Local Explanations for Deep Learning Models - Fall 2023

Learn to understand and implement local explanation methods for deep learning models through this comprehensive university course covering gradient-based attribution, uncertainty estimation, prompting techniques, and human-AI trust evaluation. Explore data preparation and transformer architectures before diving into pretrain-finetune methodologies and uncertainty estimation techniques across two detailed sessions. Master prompting strategies including chain-of-thoughts approaches and learn to evaluate free-text explanations effectively. Develop skills in gradient-based input attribution methods and their proper evaluation, while understanding select-then-predict frameworks and pairwise feature importance through effective attention mechanisms. Investigate data influence techniques and contrastive explanation approaches to better interpret model decisions. Examine application-grounded evaluation methods for explanations and analyze the critical role of human trust in AI systems. Address the challenges involved in fostering appropriate trust and distrust in AI technologies, while exploring explainability as an interactive dialogue between humans and machines. Conclude with a comprehensive overview of local explainability methods and their evaluation frameworks, providing a complete foundation for implementing interpretable deep learning solutions in real-world applications.