Deep Learning in JS
summary
Claims that JS can run real deep learning with deeplearn.js / tensorflow.js and that models can stay in-browser for privacy. Demonstrations show WebGL browser inference running at about half C++ speed, Node CPU bindings now and C++ backend promised. Concrete build: MobileNet transfer learning with Adam optimizer and softmax head turned into a webcam Pac‑Man controller using four classes; it works reliably until input distribution shifts. Robots are impressive and also fundamentally limited.
AI-written summary · about
chapters (10)
- 0:00Start
- 8:10Stochastic Gradient Descent
- 12:15Gradient Descent
- 14:00Sigmoid Jewel Functions
- 15:05The Rectified Linear Unit
- 18:51Transfer Learning
- 22:37Soft Max Layer
- 23:35Categorical Cross Entropy
- 25:47Normalization Layers
- 26:16Convolutions
related talks
35mThe Future of JavaScript & Machine Learning
54mDeep Learning in JS
48mDeep Learning in the Browser
39mRun Deep Learning models in the browser with JavaScript and ConvNetJS
15mThe future of machine learning using JavaScript
34mA Powerful Introduction to Machine Learning with Js
34mThe Browser, The Machine Learning and The Future we might not have imagined just less than a year ago
31mDeep Learning at the Edge with TensorFlow.js