Keywords

Explainable Artificial Intelligence, Artificial Intelligence, Transfer Entropy, Neural Networks, Large Language Models

Abstract

The rapid growth of artificial intelligence, in terms of both power and prevalence, motivates the need for new ways to study explainability in deep neural networks. This is especially true for transformer architectures and large language models, which are the driving force behind the current surge in artificial intelligence. The complexity of transformer architecture foils many pre-existing methods for studying explainability in deep neural networks. To address this gap in knowledge, this dissertation proposes a new method of studying explainability in transformer architectures that leverages insights from neuroscience, employing transfer entropy to map information flows between components of the transformer architecture. This dissertation will also address criticisms that transfer entropy has faced as a method, validating its use as a tool in this domain.

Completion Date

2025

Semester

Fall

Committee Chair

Garibay, Ivan

Degree

Doctor of Philosophy (Ph.D.)

College

College of Engineering and Computer Science

Department

Computer Science

Format

PDF

Identifier

DP0029770

Document Type

Thesis

Campus Location

Orlando (Main) Campus

Download

Share

COinS