Out Of Distribution, Out Of Control? Understanding Safety Challenges In AI

Overview

Explore the safety challenges in AI, focusing on out-of-distribution issues and safety guarantees for large language models with Aditi Raghunathan.

Syllabus

• Introduction to AI Safety
Overview of AI safety concerns
Importance of addressing safety in AI systems
• Out-of-Distribution (OOD) Issues
Definition and examples of OOD
Impact of OOD on AI system performance
Strategies for detecting OOD data
• Theoretical Foundations
Statistical and probabilistic foundations of OOD
Robustness in AI models
Evaluation metrics for OOD scenarios
• Large Language Models (LLMs)
Introduction to large language models
Common use cases and applications
Limitations and failure modes
• Safety Guarantees in AI
Definition and examples of safety guarantees
Approaches for ensuring safety in AI models
Verification and validation techniques
• Techniques for Enhancing Safety
Robust training methods
Adversarial training and defenses
Model interpretability and trustworthiness
• Case Studies
Analysis of real-world AI failures
Lessons learned and safety improvements
• Ethical Considerations
Ethical implications of AI safety
Balancing performance with safety
• Practical Workshops
Hands-on exercises with open-source tools
Simulations of OOD scenarios and safety assessments
• Future Directions and Open Research Challenges
Emerging trends in AI safety
Key areas for further research and development
• Course Wrap-up
Review and discussion of key concepts
Final thoughts on the future of AI safety and OOD challenges

---

Subjects

Computer Science