Our research presents a k-Nearest Neighbors (kNN) particle filter for state estimation of a Darwin-OP humanoid robot. Robots need quick and accurate state estimation for real-time control; you can not move if you do not know where you are. While many current methods rely on extremely accurate but often cost-prohibitive sensors, a good algorithm can make up for cheap equipment and inaccurate models. We have chosen to use a particle filter (PF), which works by sampling a large number of “particles” to capture the probability distribution of possible positions. This makes it a good choice since it attempts to estimate the true distribution rather than approximate it with a Gaussian and is a global estimator making more robust. Our research expands on the basic PF idea by combining with a kNN database, which precomputes the particles offline and caches the data. During online estimation, we can then compare our sensor data to particles in the kNN database to form a position estimate rather than simulating the dynamics of each particle, thus increasing computational efficiency. This increase in efficiency allows us to sample a greater number of particles, increasing the accuracy and robustness of our estimator. We first constructed a basic working PF and experimented with various weighting schemes and distance functions to produce the most accurate results. We then built a kNN database and fill it with gathered position and sensor data. Our algorithm’s estimation ability is evaluated by comparing it’s performance with our previous work of an Unscented Kalman Filter. While the general particle filtering technique is useful for estimation, adding a kNN database broadens its application to dynamical systems. Future research will focus on such applications such as system identification and real-time control of the Darwin.

Object recognition in images is used in a variety of applications from the everyday example of consumer gaming technology to the more erudite autonomous vehicles and industrial robotics. Recently, technical breakthroughs have increased the accuracy rate of these technologies. Even so, there are still numerous situations where this technology cannot yet recognize all the objects present; the smallest change in lighting or overlapping objects can alter the predictive outcome. We should strive for object recognition techniques that are not only highly accurate but also fast enough to be embedded in real-time applications. This project compared such diverse techniques as Convolutional Neural Networks (CNN), Decision Trees, Histogram of Gradients, and Graph Matching and measured how each of them performs according to the following metrics: precision, recall and milliseconds. We anticipated that using CNNs and their permutations will have a higher accuracy than the more traditional methods. We asserted that the rate of adoption and proliferation of these object recognition technologies will increase. Now that these techniques have reached a modicum of accuracy, researchers will continue to pursue ever-more-accurate algorithms. Furthermore, because accuracy correlates positively with safety, this pursuit will become even more relentless. This project supported the fact that the CNN should be preferred over other current object recognition technology until a higher accuracy technology is found; future work can build upon this project in new areas such as the medical field or space exploration.

A standard QWERTY keyboard for text entry may not be suitable for people with mobility related disabilities. As such, people with mobility related disabilities may need to find alternative ways to input text. Alternative and Augmentative Communication (AAC) is a field that researches and develops communication methods for people with complex communication needs. The AAC field studies technological solutions and its social implications. Strategic competence is an area of AAC research that focuses on managing the functional constraints of AAC devices such as text entry speed and input accuracy. Social competence is another area of AAC research that focuses on establishing and maintaining interpersonal communication. The ability to effectively communicate emotion plays a large role in one’s social competence. An ideal alternative text entry input system should not only consider strategic competence, but also the social competence. This research will survey existing alternative text entry systems and how emotion is conveyed through standard text/text related systems. The results from both surveys will be synthesized to inform an investigation of new methods for facilitating the communication of emotion through alternative text entry systems. I anticipate discovering new methods of conveying emotion through alternative text entry systems.

Industries have been slow to digitize their equipment and processes. Until now, doing so has not been as accessible due to factors such as price along with the size and efficiency of the equipment with which to do so. The Internet of Things (IoT) industry is projected to grow to over 30.7 billion devices by 2020. Such a huge number of devices have the potential to become a threat to Internet security. This was seen in the attack that occurred on Oct. 21, 2016, when hackers used IoT devices to carry out large-scale DNS attacks. This example illustrates the importance of gathering and analyzing data from these IoT devices. To better understand the problem, we are implementing the use of IoT devices to gather data and populate it into a data warehouse for further analysis. We used wireless antennas to gather network data, then populated data into a data warehouse. Finally, we extracted data from this warehouse to answer questions such as, "how crowded is the wireless spectrum" and "what is the distributribution of traffic on a network"? The sensor used in this project was a Raspberry Pi paired with a USB wireless antenna. The data warehouse used an unstructured, NoSQL database. The analysis tools used were Excel and Python. The Python programming language served as the link between gathering, population and analysis of data.

Spirometry is among the most common tests for characterizing pulmonary function and plays a critical role in detecting and improving outcomes related to chronic lung disease. During a spriometry maneuver, a patient rapidly exhales into a machine which records the total volume and flow of air over time. However, patient error in performing the spirometry maneuver, such as by coughing or taking multiple breaths, can lead to clinically misleading results. As a result, spirometry must take place under the supervision of a trained specialist who can detect and correct patient errors. To reduce the need of specialists to coach patients during spirometry, we demonstrate the ability to automatically detect four common patient errors. Applying machine learning techniques on features derived from spirometry data, we were able to automatically classify errors from spirometry data with an F-score between 0.784 and 0.907. Our work provides the first step toward reducing the need for trained individuals to administer spirometry tests by demonstrating the ability to automatically detect specific errors and provide appropriate patient feedback. This will increase the availability of spirometry, especially in low resource and telemedicine contexts.

Financial inclusion - the expansion of access to formal bank accounts, particularly to the poor - has recently moved up the ranks as a priority for development. After the unexpected success of M-PESA - a mobile phone based money transfer service started in Kenya - many quickly joined the camp of mobile money advocates believing that mobile money would be the solution to this newly minted UN sustainable development goal. However, with such nuanced characteristics in each country and even more distinctions within each financial system and economy, mobile money as a solution may not be applicable in every country aiming for financial inclusion. Therefore, this research aims to primarily understand the financial practices and needs of Ghanaians. To do so, I conduct surveys in five different cities ranging from small towns to large cities to evaluate an individuals current financial practices, existing and projected needs for financial services, and the role that mobile money may play in fulfilling either of those goals. 

The ability to form an understanding of an environment is essential for a mobile robot, enabling it to plan and execute actions efficiently. Such understanding should consist not only of geometry of space, but include notion of high-level semantic spatial concepts, such as objects, landmarks, or functional room categories. The goal of this project is to enable mobile robots to learn a unified deep generative model of the robot's environment that spans multiple levels of abstraction. Training deep models for robotics applications is often challenging. It requires large amounts of annotated data which is tedious and costly to collect in real environments. In this work, we investigate the benefits of utilizing knowledge generated in simulated environments, in which robots can be deployed significantly faster and in parallel, for improving the quality of a learned spatial model. Specifically, we generated random artificial simulated worlds and perform initial pre-training of the model there. To this end, we made following contributions. First, we improved the robustness of the navigation system of our robot DUB-E, which allowed us to deploy the robot in cluttered, human-populated environments. Second, we implemented an algorithm generating random simulated worlds for automatic data collection. Third, we designed a low-level representation of geometry of space around the robot, which serves as input to the deep spatial model. Finally, we collected and processed a large dataset of sensory information captured in simulation and the real world. We demonstrated the impact of the additional data generated in simulation on the performance of the deep spatial model.

Modern linguistic theory has held that language is arbitrary; that is, there is no connection between the physical form (spelling or pronunciation) of a word and its meaning. However, recent work has used distributional semantic vector representations of words to verify non-arbitrary form-meaning pairings (called phonesthemes) in English that were proposed in past psycholinguistics literature. One notable phonestheme is the prefix "gl-", which occurs at the beginning of several English words relating to light or vision (e.g. "glow", "glint", "glisten"). In this work, we use sparse regression to quantify this non-arbitrariness by extracting phonesthemes from word vectors. We also conduct human subject studies with a cloze-style task with non-English words containing phonesthemes, in order to verify that candidate phonesthemes have psychological reality in the lexicon of native speakers. In addition, we explore phonesthemes of other languages and across languages through multilingual word vectors. Uncovering the true extent of non-arbitrariness in human languages is important for furthering understanding of language comprehension and production, as well as advancing pre-existing linguistic theory.

In order to provide high availability and operate in unreliable environments, many critical applications are implemented as distributed systems. Unfortunately, the need to handle unreliable networks, machine crashes, and churn--the spontaneous arrival and departure of nodes to and from a network--has made these systems difficult to implement correctly. Recent work has proved the correctness of such implementations using proof assistant software. However, these frameworks do not support reasoning about churn. To address this limitation, we introduced support for implementing and verifying systems under churn to the Coq-based Verdi framework. With a model for the network in hand we could implement and begin to verify systems. However, for most systems under churn we need to prove punctuated safety properties. These properties allow a system to break some promises during and after churn, but stipulate that safe behavior be restored sometime after churn stops. Such a property is best proved by reasoning about infinitely long system executions, but this was impossible in Verdi and always awkward in Coq. We built tools for high-level reasoning about infinite sequences of events in Verdi using existing ideas from linear temporal logic in order to prove punctuated safety. To establish the efficacy of all these new tools, we have built an implementation of the Chord distributed lookup protocol and are currently proving it correct. Chord was designed to evenly distribute resources or tasks across a set of nodes even as some enter and leave the network. Recently, a study using a formal model of Chord's protocol showed that certain kinds of node failures or joins could cause system failure. While a new specification has been developed which avoids these problems, it has not been implemented or fully verified. We are developing the first verified implementation of Chord.

Augmented reality (AR) technologies are being actively developed, from head-mounted displays such as Microsoft’s HoloLens to AR-enabled car windshields by companies such as Hyundai. These emerging technologies, which integrate virtual content with the user’s perception of the real world, have the potential to greatly benefit their users. Applications running on an AR platform, however, may be malicious or malfunctioning, and so could do harm to the user: for example, occluding safety-critical real world information, startling the user, or collecting sensitive data about the user’s surroundings. These risks may be further compounded by multiple applications running concurrently. Though previous work has examined the risks of applications receiving unfiltered sensor input, here we examine the risks of unregulated application output. We propose enabling the operating system of an AR platform to enforce safety and security properties of applications in dynamic, multi-app contexts. In this work, we prototype an output policy module in a simulated AR operating system. The module mediates applications’ visual output, filtering based on policies specified by the system, user, and applications: for instance, blocking rapid resizing of objects or occlusion of pedestrians crossing the road. We deploy the simulated operating system in several mock scenarios that reflect proposed everyday uses of AR, including home, office, and automotive settings. Our results demonstrate that the output policy module successfully blocks negative application behaviors while permitting desirable functionality. We show that our solution is both effective and practical: our output policy framework is expressive but does not impose an undue burden on developers, and the computational overhead for application performance is modest. Our solution thus presents a promising step toward designing more secure AR systems.

Today's problems in fields like biomedical research require complex analysis of vast amounts of data in a short window of time, and the state of the art in computer science is not up to the challenge. The status-quo approach to such computationally demanding problems has been to design a custom algorithm, but this process is difficult, time-consuming, error-prone, expensive, and rarely yields the best solution. To overcome these challenges, computer science is undergoing a paradigm shift, from solving problems with individual algorithms to solving entire domains of problems with automatic generation, or synthesis, of algorithms. In particular, many problems boil down to scheduling or, in other words, planning when to compute what. For instance, a data center must plan how to divide a given complicated task across multiple machines. When this plan or schedule is created ahead of time, the process is known as static scheduling, and this schedule constitutes an algorithm specialized to the problem. Until now, static scheduling for every new problem domain has required either a complex, custom algorithm or a complex, custom model for a logic solver. Even in simple cases, these existing approaches are highly susceptible to human misinterpretation of requirements and necessitate complex, error-prone engineering to scale to real-world usage. The situation is worse still if the schedule needs to be optimized for efficiency. My research with Professor Bodik automates the creation of scalable tools for static scheduling from simple code to interpret the meaning of a candidate schedule. This work enables application of static scheduling to sophisticated problem domains while maintaining high confidence in schedule correctness. I applied this framework to create a static scheduling tool for hierarchical computation, supporting advanced and notoriously difficult techniques like parallel and incremental computation. With this tool, I created high-performance algorithms for data visualization and web browsers.

The Darkweb is a section of the internet only accessible by special TOR browsers that anonymize users' IP addresses. There are markets on the Darkweb where people buy and sell illicit goods and services using Bitcoin as their currency. Both the Darkweb and the Bitcoin blockchain are considered anonymous and independent of each other. There are patterns that we believe present themselves similarly in both networks through the growth in popularity and activity in nodes of the networks. The first step necessary is to create “personas” that represent the actual people behind the different users on all the Darkweb forums in order to create a network graph. We are then able to create the network graphs that we will be able to use the Page-Rank algorithm and other similar network centrality analysis to determine the growth of the two networks over time. All of this will allow us to determine which subsets of Bitcoin transactions are related to the Darkweb. After an initial thematic analysis our team was able to determine different interactions that would lead up to a sale of a product and thus a Bitcoin transaction. All of this work allows our team to start and determine exactly how well the security of the crypto currency and crypto markets work in conjunction with each other. By determining how easily we can connect these two groups of users we either show that the security measures used within the systems and communities are as secure as believed or that they can be defeated. This leaves an opening for future work into similar crypto markets and currencies that usual similar systems to that of Bitcoin.

Additive manufacturing - the process behind 3D printing - has been growing in widespread use since the introduction of the first commercially-available desktop printer and later improvements in cost, quality, and accessibility. Expanding horizontally and vertically in the global market, the 3D printing industry has been one that not only appeals to hobbyists, but has demonstrated potential as a solution to a variety of challenges faced around the world today, including those in biotechnology, aerospace, and industrial engineering. Seeking to take advantage of the increasing demands in 3D printing technologies, a number of companies and open-source organizations are offering new types of 3D printers and associated software. However, the rapid growth and diversity of the products being released have resulted in a loss of universal standards. Often indented to be accessible and reliable regardless of a user’s technical background, 3D printing technologies have been known to require consumers to learn the more technical aspects of their printer, G-Code, slicing software, materials, etc. in order to configure their overall setup and reduce the prevalence of errors. Our project aims at using advanced parsing strategies, program optimization, and statistical analysis to provide a universal, easy-to-use tool that both analyzes and optimizes G-Code files prior to printing. By automating the common fixes users are required to implement in their G-Code files and identifying common factors in 3D prints (material warping, inadequate temperatures, speed of extruder heads, etc.), this tool is intended to be usable for any consumer, regardless of engineering or design experience. Ultimately, our goal is to help provide a universal solution to the lack of consistency between products, and help leverage the promising field of 3D printing to reach its potential as both a home essential and reliable solution in fields such as biotechnology and industrial engineering.