and wearable electrophysiology. The second contribution describes using GPU "Tensor Core" matrix multiply hardware to accelerate the computation of discrete Fourier transforms of sizes which are prime or have large prime factors. This includes mapping these sizes to the power-of-2-size Tensor Cores and emulating higher-precision arithmetic with lower-precision floating point numbers. For large batch sizes and for certain transform sizes which are odd or an odd number times 2 or 4, this approach produced state-of-the-art Fourier transform throughput. Finally, two papers on algorithms design in real-time acoustic modeling for an audio spatialization system are presented. Two perceptually relevant types of diffraction are simulated with ray-based models of sound propagation. Existing methods have accuracy or performance limitations, especially in dynamic applications. A set of algorithms called Volumetric Diffraction and Transmission (VDaT) is introduced to approximate shadowed or near-shadowed diffraction by an occluding object. Similarly, Spatially Sampled Near-Reflective Diffraction (SSNRD) handles near-reflective diffraction involving the edges of reflecting objects. Both methods use ray tracing to spatially sample the scene, approximate ground truth results to within 1-3 dB,and have fast performance suitable for real-time applications. SSNRD also incorporates path generation algorithms, uses a small deep neural network (DNN) to compute the response of each acoustical path, and applies the GPU "RT core" real-time ray tracing hardware to spatial computing tasks beyond traditional ray tracing.

Maximum Entropy Reinforcement Learning (MaxEnt RL) algorithms such as Soft Q-Learning (SQL) trade off reward and policy entropy, which has the potential to improve training stability and robustness. Most MaxEnt RL methods, however, use a constant tradeoff coefficient (temperature), contrary to the intuition that the temperature should be high early in training to avoid overfitting to noisy value estimates and decrease later in training as we increasingly trust high value estimates to truly lead to good rewards. Moreover, our confidence in value estimates is state-dependent, increasing every time we use more evidence to update a state's value estimate. In this paper, we present a simple state-based temperature scheduling approach and instantiate it for SQL as StaTeS-SQL. We prove the convergence of this method in the tabular case, describe how to use pseudo-counts generated by a density model to schedule the state-dependent temperature in large state spaces, and propose a combination of our method with advanced techniques collectively known as Rainbow. We evaluate our approach on the Atari Learning Environment benchmark and outperform Rainbow in 18 of 20 domains.

Similar to social media platforms, people use emojis in software development-related com- munication to enrich the context and convey additional emotion. With the increasing emoji use in software development-related communication, it has become important to understand why software developers are using emojis and their impact. Gaining a deeper understanding is essential because the intention of emoji usage might be affected by the demographics and experience of developers; also, frequency and the distribution of emoji usage might change depending on the activity, stage of the development, and nature of the conversation, etc. In this paper, we present a large-scale empirical study on the intention of emoji usage conducted on 2,712 Open Source Software (OSS) projects. We build a machine learning model to auto- mate classifying the intentions behind emoji usage in 39,980 posts. We also study the impact of emoji usage on pull request acceptance and examine if emoji usages can help facilitate the speed of getting a response. To validate the outcomes, we surveyed 60 open-source software developers from 17 countries to understand developers’ perceptions of why and when emojis are used. Our results show that we can classify the intention of emoji usage with high accuracy (AUC of 0.97). In addition, the results indicate that developers use emoji for varying intentions, and emoji usage intention changes throughout a conversation. Our results also show that emoji usage has a significant association with quicker responses which leads us to posit that an emoji conveying a positive sentiment will speed up the response time than an emoji conveying a negative sentiment. However, our results does not show that emoji usage has an impact on pull request acceptance. Our study opens a new avenue in Software Engineering research related to automatically identifying the intention of the emoji use that can help improve the communication efficiency and help project maintainers monitor and ensure the quality of communication. Another thread of future research could look into what intentions of emoji usage or what kind of emojis are more likely to attract users and how that is associated with emoji usage diffusion in different levels (threads, projects, etc.)

I propose four network architectures for improving security in machine learning model training where training data must remain secure from the model trainer and the model must be secure from the data owner. The architectures are numbered 1 through 4 with the first being the least secure and 4 being the most. Architecture 4 is the final iteration implemented in practice and facilitates end-to-end security for the Provider and one or more Consumer’s.Each architecture is comprised of multiple processes running concurrently and performing distinct tasks. The processes are deployed in docker containers with strictly defined networks to allow communications between processes only where necessary to uphold security objectives. Techniques such as attestation and cryptography are applied among various other technologies including FarmStack Trusted Connectors and Docker. All of which were employed to reduce data and application security threats further. The important contributions • Plausible architectures, which provide security to the Provider of the raw data and Con- sumer of that data, for training models in the agricultural industry. • A previously untested application of FarmStack Trusted Connectors. • The development of the Internal Data Handler (IDH), which replaced disk storage as an internal cryptographic file storage server. My results prove that of the range of architectures described in this paper, the only worth deploying are architectures 2 and 4. Architecture 2 has far from the same security guaran- tees as architecture 4 but is approximately 3.7 times faster than architecture 4 in its current state. Architecture 4, while slower, addresses all objectives of the threat model. There are many objectives in the threat model not addressed by Architecture 2. Hence, Architecture 4 is recommended over Architecture 2.

Recently, bound propagation based certified robust training methods have been proposed for training neural networks with certifiable robustness guarantees. Despite that state-of-the-art (SOTA) methods including interval bound propagation (IBP) and CROWN-IBP have succeeded in providing certified robustness with efficient per-batch training complexity, there are several challenges faced by these certified robust training methods. First, they usually use a long warmup schedule with hundreds or thousands epochs to increase the perturbation radius for SOTA performance and are thus still costly. Second, the convergence of IBP training remains unknown. In this paper, we identify two important issues related to slow warmup schedule for IBP training, namely exploded bounds at initialization, and the imbalance in ReLU activation states. These two issues make certified training difficult and unstable, and thereby long warmup schedules were needed in prior works. We proposed improvements to mitigate these issues and we are able to obtain \textbf{65.03\%} verified error on CIFAR-10 ($\epsilon=\frac{8}{255}$) using very short training schedules. For the convergence problem, we show that for a randomly initialized two-layer ReLU neural network with logistic loss, with sufficiently small perturbation radius and large network width, gradient descent for IBP training can converge to zero training robust error with a linear convergence rate with a high probability, and at this convergence state the robustness certification by IBP can accurately reflect the true robustness of the network.

The purpose of this paper is to analyze how to predict a business’s operation status during the spread of COVID-19 between the years 2020 to 2022. Exploratory Data Analysis techniques including barplot, geographical visualization, and word cloud are utilized. Three models are chosen to fit and predict business operation status: Logistic Regression, Support Vector Machine, and Bidirectional Long Short Term Memory. After tuning the probability cutoff threshold to 0.0510 with the F1 score, the logistic regression using the solver Newton's Conjugate Gradient generates an accuracy score of around 0.8120. Support Vector Machine algorithm leads to an accuracy score of around 0.9763. By using GloVe for word embedding, bidirectional LSTM results in a loss of 0.1113 and an accuracy score of 0.9758 after 2 epochs. From the performance of the three models, SVM seems to be the most effective and efficient in predicting business operation status through user reviews.

The similarity of quantum states has long been studied and some good measures such as Fidelity were developed. In this paper, we aim to come up with a similarity measure for quantum operations based on a metric called normalized Schatten 2-norm. Given two quantum operations in circuit forms, we show a quantum sampling circuit that estimates their normalized Schatten 2-norm using one clean qubit and one classical random variable. We prove a $Poly(1/\epsilon)$ upper bound on the sample complexity which is independent of the size of the quantum system. We then extend this sampling method to mixed quantum operations, i.e. a linear combination of unitary quantum operations. In addition, we build a connection between the normalized Schatten 2-norm and the similarity measure of two quantum operations. Inspired by the connection, we then develop an algorithm for quantum circuit learning.

As one of the most popular programming languages, C++ is characterized by its unique header-file mechanism that provides an effective way to access the interface of a library. However, this header-based mechanism also has its weaknesses. For instance, compilers have to perform redundant work which leads to poor compiling performance, developers should write their code carefully to avoid macro collisions, declarations and implementations are separated into multiple files which increase the complexity of the project, etc. To mitigate these challenges, the C++ standards committee proposed the modules feature of C++20 in 2020, which is introduced as an improvement of the traditional header file mechanism. C++20 modules provide a better way to encapsulate codes and address most deficiencies of header files. However, since the module feature is quite new, its advantages and potential challenges are not well-understood. On the other hand, most existing C++ applications are still based on header files and the include model, and there are not enough instructions on how to modularize a header-based app into a module-based app. To bridge these gaps, the paper discusses the influence of C++20 modules and proposes H2M, a new approach for conversion of a header-based C++ app to a module-based C++ app with better compiling performance. H2M starts by determining candidate source files for modularization. Next, it bundles up similar candidate source files and identifies appropriate dependencies. Finally, H2M generates the corresponding module-based app of the given header-based app. Our empirical studies verify the effectiveness of C++ 20 modules in improving compiling performance and the feasibility of code migration. Besides, the empirical studies on several header-based C++ applications demonstrate the effectiveness of H2M.

Methods for measuring fish populations are incredibly outdated. Even with rapid improvements in battery technology, compute power and camera systems in the last few years, the most popular techniques for counting fish and measuring their lengths employ humans. In this paper, I present FishSense as an alternative. By utilizing a compact 3-D camera, powerful embedded compute system, and easily-available off-the-shelf parts for assembly, the FishSense modules are sleek, cost-effective and modular systems able to slot into a wide variety of applications. By capturing both depth and color image data, these modules are able to measure the length of individual fish as well, offering an effective non-invasive solution to length \& biomass estimation. The hardware design allows for a high degree of modularity, with options for longer battery life, better performance, and applications in longer-term passive monitoring. The FishSense handheld model has been rigorously tested in several deployments, ranging from a small testing pool with toy fish, to a 70,000 gallon aquarium tank, to an active aquaculture center, with more planned in upcoming months, and preliminary software results from these deployments demonstrate the efficacy of the FishSense modules at measuring fish populations and sizes.

Starting in the 15th century, intellectuals and theorists began developing the techniques and visual language necessary to study, understand, review, and communicate spatial concepts through drawings and models. Drawing and drafting on paper has since remained the generally preferred method of conceptual design exploration, and designers and architects have honed the necessary techniques to explore, understand, and represent three-dimensional space with this two-dimensional medium. Today, design and especially conceptual design, has evolved into a highly iterative process that usually starts with rough doodles, sketches, and other methods for expressive, quick, and inexpensive exploration before moving the work into a more precise and less forgiving environment: the computer. Digital models already have the ability to offer designers more insights through the rapid exploration of variation, even automated iteration, and powerful simulation and validation, as well as the ability to routinely and to easily generate drawings. Even still, rough physical models and drawings are widely encouraged by educators and professionals prior to entering the digital realm, and drawing on paper endures as the de facto method of choice for designers to record and develop their ideas. Our experience and research suggests that the main issues preventing designers from fully embracing digital methods as viable alternatives or even improvements on their traditional mediums, particularly in the early-phases of the design process, are the limitations of the customary input technologies and methodologies as well as the subsequent interaction and representation capabilities available with and of the inputted geometry. Pen and multi-touch technologies have made big leaps in recent years though hardware specifications and costs still remain among the biggest hindrances; most importantly, currently affordable display surface areas are not sufficiently large or of adequate resolution for design type explorations. Our prototype 3-dimensional drawing platform was developed using a large format Microsoft Surface Hub, which we believe is the closest hardware solution currently available to an ideal high-resolution, multi-touch, drafting board. Moreover, Windows 10’s intrinsic cross-platform characteristics provide a framework for our prototype to also work on more inexpensive personal computers, such as Microsoft’s Surface Book, while the larger format displays remain cost-prohibitive. This thesis recognizes that a viable digital alternative to drawing on paper must look and feel good and must support pen and multi-touch input. Of course, the full range of possibilities that pens, pencils, markers, and paper mediums offer designers are very difficult to accurately simulate and render digitally, especially when drawing is considered a personal experience that can vary greatly from designer to designer. While we acknowledge that our platform will not be able to provide the full range of possibilities available in expression through traditional drawing, we believe that the added benefit of a more integrated workflow will overcome some of the shortcomings in expression and demonstrate a clear path for future research and development. Drawing plays different roles throughout the different stages of the design process and is also used to move ideas and intentions back and forth between those stages; from understanding others’ existing designs, to recording and developing one’s own ideas, to voluntarily misreading and reimagining those ideas. Trace paper or “red-lining” are popular common tools that allow designers and architects to draw and redraw over previous sketches or detailed drawings as they iterate, reference, and rethink past design decisions. This thesis demonstrates a more integrated and therefore more powerful way of drawing that recognizes the iterative nature of the design process and the apprehensions preventing designers from embracing digital drawing. By providing a flexible framework for iterative exploration within a 3-dimensional context, we demonstrate a valuable design solution and workflow that is not currently accessible by designers; one that blurs traditional distinctions between drawing and modeling. This thesis proposes that drawing can be more powerful in 3-dimensions, while acknowledging that 3d modeling is not drawing. We propose a 3d drawing tool, built over an existing 3d modeling ecosystem, which allows designers to do drawing, but within a 3d environment. More specifically, we have developed a system that allows 2-dimensional and 3-dimensional sketching to happen within a 3d scene, one that intrinsically provides a basis to transition back and forth more fluidly and intuitively between a hand drawing and a virtual model. We feel strongly that the best solution is one that works with existing tools that designers already find comfortable and that are an essential part of their workflows and processes. We therefore built our prototype 3d drawing platform (Cuttlefish 3D) over a popular 3d modeling software ecosystem (Rhinoceros 3D) that has a powerful feature set, a low cost of entry, and a passionate and dedicated user and developer base. In this way, a working drawing can directly become a working digital model and therefore be further and fully explored and developed within powerful 3d modeling software. We further acknowledge that most continued design explorations are part of an iterative process and that any viable solution should therefore provide the designer the ability to then sketch over and within a working digital 3d model and scene.

Starting in the 15th century, intellectuals and theorists began developing the techniques and visual language necessary to study, understand, review, and communicate spatial concepts through drawings and models. Drawing and drafting on paper has since remained the generally preferred method of conceptual design exploration, and designers and architects have honed the necessary techniques to explore, understand, and represent three-dimensional space with this two-dimensional medium. Today, design and especially conceptual design, has evolved into a highly iterative process that usually starts with rough doodles, sketches, and other methods for expressive, quick, and inexpensive exploration before moving the work into a more precise and less forgiving environment: the computer. Digital models already have the ability to offer designers more insights through the rapid exploration of variation, even automated iteration, and powerful simulation and validation, as well as the ability to routinely and to easily generate drawings. Even still, rough physical models and drawings are widely encouraged by educators and professionals prior to entering the digital realm, and drawing on paper endures as the de facto method of choice for designers to record and develop their ideas. Our experience and research suggests that the main issues preventing designers from fully embracing digital methods as viable alternatives or even improvements on their traditional mediums, particularly in the early-phases of the design process, are the limitations of the customary input technologies and methodologies as well as the subsequent interaction and representation capabilities available with and of the inputted geometry. Pen and multi-touch technologies have made big leaps in recent years though hardware specifications and costs still remain among the biggest hindrances; most importantly, currently affordable display surface areas are not sufficiently large or of adequate resolution for design type explorations. Our prototype 3-dimensional drawing platform was developed using a large format Microsoft Surface Hub, which we believe is the closest hardware solution currently available to an ideal high-resolution, multi-touch, drafting board. Moreover, Windows 10’s intrinsic cross-platform characteristics provide a framework for our prototype to also work on more inexpensive personal computers, such as Microsoft’s Surface Book, while the larger format displays remain cost-prohibitive. This thesis recognizes that a viable digital alternative to drawing on paper must look and feel good and must support pen and multi-touch input. Of course, the full range of possibilities that pens, pencils, markers, and paper mediums offer designers are very difficult to accurately simulate and render digitally, especially when drawing is considered a personal experience that can vary greatly from designer to designer. While we acknowledge that our platform will not be able to provide the full range of possibilities available in expression through traditional drawing, we believe that the added benefit of a more integrated workflow will overcome some of the shortcomings in expression and demonstrate a clear path for future research and development. Drawing plays different roles throughout the different stages of the design process and is also used to move ideas and intentions back and forth between those stages; from understanding others’ existing designs, to recording and developing one’s own ideas, to voluntarily misreading and reimagining those ideas. Trace paper or “red-lining” are popular common tools that allow designers and architects to draw and redraw over previous sketches or detailed drawings as they iterate, reference, and rethink past design decisions. This thesis demonstrates a more integrated and therefore more powerful way of drawing that recognizes the iterative nature of the design process and the apprehensions preventing designers from embracing digital drawing. By providing a flexible framework for iterative exploration within a 3-dimensional context, we demonstrate a valuable design solution and workflow that is not currently accessible by designers; one that blurs traditional distinctions between drawing and modeling. This thesis proposes that drawing can be more powerful in 3-dimensions, while acknowledging that 3d modeling is not drawing. We propose a 3d drawing tool, built over an existing 3d modeling ecosystem, which allows designers to do drawing, but within a 3d environment. More specifically, we have developed a system that allows 2-dimensional and 3-dimensional sketching to happen within a 3d scene, one that intrinsically provides a basis to transition back and forth more fluidly and intuitively between a hand drawing and a virtual model. We feel strongly that the best solution is one that works with existing tools that designers already find comfortable and that are an essential part of their workflows and processes. We therefore built our prototype 3d drawing platform (Cuttlefish 3D) over a popular 3d modeling software ecosystem (Rhinoceros 3D) that has a powerful feature set, a low cost of entry, and a passionate and dedicated user and developer base. In this way, a working drawing can directly become a working digital model and therefore be further and fully explored and developed within powerful 3d modeling software. We further acknowledge that most continued design explorations are part of an iterative process and that any viable solution should therefore provide the designer the ability to then sketch over and within a working digital 3d model and scene.

Automatic Plate Recognition plays an important roll in intelligent transportation systems. However, most license plate recognition methods work under restricted conditions like slow speed and good illumination. That is a restriction on industrial application. In this thesis, the constraints are relaxed by vanished points distortion-recovery method and denoising method. This thesis implements a license plate recognition method by morphological edge detection method and convolution neural network recognition method. The thesis is constructed contributes to several papers optimization methods. The proposed approach can be trained for recognition of country-specific license plates. More than 500 images are collected for training and over 300 images are collected for recognition test. This paper achieves 97.05% on license plate recognition for detecting total characters and numbers of the license plates.License plate recognition consists three parts, pre-processing image, locating license plate and identifying license numbers and characters. License plate location is important to obtain license images and plays a key role in identifying plates. The plate recognition has two major steps, character separation and identification. In this paper, machine learning method is applied for license plate recognition.

Briefly, this thesis paper consists of two parts, concerning spatial data management and analysis. The first part is targeted at dealing with spatial data management and visualization given some anthropogenic litter data. While the second part aims to leverage some algorithmic techniques to enhance the performance of a proposed baseline algorithm for spatial data regionalization. The combination of both works facilitates the comprehension and real practice from system design to data analysis in terms of spatial data.In the first part, anthropogenic litter data: data about waste that originates from human activities such as food waste, diapers, construction materials, used motor oil, hypodermic needles, etc, is causing growing problems for the environment and quality of life in modern cities in recent years. Such data has significant importance in the field of environmental sciences due to its important use cases that span saving marine life, reducing the risk from natural hazards, etc. In this paper, we introduce a data-driven approach that enables environmental scientists and organizations to track, manage, and model anthropogenic litter data at a large scale through smart technologies. We make a major on-going effort to collect and maintain this data worldwide from different sources through a community of environmental scientists and partner organizations. With the increasing volume of collected datasets, existing software packages, such as GIS software, do not scale to process, query, and visualize such data. To overcome this, we provide a scalable data management and visualization framework, called CleanUpOurWorld, that digests datasets from different sources, with different formats, in a scalable backend that cleans, integrates, and unifies them in a structured form. The backend includes four main modules: a data cleaner, a data integrator and loader, a data store, and a query processor. On top of this backend, frontend applications are built to visualize litter data at multiple spatial levels, from continents and oceans to street level, to enable new opportunities for both environmental scientists and organizations to track, model, and clean up litter data. The current CleanUpOurWorld implementation is based on thirty real datasets and provides different interfaces for different kinds of users.In the second part, spatial areas with multiple attributes are considered to be aggregated into larger regions. These spatial attributes in one area, when taken account of by the researchers can be divided into two categories in our case. One category is based on similarity measure, such as degree of diversity, income per area, etc. The other can be concluded as extensive threshold attributes, such as population in one area or other properties that ensure aggregation quality by giving constraints. We proposed a novel algorithm to solve the max-p-regions problem including larger initial region generation in the construction phase and less heterogeneity in the local search phase compared to previous benchmark algorithms. We conducted the experiments on multi-core platforms to ensure that parallelism is well exploited in a meta-heuristic approach. The novel algorithm provided insights in an empirical and quantitative way that can facilitate future research on this topic.

This paper proposes a conditional version of Divergence Triangle [1] as a framework to train generator, energy-based and inference models jointly with the information of labels, where the learning of the above three models are integrated perfectly in a unified probabilistic formulation. Experiments demonstrate that, within this one framework, we are able to complete the following tasks together, (1) control the fine-grained categories to generate realistic images, (2) obtain the meaningful representation of observed data in the low dimensions, and also (3) conduct label classification on unobserved data. Additionally, I also discuss a possible extension on Conditional Divergence Triangle model at the end of this paper for future work.

Many techniques for analyzing cardiovascular health rely on cardiac magnetic resonance images that have been segmented to identify various components of the heart. Manually segmenting these images is cumbersome and prone to variability, which calls for the development of accessible automation tools for cardiac researchers. In order to benefit the developing symbiosis between machine learning and medicine, such tools must be accurate, efficient and inferentially transparent. This paper introduces a U-Net-based pipeline for left ventricular segmentation of short-axis CMRs. The U-Net, a Fully Convolutional Network known for its success in biomedical image segmentation, is a natural candidate for our task. This work constitutes the core of a larger-scale project focused on improving human disease models through the acceleration of animal cardiac research. Accordingly, experiments discussed here leverage both human and animal data to explore the efficacy of image processing and model training strategies.This paper focuses on optimizing our U-Nets for resource constrained environments, and demonstrates that these models require only a fraction of their typical convolutional filters. This reduction affords efficiency with the added benefit of explainability by improving the practicality of visualizing learned features. Inference can also be further optimized by pruning trained models without any loss of accuracy or the need for retraining. Specifically, we show that U-Nets with less than 2% of their original parameters train in minutes on a single GPU and achieve Dice scores above 0.95 on multiple CMR datasets. Furthermore, inference on hundreds of images can be performed in seconds on a laptop.

This thesis considers three aspects of the numerical simulations, which arecoupling, conservation, and performance. We conduct a project and addressone challenge from each of these aspects.We propose a novel penalty force to enforce contacts with accurate Coulombfriction. The force is compatible with fully-implicit time integration and theuse of optimization-based integration. In addition to processing collisionsbetween deformable objects, the force can be used to couple rigid bodies todeformable objects or the material point method. The force naturally leads tostable stacking without drift over time, even when solvers are not run toconvergence. The force leads to an asymmetrical system, and we provide apractical solution for handling these.Next we present a new technique for transferring momentum and velocity betweenparticles and MAC grids based on the Affine-Particle-In-Cell (APIC) frameworkpreviously developed for co-locatedgrids. We extend the original APIC paper and show thatthe proposed transfers preserve linear and angular momentum and also satisfyall of the original APIC properties.Early indications in the original APIC paper suggested that APIC might besuitable for simulating high Reynolds fluids due to favorable retention ofvortices, but these properties were not studied further. We use twodimensional Fourier analysis to investigate dissipation in the limit $\dt=0$.We investigate dissipation and vortex retention numerically to quantify theeffectiveness of APIC compared with other transfer algorithms.Finally we present an efficient solver for problems typically seen inmicrofluidic applications.Microfluidic ``lab on a chip'' devices are small devices that operate on smalllength scales on small volumes of fluid. Designs for microfluidic chips aregenerally composed of standardized and often repeated components connected bylong, thin, straight fluid channels. We propose a novel discretizationalgorithm for simulating the Stokes equations on geometry with these features,which produces sparse linear systems with many repeated matrix blocks. Thediscretization is formally third order accurate for velocity and second orderaccurate for pressure in the $L^\infty$ norm. We also propose a novel linearsystem solver based on cyclic reduction, reordered sparse Gaussian elimination,and operation caching that is designed to efficiently solve systems withrepeated matrix blocks.

The segmentation of the sensory stream into discrete events is a central process in cognition, with implications for memory, learning, and attention. Event cognition can be modeled computationally as the process of inferring underlying events from an unlabeled sequence of scenes, learning the dynamics of different classes of events using the event schema abstraction, and using trained event schemata to predict future scenes and recall past events. In this paper we explore the problem of representing scenes and event schemata so that an event schema can be trained on sequences of generated scenes to learn the dynamics of a single class of events, providing a benchmark for the segmentation process. In particular, scenes are encoded with structured representations to capture their semantic content, allowing for the decodable vectorization of complex semantic relationships. Event schemata, which are transition functions that learn event dynamics, are represented as recurrent neural networks that operate on sequences of observed scene vectors to predict the next observed scene. This paper addresses problems in scene representation, event schema architecture, and training methods, arriving upon a framework for learning world dynamics that can be used as a foundation for event segmentation models.

In recent years deep learning has produced significant improvements in the field of machine learning, with some of the greatest successes resulting from the application of convolutional neural networks to Euclidean data domains such as images and audio. Many modern data sets do not have such a simple structure, however, and are accordingly more difficult to process with standard machine learning techniques. This paper provides an introduction to deep learning applied to non-Euclidean data domains. Specifically,this paper considers the generalization of essential components of convolutional neural networks to data defined over graphs. The generalization of convolutional neural networks to non-Euclidean data sets such as graphs requires a convolution operator suitable for use over graphs. Such an operator is defined by analogy with the Fourier transform through the graph Laplacian matrix. These required elements are defined and the analogy between their Euclidean counterparts is developed. The explanation of the analogy is intended to be accessible to readers with relatively limited background in the subject areas involved. In addition to the theoretical material discussed,this paper considers certain computational concerns arising in practical applications of the theories presented, and summarizes a selection of existing applications of the techniques explored. The paper concludes with a discussion of an original extension of graph convolutional neural networks developed as a part of a research project in which the author participated.

Class Imbalance is one of the numerous problems that affect the performance of a classifier. Data is said to be imbalanced when one or more classes have much fewer instances compared to other classes. In the process of learning from data, having a low number of instances of one class, results in a poor model. The Model tends to be partial to the class having a large number of instances. The issue of Imbalanced data sets is always ongoing research and there are several solutions to cope up with it in the case of two-class classification problems. However, for the multi-class imbalanced data, not every traditional technique helps to solve the problem. In the case of multi-class imbalanced data, few classes can have an extremely low number of instances and few can be moderately low. Balancing the distribution in such cases is a challenge. It is important to make sure that a model can differentiate between all available classes in the data.Apart from imbalanced data, Concept drifts also make a difference in learning a model. Concept Drifts are very likely to occur. For instance, a model that is trained on summer shopping trends can no longer be suitable to predict winter shopping trends. This paper is an analysis of how a combination of existing solutions for the above-mentioned problems behave on few synthetic and real-world data.

The human eye perceives up to 80% of all the impressions and acts as the best shield from threat. While it is believed and accepted that vision is a predominant sense in people, as per the World Health Organization, around 40 million individuals on the planet are blind, and 250 million have some type of visual disability. As a result, a lot of research and papers are being suggested to create accurate and efficient navigation models utilizing computer vision and deep learning approaches. These models should be fast and efficient, and they should be able to run on low-power mobile devices to provide real-time outdoor assistance. Our objective is to extract and categorize the information from the live stream and provide audio feedback to the user within the University campus. The classification of the objects in the stream is done by a CNN model and sent as an input for the voice feedback, which is divided into several frames using the OpenCV library and converted to audio information for the user in the real-time environment using the Google text to speech module. The results generated by the CNN model for image classification have an accuracy of over 95 percent, and real-time audio conversion is a rapid transition technique, resulting in an algorithm that performs competing with other prior state-of-art methods. We also want to integrate the application in smartphones, into our mobile app to provide a more user-friendly experience for the end-users.

This paper explores the frequency at which SQL/PHP posts on the website Stackoverflow.com contain code susceptible to SQL Injection, a common database vulnerability. Specifically, we analyze whether other users give notice of the vulnerability or provide an answer that is secure. The majority of questions analyzed were vulnerable to SQL Injection and were not corrected in their answers or brought to the attention of the original poster. To mitigate this, we present a machine learning bot which analyzes the poster’s code and alerts them of potential injection vulnerabilities, if necessary.

This paper is a study of trends in patenting of Artificial Intelligence technologies in the US using natural language techniques. Patents contain significant knowledge of the current scenario of commercial technologies. Natural language techniques can be used to extract useful information from patents which can reduce human effort to analyze them. Around 104,000 patents related to artificial intelligence technologies filed in the US were collected from the WIPO database from 2000 to 2018. Topic modeling was performed on the extracted text data. Patent applications were then categorized into five categories: Companies, Individuals, Research Institutes, Government and Universities and the trends of applicants in the field of Artificial Intelligence for each of the categories were compared.

Cumulative probability models have been introduced as a flexible alternative to linear models. However, they are computationally heavy especially when the number of distinct outcomes is large. We introduced and implemented three methods to address this problem. In the divide-and-combine method, the data are partitioned into subsets and the models are fit in parallel; in the binning and rounding methods, the outcomes are either grouped into bins or rounded (to certain significant digits or decimal places) before the models are fit. We implemented these approaches, built simulation programs, and developed an R package. We describe various challenges during the implementation and present their solutions. Examples include using ExitStack for data partition, and implementing an algorithm to merge all step functions to obtain the final estimates. We also briefly describe the results from our paper and an application to a large dataset.

miniKanren is a relational programming language embedded in Scheme. This paper studies the problem of supporting delayed goals in the context of µKanren, the core of miniKanren. It explores different representations of delayed goals and tradeoffs between them. In the presented approach, delayed goals are both accumulated and dispatched throughout the execution of a program, and the ones that remain at the end are returned to the programmer as a continuation. A novel interface, inspired by the concept of engines, allows continuously supplying µKanren code to partially computed program outputs.

This paper proposes a new toolchain for digital architecture development which is designed to leverage Fluid Pipelines, a variant of latency-insensitive (LI) systems developed at the UCSC Micro-Architecture (MASC) Lab. Prior work on Fluid Pipelines and other LI systems has shown that they have promising properties which can be leveraged in various ways to aid in digital architecture design, especially when developing and manipulating large digital architectures, but that there are also challenges when working with this type of system. This paper presents Pyrope, a toolchain designed for developing Fluid Pipelined digital architectures. The toolchain includes a custom hardware description language and optimizations for simulation, synthesis, and verification which leverage Fluid Pipelines. The final section of this document presents LiveSim, a live development environment built on top of Pyrope, which uses incremental compilation and hot binary reloading to update simulation results based on code changes in seconds, even for large architectures.

A significant aspect of communication between academic research and technology development is through various forms of publications. Many educational professionals, scholars, and students spend significant amounts of their time writing research papers, journal articles, conference papers, and other types of scholarly text. Microsoft Word is a well accepted document preparation system. To improve the usability of document authoring, this research proposes an add-on for Microsoft Word, which recommends the keywords as well as related hyperlinks.The keywords are extracted using the concept of Automatic Keyword Extraction. A domain-independent algorithm, Rapid Automatic Keyword Extraction (RAKE) is used to retrieve the keywords from the document. A custom Google search option is included in this add-on to find out the related hyperlinks for the keywords.

P4FPGA is a toolkit developed at Cornell University that allows network switch programmers to compile and test P4 programs on a variety of FPGA-based networking devices. The toolkit is open-source and the original P4FPGA paper provided a technical overview of the toolkit; this paper describes the more pragmatic aspects of the actual C++ program that constitutes the P4FPGA compiler as well presents guidance on how to use the P4FPGA compiler.

Selective pressure and molecular phenotype provide complimentary perspectives on functional properties of the human genome. In this dissertation, I develop computational methods for identifying collections of molecular phenotypes called functional classes, that are optimally informative about recent selective pressure in humans. Aggregating selective pressure across genomic positions within each functional class produces a score representing the probability that a position evincing a class-associated molecular phenotype is under selective pressure. A class’s score is interpreted as a measure of potential for fitness-influencing genomic function. Functional classes and attendant selective pressure scores are developed over the course of two papers. In the first paper, I investigate three ENCODE cell-types and develop a non-parametric representation of covariates from four genomic properties including DNase-seq, RNA-seq, chromatin state, and protein coding annotation. The resultant 624 classes and attendants scores, are shown to predict eQTL, transcription factor binding, and enhancers as well as or better than contemporary methods using high dimensional functional covariates, or selective constraint alone. The interpretation of the score as selective pressure is also shown to be consistent with previous measures of genome-wide selective pressure. In the second paper, I expand the cell-type cohort to 115 Epigenomic Roadmap cell-types and nine genomic properties including splicing, transcription factor binding, and small RNA-seq. Complexity constraints are developed to reduce the number of functional classes from more than 1.2 million possibilities to 61. The resultant functional classes, genomic segmentations, and positional scoring (FitCons2 scores) are used to detect small features including disease associated variation from HGMD and ClinVar clinical databases. FitCons2 scores are shown to have power comparable or superior to contemporary methods designed specifically to detect such features. Functional classes and scores are also shown to identify cell-type specific regulatory behavior of promoters and enhancers, while highlighting regulatory relationships between differing cell-types and developmental stages. I demonstrate how cell-type sensitivity in FitCons2 scores can be used to address an unsolved biological problem in characterizing transcription factor binding in craniofacial enhancers that are believed to differentiate neural development between humans, chimpanzees, and Neanderthals.

Scientific collaboration between researchers is very common and much influential and ground-breaking research is performed by teams comprised of scientist from different fields and organizations. In this thesis, we analyze and model a small scientific collaboration network limited to two organizations: Wright State University and the Air Force Research Laboratory. Research paper co-authorship is used for establishing the network structure.We analyze several network properties and compare them to past results from analysis of larger and more diverse collaboration networks. We show that the two-organization network we explored exhibits properties similar to those of larger networks.Guided by advances in state-of-the-art algorithms for the link prediction problem in large-scale networks, we explore modeling of the local network via similar methods. We use a variety of link prediction algorithms and models, from simple to state-of-the-art, and compare their accuracy. Results of our experiments suggest that simple and easy to calculate prediction methods produce robust results, outperforming the more complicated state-of-the-art models we explored. We observe a variety of methods producing very accurate predictions, which suggests these methods can be effectively used to solve practical real-world problems associated with small local or intra-organizational networks.

This research combined existing distributed data and graph techniques with a novel approach, refined over many experiments, to quantify social media engagement. The developed methodology scores engagement through measures including unique users reached, number of coverage items about which the user posted, and number and length of social post chains. In calculating scores, the use of quality measures goes beyond page rank's more limited concern with the centrality of a user within the social reference graph structure. The experiments used data comprised of scientific literature from the CORD-19 dataset, Scopus citation graphs relating to CORD-19, and PlumX altmetrics with corresponding Twitter posts intersecting with papers in the CORD-19 and Scopus citation graphs. The most significant outcomes of the research are two scoring algorithms, Social Media Engagement (SME) and Social Media Noise (SMN), which quantify complexity of engagement, or lack thereof, by users within social data. The incremental design allows new data to be added to a cumulative score without recalculation. SME and SMN algorithms have wide applicability for all social data at any scale or velocity and could become the basis for a new class of altmetrics.

This thesis is an investigation into the performance ramifications of making specialized component reservations for machine learning workloads in virtualized environments. The reliance of machine learning on floating-point operations makes graphics processing units an important part of processing these workloads quickly. While virtualization is one of the most widely-used consolidation techniques used in data centers of all sizes, compatibility issues between graphics processing units and virtualization have slowed the adoption of virtualization for machine learning workloads. To that end, this paper discusses the motivations and history behind virtualization and the application-specific acceleration devices used and how they are applied to machine learning on various public and private computing platforms. This is followed by a presentation of an experimental framework for testing the impact of controlling for non-uniform memory access when running machine learning workloads. Using this framework, a series of experiments were performed and documented in this thesis that test multiple placement configurations for graphics processing units in a virtualized system and how throughput of data from the host system to the device was affected. Current virtualization platforms offer recommendations to use these settings, but do not talk about the specific impacts of implementing them. Based on the results of the experiments, configuration parameters and placement recommendations are presented along with information about how these settings can help optimize the machine learning pipeline and the potential pitfalls to their use.