Tom Dietterich President, Association for the Advancement of Artificial Intelligence
Marvin Minsky (1927-2016)
Minsky: Difference between Computer Programs and People
almost any error will completely paralyze a typical computer program, whereas a person whose brain has failed at some attempt will find some other way to proceed. We rarely depend upon any one method. We usually know several different ways to do something, so that if one of them fails, there's always another.
Outline
The Need for Robust AI
Approaches toward Robust AI
Concluding Remarks
- High Stakes Applications
- Need to Act in the face of Unknown Unknowns
Approaches toward Robust AI
- Robustness to Known Unknowns
- Robustness to Unknown Unknowns
Concluding Remarks
Exciting Progress in AI: Perception
Image Captioning
Perception + Translation
Skype Translator: Speech Recognition + Translation
Exciting Progress in AI: Reasoning (SAT)
Exciting Progress: Reasoning (Heads-Up Limit Hold’Em Poker)
Exciting Progress: Chess and Go
Personal Assistants
Technical Progress is Encouraging the Development of High-Stakes Applications
Self-Driving Cars
Automated Surgical Assistants
AI Hedge Funds
AI Control of the Power Grid
Autonomous Weapons
High-Stakes Applications Require Robust AI
Why Unmodeled Phenoma?
It is impossible to model everything
It is important to not model everything
Conclusion:An AI system must act without having a complete model of the world
Digression: Uncertainty in AI
Known Unknowns
Unknown Unknowns
Outline
Robustness Lessons from Biology
Approaches to Robust AI
Idea 1: Robust Optimization
Uncertainty in the constraints
Minimax against uncertainty
Impose a Budget on the Adversary
Idea 2: Regularize the Model Regularization in ML:
Regularization can be Equivalent to Robust Optimization
Idea 3: Optimize a Risk-Sensitive Objective
Idea 3: Optimize Conditional Value at Risk
Optimizing CVaR confers robustness
Idea 4: Robust Inference
Approaches to Robust AI
Idea 5: Expand the Model
Idea 5: Expand the Model
Idea 6: Use Causal Models
Idea 7: Employ a Portfolio of Models
Portfolio Methods in SAT & CSP
SATzilla Results
Parallel Portfolios
IBM Watson / DeepQA
Knowledge-Level Redundancy
Multifaceted Understanding
Achieving Multi-Faceted Understanding
Idea 8: Watch for Anomalies
Automated Counting of Freshwater Macroinvertebrates
Open Category Object Recognition
Prediction with Anomaly Detection
Novel Class Detection via Anomaly Detection
Anomaly Detection Notes
Related Efforts
Open Questions
Concluding Remarks
