CSCI-B 654 Explainable Artificial Intelligence
3 credits
- Prerequisite(s): CSCI-B 503 and one of the following: CSCI-B 551, CSCI-B 555, CSCI-P 558, INFO-H 515, or INFO-H 518
- Delivery: On-Campus
Description
This course covers explainable artificial intelligence methodologies and techniques for effective model building. The goal is to leverage explainability design principles to build powerful, complex, and transparent models. Students learn methodologies, parametric models, nonparametric models, deep learning complexity, activation and saliency maps, attention and transformer, compliance and ethics, and applications.
Topics
Introduction of explainable AI
- Fast pace of the field: large-scale models, multimodal models, various industries, etc.
- Pressing need of AI explainability
- Explaining AI learning principles for powerful model building
- Explaining AI learning principles for transparent model building
- Challenges in AI explainability
- Explainability in machine learning model building
- Explainability in deep learning model building
- Co-considering model performance, transparency, compliance and ethics
Explainability in machine learning
- Parametric AI models
- Probabilistic interpretability with Bayes decision theory
- Explainability of decision-making risk
- Non-parametric AI models
- Non-parametric AI model learning principles
- Data representation and features in the learning process
- Feature importance analysis
- Explainability in supervised and unsupervised learning
Explainability in deep learning
- Neural network complexity and explainability
- Interpretating and boosting neural network performance
- Activation map analysis
- Gradient-weighted class activation mapping
- Saliency map analysis
- Intermediate neural-layer interpretations
- Attention and transformer in deep learning and large models
- Complex deep learning model analysis and boosting
Explainable AI for transparency, compliance and ethics
- Enhancing transparency and trust
- Regulation compliance
- Ethical considerations
- Co-considering model performance, transparency, compliance and ethics
Learning principles in AI
- Problem formulation in AI
- Hypothesis space
- VC-dimension
- Linear and nonlinear AI models
- Large AI models
Learning Outcomes
- Analyze and compare major methods for explainability in machine learning.CS 4
- Assess the various explainability methods and techniques.CS 4
- Evaluate deep learning models with explainability design and analysis methods.CS 4
- Analyze and optimize the tradeoffs between model performance, transparency, compliance, and ethics in machine learning systems.CS 4 CS 6
- Design and implement the neural network model with the explainability principles, with a course project.CS 2 CS 6 CS 7
Policies and Procedures
Please be aware of the following linked policies and procedures. Note that in individual courses instructors will have stipulations specific to their course.