Tuesday, May 01, 2018

TensorFlow.js: ML in web browsers


"A WebGL accelerated, browser based JavaScript library for training and deploying ML models."
  • Develop ML in the Browser
  • Run Existing models
  • Retrain Existing models

A very nice animation of neural network "learning" process.

A Neural Network Playground

References to free online bools about Machine Learning:

podcast interview with Director of Engineering in Google working on Tensorflow:

TensorFlow - YouTube - YouTube

Teachable Machine

Porting of a complex application to web browser could be done by

  1. rewriting it in JavaScript, apparently what Google has done, or
  2. cross-compiling C++ binary to WebAssembly 
Here is an attempt of the second approach:

tomasreimers/tensorflow-emscripten: @ GitHub

MXnet, another major machine learning library, used and supported by Amazon
is available cross-compiled:

dmlc/mxnet.js: MXNetJS: Javascript Package for Deep Learning in Browser (without server) @ GitHub

A predictable future: Microsoft's CNTK, a ML library similar to Tensorflow and MXnet,
will be cross-compiled to WebAssembly soon :) If whole .NET can run in the browser, so can ML tools.

Microsoft/CNTK: Microsoft Cognitive Toolkit (CNTK), an open source deep-learning toolkit @ GitHub

No comments: