AI in JavaScript

---

TensorFlow

• TensorFlow
• tensorflow.js (and github) (formerly deeplearn.js)
• MobileNet (see github) (Convolutional Neural Networks for lightweight machine vision) using tensorflow.js
• MobileNet is used in ML5.

---

ML5

• Inspired by P5. Built on tensorflow.js. Includes tensorflow utilities.
• ml5js.org
• docs.ml5js.org

• github
• npmjs

• releases
• ml5.min.js (v0.6.1)
• ml5.min.js.map (v0.6.1)

• ML5 examples on P5 web editor

How ML5 loads pre-trained AI models

• ML5 can get a pre-trained AI model at run-time.
• Example in the ML5 World: "Recognise any image".
• See code:

   classifier = ml5.imageClassifier ( 'MobileNet' .. 

• ml5.imageClassifier docs
• The argument is one of pre-defined models or else URL of some model.json.

• How to see the download of the AI model:
  • In Chrome dev tools: Network - Reload. You see model.json with lots of extra arguments. "Open in new tab".
  • This is the "model manifest", defining the structure of the neural network, and list of the files of weights:

     ["group1-shard1of4.bin", "group1-shard2of4.bin", "group1-shard3of4.bin", "group1-shard4of4.bin"] 

  • We can also see these shards in the "Network" list. They are downloaded (over 10 M). "Copy URL" or "Open in new tab".
  • If they are coming from cache, it does not show filesize. Click "Disable cache" checkbox and run again.
  • The shards combine to form the working neural network which is therefore downloaded and runs locally.

---

ConvNetJS

• Andrej Karpathy
  • site
  • ConvNetJS (Convolutional Neural Networks library) (see github)
  • ConvNetJs2
  • REINFORCEjs

---

Other

• Awesome Machine Learning list of libraries by language
  • JavaScript

• Other:
  • WebCNN (see demo)
  • synaptic.js
  • Keras.js
  • Limdu