Comprehensive Exam Presentation - Graham Annett

Explainability for Deep Learning

Graham Annett – Computer Science

Zoom link

Abstract:

While data driven models are nothing new, the recent advancements of deep learning in particular has resulted in the need for new metrics and methodologies related to explainability. These models which are combinations of layers of differing types of transformations result in highly predictive models while consequently being highly parameterized. While explainability in relation to machine learning and statistics is a well studied topic, many of these well tested previous methods are specific to more traditional machine learning models and provide much more intuitive and clear explanations of what is happening. Understanding how deep learning models not only transform an input into an output, but what information these parameters contain is an active area of research that when related to deep learning models presents many new topics of research.

Committee:

Tim Andersen, Ph.D. Chair
Casey Kennington, Ph.D.
Hoda Mehrpouyan, Ph.D.
Grady Wright, Ph.D.
Sole Pera, Ph.D.