Will Machine Learning kill Business Logic?
summary
A small insurance demo app takes speed age and miles and outputs risk using models trained on ~1,500 labeled records, showing that hand written rules score ~43% while a scikit‑learn decision tree reaches ~70% and a tiny TensorFlow neural net about ~80% on held‑out data. Code runs in Google Colab on a K80 GPU and can export to TensorFlow.js for the browser. Key failure modes are empty feature regions and overconfidence so add uncertainty estimates, holdout validation and plan for ~5,000+ samples for reliable results.
AI-written summary · about
chapters (27)
- 0:00Intro
- 1:15Scenario
- 2:07Building the system
- 3:35Requirements
- 4:57Rules
- 6:51Better than guessing
- 7:50Machine Learning
- 9:28Visual Experiment
- 15:15Decision Boundaries
- 19:46Decision Trees
- 21:57Code
- 25:58Results
- 27:10Good generalization
- 29:09Overfitting
- 29:51Reduce Depth
- 30:30Model Results
- 31:15Model Code
- 34:12What to do
- 35:48What works
- 38:00German Autobahn
- 39:19Supervised ML
- 39:40Other types
- 40:17What doesnt work
- 42:13Machine Learning vs Software Engineering
- 43:33Uncertainty
- 45:37Friedman
- 46:11Conclusion
related talks
35mBreaking black-box AI
01h 39mApplying Rigorous Machine Learning Methods in Business Strategy
30mMachine Learning: Beyond the Tutorial
54mMachine Learning explained with Examples
50mBuilding Intelligent Mobile Apps using Machine Learning
56mData, AI, Hype and Fear - Recipe for Disaster?
01h 05mMachine Learning Models and Productionization
52mBreaking black-box AI