|
" Domain Knowledge Aided Explainable Artificial Intelligence "
Islam, Sheikh Rabiul
Eberle, William
| Document Type
|
:
|
Latin Dissertation
|
| Language of Document
|
:
|
English
|
| Record Number
|
:
|
1106747
|
| Doc. No
|
:
|
TLpq2411479372
|
| Main Entry
|
:
|
Eberle, William
|
|
|
:
|
Islam, Sheikh Rabiul
|
| Title & Author
|
:
|
Domain Knowledge Aided Explainable Artificial Intelligence\ Islam, Sheikh RabiulEberle, William
|
| College
|
:
|
Tennessee Technological University
|
| Date
|
:
|
2020
|
| student score
|
:
|
2020
|
| Degree
|
:
|
Ph.D.
|
| Page No
|
:
|
126
|
| Abstract
|
:
|
In the age of the Internet of Things (IoT) and ''Big Data'', a voluminous, heterogeneous stream of data from billions of internet-connected devices has necessitated the adoption of Artificial Intelligence (AI) based models in many real-world applications. Although the capability of learning very complex functions has made these models robust, most of the successful models are ''black box'' in nature as they lack the ability to explain the decision process in human terms. As a result, this leads to ethical and trust issues in critical applications of relevant domains (e.g., Health-care, Security, and Finance) for potential implications to human interests, rights, and lives. Research suggests that a multidisciplinary effort and leveraging of useful domain knowledge could lead to a viable solution to explainable AI systems that produce human-friendly decisions. However, Explainable Artificial Intelligence (XAI) is still an emerging field of research, where uncovering domain-specific useful knowledge is challenging, and incorporating domain knowledge in a ''black box'' system for better explainability is underutilized. To address these issues, we propose and demonstrate a way to extract useful domain knowledge from the application domain, and incorporate that into a ''black box" model for better explanations of decisions. Our understanding from experiments on bankruptcy prediction and intrusion detection reveals that the incorporation of domain knowledge makes the output of the ''black box'' model more explainable with negligible deviation in performance. In addition, the introduced generalization provides better execution time and resilience with unknown cases (e.g., attacks) while retaining most of the important information. Finally, although XAI lacks formalization and is an open problem, we propose and formulate an approach that suggests a reasonable and model-agnostic way to quantify the extent of explainability in XAI methods.
|
| Subject
|
:
|
Computer science
|
| |