Run Deep Learning models in the browser with JavaScript and ConvNetJS
summary
Browser-based inference and even training run end-to-end using JavaScript: models exported from Caffe/Keras load into ConvNetJS-style runtimes, use typed arrays, WebGL shaders and webworkers for GPU-accelerated tensors, and handle backpropagation for demos. Large models (AlexNet ~240 MB) are impractical but modern nets get small (GoogleNet ~7M params, 5 MB SqueezeNet). Working demos include webcam-based on-device training (teachable-machine style), deep dream/style transfer, GANs and interpolating autoencoders; core libraries shown include ConvNetJS and deeplearn.js.
AI-written summary · about
chapters (19)
- 0:00Introduction
- 2:20Why run deep learning models in JavaScript
- 7:00Introduction to neural nets
- 13:09Layers of neural nets
- 13:46Playground
- 14:34Ingredients to Deep Learning
- 16:56Example
- 18:07Models getting deeper
- 20:23Deep Dream demo
- 21:41Deep Dream in the browser
- 22:37ConvNet Models
- 25:26JavaScript
- 26:54Challenges
- 27:34Models available today
- 29:53Demo
- 33:22Other models
- 36:27Additional resources
- 36:45Takehome messages
- 37:44Wrap up
related talks
32mDeep Learning in JS
48mDeep Learning in the Browser
54mDeep Learning in JS
34mA Powerful Introduction to Machine Learning with Js
35mThe Future of JavaScript & Machine Learning
48mTensorFlow.js
26mMachine Learning in the Browser with Deep Neural Networks
20mTensorFlow.js Bringing Machine Learning to the Web and Beyond