CSSE Senior Thesis Poster Presentations

• Samuel Flickinger
  • FITting Tree: A Data-Aware Index
  • Abstract. In modern day database management systems (DBMSs), indices are vital structures used to optimize searches. It is estimated that up to 55% of the total memory consumed by a DBMS can be dedicated to these index structures. In an attempt to reduce overhead, the FITting tree was born; in a FITting tree, blocks of data can be grouped into segments that can be described by linear approximations simply composed of a calculated slope and starting key. To find improvement upon other types of indices, we compare a cached, disk-resident FITting tree to a cached, disk-resident B+ tree through the use of TPC-C inspired testing.
  • poster
  • 5-min narration of poster
• Cherise McMahon
  • Reimagining Password Creation: Creating Strength Through Prediction
  • Abstract. Users struggle to create strong and unpredictable passwords due to misconceptions, laziness, and other factors. These misconceptions can often be attributed to their time interacting with password creation tools and complexity requirements, which begs the question if adjusting tools to have negative, rather than positive, feedback can create stronger passwords and avoid misconceptions. Many users do want to have a strong password, but they struggle to understand what makes passwords secure from what makes passwords predictable. By creating a password tool with explicit feedback, users do not have to go through the process of asking themselves what is secure and instead can work with the tool to develop a secure password. This cohesion with the tool makes the predictable concerns explicit, so users can instead understand what is secure. Preliminary results show that users make strong passwords with the tool and after using it, yet also make weaker passwords with the tool when having created a password before using it.
  • poster
  • 5-min narration of poster
• Jessica Myers
  • Music Generation With Deep Neural Networks Using Flattened Multi-Channel Skip-3 Softmax and Cross Entropy
  • Abstract. This thesis is an investigation of whether a computer program can generate human-plausible, goal-oriented music. Deep neural networks were chosen as a focus because the way in which they learn seems to parallel the way a person passively learns music throughout life, whether or not that person knows music theory. The raw dataset of MIDI files was acquired from 17,216 song clips by 4825 artists from Hooktheory's TheoryTab Database. A custom dataset was created by encoding the MIDI files into sparse matrices and sliding a fixed window over each song to generate sequences. A novel approach within the domain of music generation was employed using a custom softmax activation function. The current results of generating music, given a seed, using a fully-connected network and a convolutional network showed some evidence of rhythmic and harmonic patterns, but lacked melodic elements.
  • poster
  • 5-min narration of poster
• Vanshika Reddy
  • An Emotionally Aware Dialogue System with Memory
  • Abstract. This project proposes a unique way to build a memory along with analyzing the user’s input to generate emotionally appropriate responses. The model contains two important and distinct features: generating a mental model, which acts as memory for the system and analyzing the intensity of the emotion. By remembering instances and intensities of an emotional event, the model tries to etch out an emotional profile of the user which is a key input in the response generation process. The contribution of this project is a system to better understand human emotions and provide human-like emotional assistance
  • poster
  • 5-min narration of poster
• Andy Sadler
  • Examining Memory Safety Regressions in Open Source Software"
  • Abstract. Memory safety regressions are the intersection of two important topics within software: memory safety bugs, which account for over 70% of bugs seen by both Microsoft and Google Chrome, and software regressions, which can indicate underlying problems with your code. I investigate the rate that memory safety regressions occur within Firefox. This was accomplished by mining bugs within Firefox's bug tracker and analyzing them to determine the regression rates within Firefox and how long it took for these bugs to regress. From this, I conclude that the rate at which these regressions have been occurring has been increasing across the history of Firefox.
  • poster
  • 5-min narration of poster
• Austin Swatek
  • Music Classification and Recommendations System Avoiding Genre Selections
  • Abstract. The most common method of recommending songs to users on a music platform is done through collaborative filtering, a system that relies on the data of both the primary user and other users on the platform. A main flaw of a system that relies on analyzing the information of other users to recommend music is that it is not actually personalized for the primary user. Genres are often too broad, even when broken down to subgenres. By analyzing the actual audio characteristics of different songs, it is possible to identify songs with related characteristics that would be labelled as separate genres, even though they sound similar, and might not be identified by a traditional recommendation system.
  • poster
  • 5-min narration of poster
• Max Wang
  • Data-Efficient Reinforcement Learning Based on Koopman Operator Theory
  • Abstract. Reinforcement Learning has shown great potential in decision making and control problems, and has been widely used in fields such as robotics, autonomous vehicles, and natural language processing. It still has the disadvantages of extensive training time and low data efficiency. Compared to model-free reinforcement learning, model-based reinforcement learning achieves higher data efficiency, but its performance depends heavily on how well the model is developed. Good models are generally hard to learn, and they also increase the complexity of the underlying control problem. We propose a data-efficient model-based reinforcement learning methodology, using Koopman Operator Theory to solve the optimal control problem of nonlinear dynamics systems. We also introduce Operator Sampling to our approach, creating a better estimation of the global model using multiple samples of local models.
  • poster
  • 5-min narration of poster
• Tony Xi
  • Dynamic Auction End-price Forecast With Neural Networks
  • Abstract. The purpose of this research is to develop a system that can dynamically forecast the outcome of an auction when given the historical bids. The underlying architecture of the system is a Long-Short Term Memory Network with an Encoder-decoder structure. The system also includes a feature engineering unit that extracts functional objects from discrete bid amounts to form price velocity and acceleration, which are then used as parallel inputs for the neural network. The model has an accuracy around 70% when predicting the auction dynamics continuously, and an accuracy of 85% when predicting only the end-price.
  • poster
  • 5-min narration of poster