Machine Learning Sucks
summary
Machine learning needs lots of labeled and cleaned data so projects spend ~75% of time on data cleaning, and labels are costly; Facebook gets ~97% face recognition accuracy by crowd tagging while older systems hit ~85%. Anomaly detection uses next-day prediction with a tolerance band on six-day averages. Overfitting appears when ten noisy points are fit to many parameters — 100% fit that misses the underlying sine wave. Self-driving fleets cut transport cost from €650 to <€100/month and 12 planting robots replace 24 workers.
AI-written summary · about
chapters (14)
- 0:00Intro
- 0:21Machine Learning Sucks
- 4:20Data
- 10:43Over-fitting
- 15:14Extrapolation
- 16:13Features
- 18:09Adverserial Examples
- 19:16Technology
- 20:36Job market
- 26:36Democracy
- 27:02Start-ups
- 28:25Business
- 34:25Conclusion
- 35:12healthskouts
related talks
25mAutomating Operations with ML
54mMachine Learning explained with Examples
30mActive Learning and Human-in-the-Loop Machine Learning
15mMachine learning: Necessity vs obligation
45mApplied machine learning: a few lessons I learned the hard way
33mThe Role of Software Developers in Machine Learning
30mMachine Learning: Beyond the Tutorial
37mMachine Learning, The Right Way