The Evolution Revolution
summary
Argues that machine learning ops should implement once by writing core math in one language and transpiling to GPUs and runtimes, showing working JavaScript examples using brain.js and gpu.js (colorblind filter, green‑screen compositing, Mandelbrot, convolution/backprop kernels, XOR feedforward net running on GPU). Claims brain.js once beat Dialogflow/LUIS/Watson on an NLP test and notes dual‑optimizer backprop; shows a brewing classifier prototype (Welton) and reports that GPU-backed feedforward training and inference actually run and simplify engineering when the math stays exposed — embrace the math.
AI-written summary · about
chapters (11)
- 0:00Introduction to innovation
- 5:43Divergence in machine learning
- 7:28Implement once revolution
- 9:13Rebooting hardware thinking
- 11:00Direct approach benefits
- 13:08Canned vs direct methods
- 15:36Practical code examples
- 17:00End-to-end brain.js demo
- 18:46Evolution of ideas
- 20:13Embrace the math
- 20:57Q&A session
related talks
15mThe future of machine learning using JavaScript
34mThe Browser, The Machine Learning and The Future we might not have imagined just less than a year ago
35mThe Future of JavaScript & Machine Learning
22mKeynote talk - The future of machine learning & JS
24mThe Future of Machine Learning & Javascript
32mDeep Learning in JS
18mMachine Learning with Javascript
41mEscaping the Cloud: High-Performance AI in your Browser