An introduction to Machine Learning using LEGO
summary
A small LEGO Mindstorms car uses a smartphone camera and a custom app to collect training pairs (a single row of pixels as features and a steering angle label 0–100%), then fits a linear model with feature weights via gradient descent on-device. About 20 minutes of driving data produce a workable model after testing squared features and simple preprocessing; LEGO Mindstorms hardware limits steering responsiveness, and experiments show a learning rate of 1 converges in 500 iterations, 0.001 is slow, 2.5 diverges and 100 overflows.
AI-written summary · about
chapters (13)
- 0:00Start
- 2:41SELF-DRIVING LEGO
- 11:52HYPOTHESIS
- 15:17MINIMIZING COST FUNCTION
- 17:14BACK TO LEGO
- 17:47COLLECTING DATA
- 20:27FEATURES
- 28:06GRADIENT DESCENT
- 38:00SOLUTION - BAD MODEL
- 41:40SOLUTION - BAD HARDWARE
- 49:39WHAT HAVE WE LEARNED?
- 51:29IMPROVEMENTS
- 54:21MORE REALISTIC STEERING
related talks
58mAn introduction to Machine Learning using LEGO
43mLessons Learned Using Machine Learning in Java
19mDeep Learning is Just LEGO: & Other Hands on Machine Learning Activities
48mLessons learned using Machine Learning in Java
01h 25mHow to Build Your Own Self Driving Toy Car
58mThe Modern Trolley Problem - Responsible AI Principles
51mMachine Learning Exposed Introduction to Machine Learning
42mThe Lego Model for Machine Learning