Organizational Unit:

Undergraduate Research Opportunities Program

Permanent Link

https://hdl.handle.net/1853/70802

Parent Organization

Organizational Unit

Office of Undergraduate Education

ArchiveSpace Name Record

https://finding-aids.library.gatech.edu/agents/corporate_entities/1180

Full item page

Publication Search Results

Now showing 1 - 10 of 216

Intelligent Buffer Pool Prefetching

(Georgia Institute of Technology, 2023-01-18) Suresh, Sylesh Kyle

Buffer pools are essential for disk-based database management system (DBMS) performance as accessing memory on disk is orders of magnitude more expensive than accessing data in-memory. As such, one of the most important techniques for DBMS performance improvement is proper buffer pool management. Although much work has already gone into page replacement policies for buffer pools, relatively little attention has been paid to developing intelligent page prefetching strategies. Commonly used sequential prefetching strategies only handle sequential accesses but fail to predict more complex page reference patterns. More complex prediction techniques exist---particularly those that leverage the predictive power of deep learning. Although such models can achieve a high prediction accuracy, due to their size and complexity, they cannot deliver predictions in time for the corresponding pages to be prefetched. With the tension between timeliness and prediction accuracy in mind, in this work, we introduce a machine learning-based strategy capable of predicting useful pages to prefetch for complex memory access patterns with an inference latency low enough for its predictions to be delivered in time. When evaluated on a subset of the TPC-C benchmark, our strategy is capable of reducing execution time by up to 13% while a commonly-used sequential prefetching yields only a 6% reduction.
Forecasting Atlanta Gentrification with Transformers

(Georgia Institute of Technology, 2023-01-18) Sett, Gaurav

Gentrification is an impactful trend in American cities, yet our ability to measure and predict this process remains weak. This thesis explores the use of machine learning to predict gentrification in Atlanta, Georgia. We use a dataset of land parcels collected by Fulton County Tax Assessors and set out to forecast changes in local land value. Inspired by progress in natural language processing, we apply a machine learning model called a transformer to forecast gentrification. We find our model outperforms typical methods for time-series forecasting of gentrification, and we discuss the implications of our findings for future research.
Usability Study of ImageScape Integration in the Prolonged Exposure Collective Sensing System (PECSS)

(Georgia Institute of Technology, 2023-01-18) Zou, Chunhao

We studied the usability and usefulness of the Prolonged Exposure Collective Sensing System (PECSS) for clinicians conducting prolonged exposure (PE) therapy with a special focus on one functional component of PECSS called ImageScape, which allowed patients to share photos of their in-vivo tasks with their clinicians. The usability study returned positive results for the PECSS system as a whole, but yielded mixed results for ImageScape in particular, suggesting that further research into the utility of ImageScape is needed. We begin by introducing relevant psychological backgrounds such as PTSD and prolonged exposure therapy. Then, we include an in-depth description of the PECSS ImageScape software design and implementation. This is followed by the design of a usability evaluation protocol for the clinician, which attempts to measure the importance of ImageScape both subjectively and objectively, quantitatively and qualitatively. Next, we provide the results of the usability study with one of the clinicians from our partner university. We will discuss the implications of our study, compare how our system with other existing systems for PTSD treatment, address the limits of our study, and discuss the next steps and future studies in the design and implementation of PECSS. Finally, We conclude our study with a summarization of our contributions.
Efficient Calculation of Frame Level Complex Predicates in Video Analytics

(Georgia Institute of Technology, 2023-01-18) Sengupta, Aubhro

The field of video analytics focuses on extracting useful information from video. Lets consider a scenario in which we have a large amount of video from a traffic camera at a certain busy intersection and we are looking for a black sedan. State of the art object detectors such as FasterRCNN [3] utilize computationally expensive methods like convolutional neural networks that analyze a frame of video and estimate the number of the object of interest and the locations of every instance of that object in the frame. The most basic approach to solving this problem would simply be to execute the object detector on all frames of the video and collect the frames which contain at least one black sedan to return to the user. However, this approach is impractical on longer videos as CNNs are computationally expensive and thus too slow. Instead the number of frames evaluated by the object detector must be limited. This field focuses on developing strategies for doing so, such as sampling, filtering, proxy models, and clustering.
Fighting on Facebook: Political Conversations Between Strong and Weak Ties

(Georgia Institute of Technology, 2023-01-18) Bernstein, Lily

Most Facebook users have observed a heated disagreement on the site, or been involved in such a disagreement themselves. Why do we get into these disagreements and who are we fighting with? Are we learning anything from the disagreements we have online? In this research, we interviewed eighteen regular Facebook users about their experiences with conflict on the site. We found that people get into disagreements online because their \textit{expectations} about how the people in their network will react were violated. Conflict was often a result not of disagreement, but of breach of expectations. When conflict ensued, participants sometimes dissolved their relationship with their argument counterpart, and sometimes did not. We review strategies used for relationship maintenance. Weak-tie relationships were more susceptible to being dissolved, and strong-tie relationships were often salvaged by "agreeing to disagree" or ceasing political discussion. Participants report sometimes learning from these hard conversations, but not often. We follow-up with design recommendations for social media platforms that could help mitigate disagreements or give people access to tools that help them have productive, hard conversations.
Oracle Guided Image Synthesis with Relative Queries

(Georgia Institute of Technology, 2023-01-18) Helbling, Alec Fisher

Isolating and controlling specific features in the outputs of generative models in a user-friendly way is a difficult and open-ended problem. We develop techniques that allow a user to generate an image they are envisioning in their head by answering a sequence of relative queries of the form "do you prefer image a or image b?" Our framework consists of a Conditional VAE that uses the collected relative queries to partition the latent space into preference-relevant features and non-preference-relevant features. We then use the user's responses to relative queries to determine the preference-relevant features that correspond to their envisioned output image. Additionally, we develop techniques for modeling the uncertainty in images' predicted preference-relevant features, allowing our framework to generalize to scenarios in which the relative query training set contains noise.
Evaluating the Predictive Performance of Genomic Data-based Machine Learning Models for 4 Different Mental Health Disorders

(Georgia Institute of Technology, 2023-01-18) Choi, Katherine Elena

Clinical psychiatry can greatly benefit from using polygenic risk scores (PRS) to assess the risk of developing certain mental health disorders. While the PRS performance can be evaluated considering exclusively the disorder, we aim to leverage recent findings which state that mental health disorders may share genetic variants and have created features sets that are not only disorder-specific, but also encompass multiple mental health disorders. To evaluate the performance of these different features sets, we developed an automated polygenic risk score script that calculates the PRS of each patient in the UK Biobank, and a logistic regression script that utilizes a linear model to evaluate the performance. The predictive performance of schizophrenia and bipolar disorder showed significant improvement from the disorder-specific features set vs. the general ’Mental Health Disorders’ features set, suggesting that these two disorders may possess an overlapping polygenic architecture. This finding may help PRS become a robust tool used in clinical psychiatry to encourage earlier diagnosis of these disorders that greatly benefit from early treatment/intervention.
BrainBraille: Towards Passive Training in Brain-Computer Interfaces using fNIRS

(Georgia Institute of Technology, 2023-01-18) Gemicioglu, Tan

Amyotrophic Lateral Sclerosis (ALS) is a debilitating movement disability that causes patients to gradually lose their ability to voluntarily control their muscles. In some cases, patients who are "locked-in" are unable to move any muscles, leaving them with no means of communicating with caregivers. Brain-computer interfaces (BCIs) attempt to create a means of communication directly through brain activity, removing the need for movement. BrainBraille is a novel interaction method for BCIs, enabling complex text-based communication using attempted movements with a six-region pseudo-binary encoding. In this dissertation, I explore a wearable BCI using functional near-infrared scanning (fNIRS) to make BrainBraille mobile. In an early study, I show that transitional gestures based on executed movements of two hands can be classified in two participants with up to 93% accuracy. I explore how transitional gestures can benefit BrainBraille by expanding the vocabulary and enabling faster responses. Finally, I evaluate future paths for integrating passive haptic training into BrainBraille to reduce the physical exertion needed to learn a BCI for ALS patients.
Multidimensional Allocation: In Apportionment and Bin Packing

(Georgia Institute of Technology, 2022-08) Chiu, Alvin

In this thesis, we deal with two problems on multidimensional allocation, specifically in apportionment and in bin packing. The apportionment problem models the allocation of seats in a House of Representatives such that it is proportional to the dimensions being represented. One example is the allocation of the 435 House seats to the 50 U.S. states, which demands being proportional in the one dimension of state population. It is also common to demand proportionality in both state population and political affiliation, where we now have to allocate to two dimensions simultaneously. We begin by investigating what it means for an 1-D apportionment to be "fair", and use this to judge the various methods of apportionment that have been used throughout history. This motivates the study of divisor methods, a certain class of apportionment methods that avoid any paradoxes. We then formally tackle the problem of 1-dimensional and 2-dimensional apportionment with divisor methods through the lens of optimization. The optimization approach generalizes well to higher dimensions, but a proportional apportionment is not always possible in 3 or more dimensions. Our thesis outlines the current method for finding "approximate" apportionments and improves it in certain regimes. As for bin packing, we model the allocation of virtual machines to servers (with limited capacity) in cloud computing, with the goal of designing and analyzing efficient algorithms that optimize the expected cost of the allocation. This builds off previous work that only considered the case where the items being packed had a one-dimensional size. We extend some of those results to items with multi-dimensional size in this thesis.
Six Feet Apart: Online Payments During the COVID-19 Pandemic

(Georgia Institute of Technology, 2022-05) Shaikh, Omar

Since the COVID-19 pandemic, businesses have faced unprecedented challenges when trying to remain open. Because COVID-19 spreads through aerosolized droplets, businesses were forced to distance their services; in some cases, distancing may have involved moving business services online. In this work, we explore digitization strategies used by small businesses that remained open during the pandemic, and survey/interview small businesses owners to understand preliminary challenges associated with moving online. Furthermore, we analyze payments from 400K businesses across Japan, Australia, United States, Great Britain, and Canada. Following initial government interventions, we observe (at minimum for each country) a 47% increase in digitizing businesses compared to pre-pandemic levels, with about 80% of surveyed businesses digitizing in under a week. From both our quantitative models and our surveys/interviews, we find that businesses rapidly digitized at the start of the pandemic in preparation of future uncertainty. We also conduct a case-study of initial digitization in the United States, examining finer relationships between specific government interventions, business sectors, political orientation, and resulting digitization shifts. Finally, we discuss the implications of rapid & widespread digitization for small businesses in the context of usability challenges and interpersonal interactions, while highlighting potential shifts in pre-existing social norms.