Dissertation Presentation

Dissertation Research

 

Title: An Initial Exploration into the Applicability of Transformer Architectures for Eye Movement Biometrics

 

Presenter:  Dillon Lohr

 

Advisor:  Dr. Oleg Komogortsev

 

Date/Time:  Thursday, December 8th @12:00 noon

 

Location:  https://txstate.zoom.us/j/91243653692

 

Abstract:

 

Transformer architectures have consistently outperformed other deep learning architectures in various applications, including image classification, speech and face recognition, and language modeling to name a few. Transformers typically require massive amounts of data before they can outperform other common architectures such as convolutional networks, though, due in large part to their relaxed inductive biases. The field of eye movement biometrics has seen great success with deep learning models, particularly convolutional neural networks, but it begs the question: could we achieve even better performance with a Transformer? We conduct an initial exploration into this question, using a convolutional head prior to a Transformer to inject some inductive bias back into the model and hopefully relax the data requirements. We find that we cannot match the performance of the state-of-the-art convolutional model, and the Transformer requires significantly more computing resources as well.

 

Deadline: Jan. 4, 2023, midnight