CSSE Senior Thesis Archive

2020/21

• Brionna Slaughter.
  • Handwritten Authorship Attribution Using Both Image Recognition and Natural Language Processing Techniques
  • Abstract. Being able to attribute authorship of a short, handwritten note question could expand the possibilities of using other information besides textual content to achieve the high accuracy that has previously been achieved by using a large dataset. When looking at the potential impact of this authorship attribution software, it is important to note that in modern times, handwritten notes are often concise, thus this project focuses on maintaining accuracy for a small amount of short notes for each author. For the neural networks trained in this project, a Recurrent Neural Network was used as our main tool for processing the text, and a Convolutional Neural network utilizing ImageNet for transfer learning was used to analyze the corresponding handwriting. In order to test the results of our program, a variation of the IAM Online Handwriting Database was used, resulting in a final average accuracy of 72.4%.
  • A one-minute video introducing the work.
  • Thesis

• Sam Dunaway.
  • Comparing Transfer Learning Methods Deep Q Networks vs. NeuroEvolution of Augmenting Topologies
  • Abstract. Transfer learning is a field that significantly opens up opportunities to use artificial intelligence. Successful transfer can enable companies and researchers to use artificial intelligence where it may normally be prohibitively expensive, as gathering enough data to make the networks worth it can be prohibitively expensive. Transfer learning addresses this by learning an easier problem, and using that knowledge to learn a different, often more difficult problem. Transfer learning can be done with many different methods, and used with many types of problems. This paper compares how well Deep Q Networks (DQN) and NeuroEvolution of Augmenting Topologies (NEAT) transfer from a basic car racing game to a more difficult version, comparing both the initial transfer and the transfer after having trained on the more difficult version. Since the approaches are intrinsically different, I will discuss the experimental designs and my approach to provide fair comparisons to avoid misleading results.
  • A one-minute video introducing the work
  • Thesis