Traps to avoid when working with AI
summary
Three concrete traps and what actually fixes them: the accuracy paradox with 99% accuracy on highly imbalanced defect images and augmentation that accidentally teaches the model to spot augmented vs real data — fix by gathering minority-class examples, resampling, changing the metric or penalizing errors. The evaluation function for behavioral tasks breaks when it’s implicit or singular (distance that allows rolling, or a divide-by-zero that makes spiders fly) — specify explicit, robust objectives. For overfitting in simulation, add randomness and test on real robots early.
AI-written summary · about
chapters (10)
- 0:00Intro
- 1:10Fault detection
- 3:08The accuracy paradox and imbalanced classes
- 5:08Data augmentation
- 6:18The choice of the evaluation function Regression
- 6:43The choice of the evaluation function Behavioral task
- 9:29rd try: idem but after the bug correction
- 10:34Rule of thumb: speak to a child
- 11:34Overfitting spotting
- 12:19Teaching a robot how to avoid walls
related talks
49mAn AI with an Agenda
31mAvoid mistakes building AI products
52mSome pitfalls in AI
01h 05mNavigating the Security and Privacy Landscape of Modern AI
56mAn AI with an Agenda: How Our Cognitive Biases Leak Into Machine Learning
31mThe Dark Side of AI
41mWhen AI Failures Reflect Our Own Social Biases
47mFight AI Bias With… Bias