2019 Poster Session

Despite recent improvements, we still see women’s underrepresentation in computing fields. By presenting challenging courses in a way that makes learning more enjoyable, students will be more likely to consider computing careers later in life. This quantitative study assesses if gamification-using game elements-can improve students’ performance in a gamified online learning environment.

This study describes an initiative for the inclusiveness of the female undergrad CS students into competitive programming in a developing country through “female only” environment. Many of the female CS students face visible and invisible barriers on their paths to comply with their computing career. This initiative was designed to focus on how they can overcome those barriers of participation.

The study examined how faculty and graduate student mentors including undergraduate women enrolled in CS collectively articulated of the barriers and challenges women encounter in CS including ways to ameliorate these problems. Findings suggest that the mentors and students were equally aware of the challenges women faced in CS, and have crafted best practices and strategies to retain women in CS.

The participation of female students in computing fields remain low. This study examined the most important factors influencing female students computing occupation. The impact of family and friends on four different racial groups including White, Black, Hispanic, and Asian (n=1650) were examined. Results indicated friends had a positive significant role on White female students major decisions.

More than 1 billion cars travel the streets of the world today. While cars have made our lives easier and comfortable, they have also made our lives riskier. Autonomous vehicles bring structure to the chaos. With the advancements in automotive industry and technologies like Blockchain, self-driving cars are getting closer to becoming a viable method of transportation that will make driving safer.

Reckless driving severely threatens the safety of people, which accounts for around 33% of all fatalities in major vehicle accidents. Most existing efforts focus on the detection and adjustment of a vehicle’s own driving behaviors with limited effectiveness. We propose a defensive alerting system to automatically and intelligently detect and notify the threats posed by reckless vehicles.

This paper uses customer reviews to extract keywords for a product and gives sentiment score using ‘SentiWords’ on top of extracted keywords. Reviews are scored using lexical chains, and WordNet provides synsets to identify relevant words. These keywords and sentiment score are useful for providing summary of customer sentiments without having to read the entire data and help in business analysis.

We propose a parallel attention network for style change detection. Given a document, it verifies if it has at least one writing style change. Unlike the conventional recurrent networks that use word sequences to learn the language model, our model learns the hierarchical structure of English language using a pretrained statistical parser. It stays at the second rank of PAN 2018 with 82% accuracy.

Tweets majorly containing words of idiosyncratic nature are difficult to analyze. In order to prevent loss of the information introduced by context and out-of-vocabulary words appearing in the training set, we design models capturing the specific nature of Twitter posts. Our hybrid CNN+RNN model trained on fastText embeddings outperforms current state-of-the-art models predicting message deletion.

A molecular tagging system with DNA based tags can provide a secure, difficult to fake identifier. An optimal system minimizes the time needed to accurately identify DNA molbits, which are unique DNA sequences. I created an error analysis pipeline that quantifies the minimum subset of reads on a MinION Nanopore sequencer necessary to produce a similar identification error rate as the entire set.

This task aims to assign protein organelle localization labels to protein samples. The ResNet34 model has been fine-tuned using different learning rates for different parts of the network. Focal loss is used as the loss function. Output is the set of target labels, and different thresholding schemes are evaluated for obtaining labels from class scores. Performance is evaluated on macro F1-score.

With the increase in smartphone usage, a new mode of data collection named mobile crowdsensing (MCS) has emerged. However, its open structure allows malicious entities to interrupt a system by reporting erroneous data. The goal of this research is to show that the introduction of a reputation system in the correlated sensor-based data fusion will enhance the overall quality of the sensed data.

The massive scale of Internet of Things (IoT) device deployment enables access to a wide gamut of data generated by these devices. Anomalies may occur where the devices fail to self-diagnose themselves. We aim to build a model from heterogeneous data and differentiate good situations from the bad. Our solution to anomaly detection of smart home devices is our aggregate model learning system.

The aim is to avoid energy and coverage holes using correlation and matrix completion in a network. Nodes deployed in a corona-like structure; coalitions are made using correlation among these nodes. The data of coalitions sent to the sink is processed using matrix completion to retrieve data of all nodes at the sink; that reduced the number of messages sent to sink; thus conserving energy.

The packet time information at receivers carries useful network information. It is challenging to accurately measure this information in a cloud network due to various factors such as virtual machine (VM) scheduling. We propose a novel packet timestamping mechanism and demonstrate its application using an improved available bandwidth estimation tool.

Kinetic simulation of plasma with PIC method involves numerical constraints on simulation parameters and time-scale making it computationally prohibitive on CPUs for large problem sizes. We illustrate hybrid parallelization of PIC code and compare the performance with shared-memory based parallel code. Our technique scales efficiently with increasing problem size and number of cores in cluster.

With modern CPUs, it is possible to process larger batches in parallel in latch-free B+ tree query processing model without extra delays. With increased batch size, there will be more optimization opportunities exposed beyond parallelism, especially under highly skewed query distributions, such as avoiding redundant and unnecessary queries. We propose QTrans to efficiently solve this problem.

Graph mining is of great importance in solving real-world problems in many application domains. Community detection is a computationally challenging problem, solved efficiently with Louvain Algorithm.We design parallel algorithms using OpenMP, MPI,use dynamic load-balancing to achieve around 12-fold speedup.We also present a brief overview of our work on temporal graphs.

Intraventricular hemorrhage (IVH) is a complication in premature infants. Babies that develop hydrocephalus as a result of IVH often require a permanent ventriculoperitoneal shunt. Here, we will identify RNAs correlated with shunt placement and develop a predictive model based on RNA expression. Our predictive model could potentially assist clinical decision-making for patients with IVH.

We examine a novel virtual reality (VR) based experimental setup to study the anxiety-related responses in balance-demanding target-reaching leaning. An EEG and EKG data collection system is established to analyze the real-time responses from subject’s brain and heart. The Brain-computer interface system integrating VR and online anxiety feedback provides the technology to study task performance.

We develop a data-driven treatment recommendation framework to predict physicians’ prescription effects, learn and improve the doctors’ prescription policy by optimizing over the estimated policy model. With the recommended prescription policy, diabetic patients see significant blood glucose reduction on average and obtain a better therapeutic effect than the state-of-art deterministic algorithms.

While pervasive computing technologies becoming mainstream for observing human behavior, few formal models have been theorized of behavior based on collected sensor data. We construct Markov models that express the nature of human behavior captured by smart home sensor data. We will evaluate if behavior becomes less Markovian and more random as individuals experience cognitive decline.

Cells are multifunctional units of life that coordinate tasks by generating intracellular domains of organelles and macromolecules, such as proteins and RNA. We developed an open-source approach to precisely analyze the distribution of individual RNAs relative to centrosomes, organelles involved in cell proliferation. Our approach can be generalized for spatial analysis of other microscopy images.

A key enabling technology for spatial applications is understanding of the device’s location and the map of environment, referred as Simultaneous Localization and Mapping (SLAM). Visual SLAM algorithms suffer from perceptual aliasing and high computational cost. Since WiFi routers are ubiquitous in urban environments, we utilized the WiFi received signal strength to alleviate those challenges.

We consider the problem of distributed control in large networks of mobile robots with limited communication. By extending aggregation graph neural networks to time varying signals and networks, we learn a common local controller which exploits information from distant teammates using only local communication. We demonstrate the performance of the decentralized flocking controller in simulation.

In this poster, we highlight how an automated bot can be a great solution for incident management in a company and a better solution than traditional incident management paradigms. To support our claim, we introduce Catbot, the incident management bot for our team in Bloomberg, its architecture, and how using it has had a big impact towards making incident management more effective in our team.

Using a browser’s ability to cache resources eliminates the time to download resources and allows an application to load faster. Most engineers consider ways of using browser cache when designing a resource deployment strategy. This poster will highlight the cost of loading resources from a disk cache, the conditions when cache costs are high, and ways to reduce this cost.

Mobile app companies utilize third-party tracking service to learn their app users’ behavior. An app developer could set user attribute on tracking. This paper studies what attribute did app developers choose for tracking, whether they choose personally identifiable information (PII) such as social security number or email. We performed a study on 1,031 apps and found 125 use unencrypted PII.

Control-Flow Hijacking is the most common attack vector against C/C++ programs. Automatically specializing code for particular usage scenario is a promising new technique for software security. I present Ancile which implements required functionality analysis, achieves significantly smaller subset of functions by specializing the code, and substantially raising the bar for CFH attacks.

he end-to-end approach is one of the frequently used approaches for the autonomous driving system. In this study, we adopt the end-to-end approach to train a deep convolutional neural network to steer a car in the Udacity simulator. Our proposed model map raw pixels from cameras of three different angles and generate steering commands for autonomous driving.

Smart home, a ubiquitous environment aims to offer Ambient Assisted Living. The activity modeling framework integrates ambient intelligence into the home environment. This research proposes an assistive dementia care system through smart home that offers functional assistance to the dement occupant during critical situations without the help of caretaker.

The American public never receives information about the vast majority of cyberattacks against U.S. industrial systems that control the operations of refineries, pipelines power plants, chemical plants, and other factories. These incidents are classified by the FBI and other agencies, and kept secret by U.S. companies unless a breach involves the theft of personal identifying information.

Actigraph watches are a tool through which we can view the minute-by-minute sleep activity of unhealthy populations with minimum intrusion. While this data can be quite noisy, machine learning can help us analyze these irregular patterns. Through a machine learning approach, we predicted nighttime sleep duration for a group of post-stroke and post-traumatic brain injury patients.

Increasingly, Instagram users maintain both public- and private-facing accounts, known as rinstas and finstas, respectively. This poster explores users’ goals in maintaining multiple accounts and how accounts interact with one another. I developed a system to compare and analyze differences between users’ rinstas and finstas with respect to content, network structure, and social metrics.

In character animation, it is important for many applications to know how similar two motions are. For this, the Root Mean Square Error is a commonly used distance metric. However, it does not always correctly map to our perception of similarities. In my research, I check several types of motion errors with the same similarity values. I investigate how these motions are perceived to help develop more perceptually valid distance metrics in the future.

We present SSS, a scalable transactional key-value store deploying a novel distributed concurrency control that provides external consistency and never aborts read-only transactions due to concurrency. SSS uses a combination of vector clocks and a new technique, called snapshot-queuing. Results show significant speedup over state-of-the-art competitors in read-dominated workloads.

In this poster, we will discuss about the problem of detecting latent concurrency bugs in partially synchronized distributed systems. Partially synchronized distributed systems do not have a specific total order of events in the system. So evaluating all possible serializations of events in the system to guarantee correctness is a challenging task and we propose to use SMT solvers to handle this.

Low-level languages for circuit descriptions, like VHDL, easily lead to errors in large designs. Much higher-level tools like Simulink yield fewer errors but lack modularity and efficiency. We show how language features of Haskell, like high-order functions, parametric polymorphism, and type classes, can provide a high-level design experience that reduces errors without sacrificing efficiency.

Over the last two decades, the number of computer systems has exponentially increased. The number of mobile users has quintupled since the early 2000s. With this increase is the increased demand for power — a valuable commodity. This poster shows how software and hardware optimizations are used to design green, energy efficient systems that meet the aggressive low power requirements of desktop computers.

Fabrics are essential in our daily lives so designing and modeling them virtually is important. Capturing high-detailed appearance is challenging while ignoring the fiber-level details yield to a fake look for cloth. We introduced a new-of-a-kind technique to model cloth under physical forces. Our model accurately reproduce the change of appearance when fabrics interact with the environment.

Before films are released to the public, they must be inspected for defects. Currently, quality control teams must perform the majority of these examinations manually, which is not only time-intensive but also prone to human error. At Walt Disney Studios, we have developed an automated tool for aspect ratio shift detection that outperforms state-of-the-art, out-of-the- box software applications.