Explainable parts-based concept modeling and reasoning
No Thumbnail Available
Authors
Meeting name
Sponsors
Date
Journal Title
Format
Thesis
Subject
Abstract
State-of-the-art artificial intelligence (AI) learning algorithms heavily rely on deep learning methods that exploit correlation between inputs and outputs. While effective, these methods typically provide little insight to the reasoning process used by the machine, which makes it difficult for human users to understand the process, trust the decisions made by the system, and control emergent behaviors in the system. One method to fix this is eXplainable AI (XAI), which aims to create algorithms that perform well while also providing explanations to users about the reasoning process to mitigate the problems outlined above. In this thesis, I focus on advancing the research around XAI techniques by introducing systems that provide explanations through the use of partsbased concept modeling and reasoning. Instead of correlating input to output, I correlate input to sub-parts or features of the overall concept being learned by the system. These features are used to model and reason about a concept using an explicitly defined structure. These structures provide explanations to the user by nature of how they are defined. Specifically, I introduce a shallow and deep Adaptive Neuro-Fuzzy Inference System (ANFIS) that can reason in noisy and uncertain contexts. ANFIS provides explanations in the form of learned rules that combine features to determine the overall output concept. I apply this system to real geospatial parts-based reasoning problems and evaluate the performance and explainability of the algorithm. I discover some drawbacks to the ANFIS system as traditionally defined due to dead and diminishing gradients. This leads me to focus on how to model parts-based concepts and their inherent uncertainty in other ways, namely through Spatially Attributed Relation Graphs (SARGs). I incorporate human feedback to refine the machine learning of concepts using SARGs. Finally, I present future directions for research to build on the progress presented in this thesis.
Table of Contents
PubMed ID
Degree
M.S.