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 13

Static Exception Checker for Java Programs

(Georgia Institute of Technology, 2017-12) Nikolenko, Liubov

The purpose of this study is to present a tool, that can check if programs, compiled to Java Virtual Machine bytecode, are exception safe. The tool performs static analysis of input programs by reducing the programs to Horn Clause systems and solving the generated systems in the automatic theorem prover. The current version of the tool is publicly available in the Bitbucket repository. The tool can be used for custom constraint verification, safety checking and Horn Clause reduction.
High Performance Algorithms for K-mer Counting and Genomic Read Overlap Finding

(Georgia Institute of Technology, 2017-12) Zhu, Shaowei

Advancements in genomics are enabling a deeper understanding of how human body works and bringing us a new era of personalized healthcare. Genome sequencing and genome assembly are important procedures that facilitate genomic understanding whose quality affect nearly all downstream analysis. Modern high throughput sequencers could generate a few billion DNA subsequences with high accuracy in one experiment, proposing challenges to existing genome assembly pipelines that could only processes millions of reads per week. For the de Bruijn graph (DBG) based pipelines whose running time do not increase much with the number of input reads, efficient k-mer counting that is stable under the presence of repeats can serve as a useful heuristic for genome reconstruction. For Overlap-Layout-Concensus (OLC) pipelines, efficient sequence overlap finding is the first and most computationally intensive step in the process that must be improved to accommodate the huge amount of reads. To address these problems, this work presents a new de novo k-mer counting method that utilizes read pairs double matching, and a new approach to constructing the sequence overlap graphs based on locality sensitive hashing (LSH). We provide theoretical performance estimation of the latter method, followed by experimental verifications on simulated read data (large dataset contains about 1.25 billion reads). Our approach is the first overlap finder that could construct billion-scale overlap graphs. The experiment results for the de novo k-mer counting algorithm indicate it can provide a quite robust upper bound of k-mer occurrences. The experiment results for the overlap finding method suggest that our approach is at least 24 times faster (on the billion node size overlapping graph) and more flexible than one of the most influential overlap finders currently available, Minimap.
Story Generation with Deep Reinforcement Learning

(Georgia Institute of Technology, 2017-12) Dass, Nathan

Storytelling has applications in areas ranging from creating books and novels to engrossing movie scripts to the gaming industry. It would be useful to create interactive plot lines for games, motivate mission generation, and ordering of events for different characters. In the military, we can make the system learn from thousands of written real-world events, incidents, and missions to create interactive simulations to train army personnel. We introduce a deep reinforcement learning approach to story generation trained on a textual story corpus. Unlike other neural network based approaches to story generation, a reward function allows a human user to control the direction that the story follows.
Domain Adaptation in Reinforcement Learning

(Georgia Institute of Technology, 2017-08) Sood, Srijan

Reinforcement learning is a powerful mechanism for training artificial and real-world agents to perform tasks. Typically, one can define a task for an agent by simply specifying rewards that reflect the agent’s performance. However, each time the task changes, one must develop a new reward specification. Our work aims to remove the necessity of designing rewards in tasks consisting of visual inputs. When humans are learning to complete tasks, we often look to other sources for inspiration or instruction. Even if the representation is different from our own, we can adapt our own representation to the task representation. This motivates our own work, where we present tasks to an agent that are from an environment different than its own. We compare the cross-domain goal representation with the agents representation to form Cross-Domain Perceptual Reward (CDPR) functions and show that these enable the agent to successfully complete its task.
Detecting Mosquitoes with Convolutional Neural Networks

(Georgia Institute of Technology, 2017-08) Moore, Lawrence S.

Mosquitoes are directly responsible for the death of more than a million people each year. Yet the ability to mitigate their deadly impact or even monitor them in the wild to better understand their behavior remains relatively limited. One of the primary reasons for this lack of progress is the difficulty in locating and tracking an individual mosquito, leading to only estimates for a population as a whole. To address this problem, this research discusses several approaches using computer vision to detect and track the flight of mosquitoes. In particular, we discuss the performance of several convolutional neural network architectures which show promising results. Once these techniques are refined to give a high enough degree of accuracy, this vision system could be used in conjunction with drones to track and eliminate mosquitoes in both an indoor and outdoor setting.
George: A visual analytics platform for social media applications

(Georgia Institute of Technology, 2017-05) Animashaun, Damilola F.

Anonymous social media applications have experienced explosive growth in recent years. These applications often organically partition users into distinct communities creating intimate communities that differ significantly from one another. Unfortunately the analytics tools need to glean insights on the user generated content of this applications is somewhat lacking. George is a prototypical corpus visualization tool that aims to fill the gap in this space. By pairing simple information visualization techniques with natural language processing George provides a medium agnostic interface to explore the textual content generated by community focused social media applications.
Blending Fuzzing and Symbolic Execution for Malware Analysis

(Georgia Institute of Technology, 2017-05) Amiri, Addison O.

Malware infections have grown at least five-fold in the past five years. With an increase in IoT devices that are lacking built-in security, this problem is likely to only continue growing. Malware analysis, meanwhile, is becoming ever more challenging. Where manual analysis, symbolic execution, or fuzzing alone are overly time consuming or unfruitful, a combination of these techniques may offer promising solutions. This paper suggests a combination of fuzzing and symbolic execution to reverse engineer malware. A framework is described to tie these components together, producing test cases that call all functionality of a malware binary. These test cases show researchers the protocol used by the malware, as well as its capabilities, and allow for a reconstruction of the C&C server as desired. The goal of this work is to allow researchers to better understand malware and how to effectively combat it.
Modeling and Inquiry Learning Application: An Exploration of Spatial Simulations in an Inquiry Learning Application

(Georgia Institute of Technology, 2017-05) Hartman, Taylor

Modeling and simulation is a common technique to explore phenomena in ecology without having to spend an enormous amount of time with data collection and waiting on an ecosystem to develop. The practice of simulating an ecological phenomenon requires a good amount of practice and ability in scripting and programming. A lack of knowledge in the domain of computer modeling and simulation makes creating a working model and simulation of an ecological phenomenon difficult. The MILA system helps to bridge this difficulty by combining a conceptual modeler with a simulator. The research presented in this work expands on the MILA system to add a spatial simulation component to the conceptual modeller and the simulation compiler to provide further flexibility for users of this system. This thesis also provides an exploration of the changes needed to make to be able to simulate phenomenon in any domain beyond ecology.
One-shot Learning for Semantic Image Segmentation

(Georgia Institute of Technology, 2017-05) Liu, Zhen

Image segmentation has been one of the central topics in computer vision. Recent progress in deep learning methods and large-scale datasets provides opportunities for learning high level semantic representation of images. The recent success in one-shot video segmentation, in which the algorithm learns how to track or segment out some object in the video only given the first frame. This study aims to design an one-shot learning algorithm for image segmentation task, in which the algorithm learns given one or several segmentation samples.
Overlapping Clustering of Contextual Bandits with NMF techniques

(Georgia Institute of Technology, 2017-05) Kwon, Jin Kyoung

We introduce a novel approach to recommendation based on item clustering using non-negative matrix factorization (NMF) techniques. We propose a new algorithm, OCB (Overlapping Clustering Bandits), that groups items into latent clusters using online user feedbacks and uses learned clusters to make recommendations. By making recommendation at cluster-level instead of at item-level, the algorithm can overcome scalability issues associated with a large number of items without compensating for long-term reward maximization. Also, by performing online clustering of items, the algorithm can learn latent topics associated with items based on user feedbacks.