8.27 Flashcards
What are the goals of the course?
Learn and implement SOTA XAI models, understand interpretability/explainability, read and discuss XAI models, implement algorithms, and conduct a research project.
Who is the instructor for this course?
Nidhi Rastogi, Assistant Professor at the Department of Software Engineering, GCCIS, RIT.
What is Explainable AI (XAI)?
A subfield of AI that focuses on making the outputs of machine learning models understandable to humans.
Inherently explainable models vs. post-hoc explanations?
Inherently explainable models are designed for transparency, while post-hoc explanations interpret complex, pre-built models.
Why is model understanding important?
It helps in debugging, detecting biases, providing recourse, and assessing when to trust model predictions.
What are some tools used in XAI?
LIME, SHAP, and TensorFlow Explain are common tools used for creating explainable models.
What are some key research interests of the instructor?
Cybersecurity, Artificial Intelligence, Explainability, Graph Networks, and Autonomous Systems.
What are some examples of when model understanding is needed?
Debugging model errors, detecting biases, offering recourse for individuals, and deciding model suitability for deployment.
What types of evaluations are used in explainability research?
Application-grounded evaluation, human-grounded evaluation, and functionally-grounded evaluation.
What are the main auxiliary criteria in XAI?
Safety, nondiscrimination, and the right to explanation.
What is the difference between incompleteness and uncertainty in XAI?
Incompleteness refers to gaps in problem formalization, while uncertainty can be quantified, such as learning from small datasets.
What are the two main approaches to achieving model understanding?
Build inherently explainable models or use post-hoc explanations for existing models.
What are some conferences relevant to XAI?
ICML, NeurIPS, ICLR, UAI, AISTATS, KDD, AAAI, FAccT, AIES, CHI, CSCW, and HCOMP.
What is the significance of explainability in high-stakes settings?
It ensures that ML systems are not just accurate but also safe, non-discriminatory, and provide the right to explanation.
What are some examples of inherently explainable models?
Decision trees, linear models, and rule-based systems are examples of inherently explainable models.
