How Can I Explain This to You? An Empirical Study of Deep Neural Net Explanation Methods - Spring 2021

Explore deep neural network explanation methods in this 37-minute lecture from the University of Central Florida's CAP6412 course. Delve into key contributions and study details, examining various visual explanation techniques such as saliency maps, scoped rules (Anchors), and SHAP. Investigate a unified representation framework for visual explanations, covering superimposition-based and training data-based methods. Learn about the study methodology, including task design, dataset selection, and model explanations. Analyze results on usability, stability, and privacy risks associated with different explanation approaches. Gain insights into the challenges of explaining complex neural networks and the implications for AI transparency and interpretability.

Intro
Outline
Motivation
Introduction
Key Contributions
Study Details
Unifying Visual Explanation Methods Across Input Domains
Saliency map
Scoped Rules (Anchors)
SHAPISHapley Additive exPlanations
A Unified Representation of Visual Explanation Frameworks
Superimposition Based Explanation Methods
Training Data Based Explanation Methods
Study Methodology
Validating Responses
Tasks & Datasets
Models and Explanations
Configuring and Optimizing Explanation Methods
Results
Usability and Stability of Explanations
Idealized vs Actualized Explanations - Superimposition Methods
Explanation-by-Example
Privacy Risks
Conclusion
Against