Steps Toward Robust Artificial Intelligence - Thomas G Dietterich, Oregon State University

Explore the critical steps towards developing robust artificial intelligence in this comprehensive lecture by Professor Thomas G. Dietterich of Oregon State University. Delve into the challenges of integrating AI technologies into high-stakes applications such as self-driving cars, robotic surgeons, and weapons systems. Examine methods for addressing known and unknown threats, including probabilistic inference, robust optimization, anomaly detection, and causal modeling. Learn about recent research on open category classification and probabilistic guarantees. Gain insights into robustness lessons from biology, decision-making under uncertainty, and the importance of employing algorithm portfolios and ensembles. Discover how to detect surprises, monitor auxiliary regularities, and use larger models to reduce the risk of unknown unknowns in AI systems.

Intro
STEPS TOWARD ROBUST ARTIFICIAL INTELLIGENCE
Marvin Minsky (1927-2016)
Minsky: Difference between Computer Programs and People
Outline
Self-Driving Cars
Automated Surgical Assistants
Autonomous Weapons
Conclusion
Robustness Lessons from Biology
Decision Making under Uncertainty
Robustness to Downside Risk
Robust Optimization • Many Al reasoning problems can be formulated as optimization problems
Impose a Budget on the Adversary
Detect Surprises
Monitor Auxiliary Regularities
Monitor Auxiliary Tasks
Open Category Object Recognition
Prediction with Anomaly Detection
Theoretical Guarantee
Related Efforts
Use a Bigger Model The risk of Unknown Unknowns may be reduced if we model more aspects of the world • Knowledge Base Construction Information Extraction & Knowledge Base Population
Use Causal Models
Employ a Portfolio of Models
Portfolio Methods in SAT & CSP
Summary