Debugging AI
summary
Claims that debugging ML is unlike traditional software: models hide failures, non‑stationary retraining multiplies failure states, and you cannot unit‑test internal learned weights. Demonstrations include AlphaGo = search + reinforcement learning + neural net, adversarial 3D‑printed turtle -> rifle that fools image classifiers, and word‑embedding gender bias (doctor -> nurse). Practical fixes are data‑quality metrics, injected synthetic cases, confusion matrices and visualizing agent internals, white‑box explanation layers, safety nets and ensemble voting when failure is acceptable.
AI-written summary · about
chapters (10)
- 0:00Start
- 13:10Playing Chess, playing Go
- 17:42Recap: Properties of Al problems
- 18:14Problems with unknown truth
- 19:33ML replicates bias in the data
- 21:21Al systems break fundamental patterns web developed as an industry
- 23:44Alls prone to hacking. So know your tools!
- 25:20Quality management on input data
- 28:22Work with test sets
- 31:08Add a safety net
related talks
51mDebugging AI
01h 05mNavigating the Security and Privacy Landscape of Modern AI
35mHacking AI Systems: How to (Still) Trick Artificial Intelligence
42mReasonable AI
46mA Short Introduction to Hacking Artificial Intelligence
32mQuality engineering in the AI age: From testing AI to leveraging GenAI
01h 01mIntroduction to Practical AI
49mAn AI with an Agenda