Schedule
Legend: 🎥 lecture recording, 🖥️ slides, 📖 notes, 📝 written questions, ⌨️ coding assignment.
Apply to participate in the course program by Jan 29th to have your assignments graded and participate in discussions and speaker events.
Background
Safety Engineering
- 3
- Risk Decomposition
- 🎥, 🖥️️
- risk analysis definitions, disaster risk equation, decomposition of safety areas, ability to cope and existential risk
- 🎥, 🖥️️
- 4
- Accident Models
- 🎥, 🖥️
- FMEA, Bow Tie model, Swiss Cheese model, defense in depth, preventative and protective measures, complex systems, nonlinear causality, emergence, STAMP
- 🎥, 🖥️
- 5
- Black Swans
- 🎥, 🖥️
- unknown unknowns, long tailed distributions, multiplicative processes, extremistan
- 🎥, 🖥️
- ►
- Review questions 📝
Robustness
- 6
- Adversarial Robustness
- 🎥, 🖥️, 📖, ⌨️
- optimization pressure, PGD, untargeted vs targeted attacks, adversarial evaluation, white box vs black box, transferability, unforeseen attacks, text attacks, robustness certificates
- 🎥, 🖥️, 📖, ⌨️
- 7
- Black Swan Robustness
- 🎥, 🖥️️, 📖
- stress tests, train-test mismatch, adversarial distribution shifts, simulated scenarios for robustness
- 🎥, 🖥️️, 📖
- 8
- Review questions 📝
Monitoring
- 8
- Anomaly Detection
- 🎥, 🖥️️, 📖, ⌨️
- AUROC/AUPR/FPR95, likelihoods and detection, MSP baseline, OE, ViM, anomaly datasets, one-class learning, detecting adversaries, error detection
- 🎥, 🖥️️, 📖, ⌨️
- 9
- Interpretable Uncertainty
- 🎥, 🖥️, 📖
- calibration vs sharpness, proper scoring rules, Brier score, RMS calibration error, reliability diagrams, confidence intervals, quantile prediction
- 🎥, 🖥️, 📖
- 10
- Transparency
- 🎥, 🖥️
- saliency maps, token heatmaps, feature visualizations, ProtoPNet
- 🎥, 🖥️
- 11
- Trojans
- 🎥, 🖥️, 📖, ⌨️
- hidden functionality from poisoning, treacherous turns
- 🎥, 🖥️, 📖, ⌨️
- 12
- Detecting Emergent Behavior
- 🎥, 🖥️, 📖
- emergent capabilities, instrumental convergence, Goodhart’s law, proxy gaming
- 🎥, 🖥️, 📖
- 13
- Review questions 📝
Alignment
- 13
- Honest Models
- 🎥, 🖥️
- truthful vs. honest, inverse scaling, instances of model dishonesty
- 🎥, 🖥️
- 14
- Power Aversion
- 🖥️
- TBC early 2023; social, economic, and governmental formalizations of power bases; power penalties
- 🖥️
- 15
- Machine Ethics
- 🎥, 🖥️, ⌨️
- normative ethics background, human values, value learning with comparisons, translating moral knowledge into action, moral parliament, value clarification
- 🎥, 🖥️, ⌨️
Systemic Safety
- 16
- ML for Improved Decision-Making
- 🎥, 🖥️, 📖
- forecasting, brainstorming
- 🎥, 🖥️, 📖
- 17
- ML for Cyberdefense
- 🎥, 🖥️
- intrusion detection, detecting malicious programs, automated patching, fuzzing
- 🎥, 🖥️
- 18
- Cooperative AI
- 🎥, 🖥️, 📖
- nash equilibria, dominant strategies, stag hunt, Pareto improvements, cooperation mechanisms, morality as cooperation, cooperative dispositions, collusion externalities
- 🎥, 🖥️, 📖
Additional Existential Risk Discussion
- 19
- X-Risk Overview
- 🎥, 🖥️
- arguments for x-risk
- 🎥, 🖥️
- 20
- Possible Existential Hazards
- 🎥, 🖥️
- weaponization, proxy gaming, treacherous turn, deceptive alignment, value lock-in, persuasive AI
- 🎥, 🖥️
- 21
- Safety-Capabilities Balance
- 🎥, 🖥️
- theories of impact, differential technological progress, capabilities externalities
- 🎥, 🖥️
- 22
- Natural Selection Favors AIs over Humans
- 🖥️
- TBC early 2023; Lewontin’s conditions, multiple agents and variation, generalized Darwinism, evolutionary mechanisms for altruism, internal constraints, incentives, Leviathan
- 🖥️
- 23
- Review and Conclusion
- 🎥, 🖥️, 📝
- pillars of ML safety research, task-train-deploy pipeline
- 🎥, 🖥️, 📝