JavaScript Machine Learning from Magic to Science
summary
Uses TensorFlow.js and Google’s QuickDraw dataset to build a 10-class “sorting hat” classifier trained from ~10,000 sketches per class reduced to ~8–9k usable images after cleaning, reaches about 90% accuracy, and produces a 400 KB model. Cleaning strategy includes an auxiliary model that flags trash sketches; initial manual cleanup fails. Convolutional feature layers are reused for transfer learning so new categories train from roughly 50 examples and converge very quickly.
AI-written summary · about
chapters (10)
- 0:00Introduction to AI in JavaScript
- 3:59Why AI is everywhere now
- 7:10Building a Harry Potter sorting hat
- 12:29Cleaning and preparing data
- 18:14Introducing the Ridiculous Dataset
- 19:18Training a neural network in JavaScript
- 23:55Convolutions and max pooling explained
- 26:05Live demo of the AI sorting hat
- 28:33Transfer learning with Star Trek logos
- 31:45Summary and next steps
related talks
36mPractical Machine Learning for Front End Developers
34mMachine learning for front-end developers
43mLessons Learned Using Machine Learning in Java
29mMachine learning for front-end developers
41mPractical Machine Learning for Front-End Developers
48mLessons learned using Machine Learning in Java
39mMachine Learning Magic for Your JavaScript Application
34mML from a software engineer's perspective