YES!! ML in JavaScript! But why?
summary
A demonstration of training and running a face-similarity pipeline entirely in-browser using TensorFlow.js and the BlazeFace detector. It explains training a small Siamese network with contrastive loss on user-collected Toloka data (training ~10–20 minutes), face alignment, using Conv/SeparableConv2D layers, model quantization (down to 4-bit) to reduce size, WebGL vs WebGPU tradeoffs, async chunked download for large models, and practical notes like disposing tensors; first-run inference ~157 ms and ~3–4 ms after warmup.
AI-written summary · about
chapters (9)
- 0:00Introduction
- 8:07Why ML in JavaScript
- 13:25TensorFlow.js basics
- 20:12Face recognition from scratch
- 28:46Building similarity models
- 32:40Other libraries and examples
- 41:09WebGPU and performance
- 47:10Summary and inspiration
- 47:53Q&A
related talks
41mTensorFlow.js 101: ML in the Browser and Beyond
15mThe future of machine learning using JavaScript
19mML at Client Side
15mMachine learning for next gen web apps with TensorFlow.js
46mArtificial Intelligence with JavaScript
39mMachine Learning Magic for Your JavaScript Application
22mVue + TensorFlow.js: Building Your First ML Powered App
54mDeep Learning in JS