Neural network 3D mesh regression
This app uses the convnet.js library to learn a Three.js sphere geometry in the browser. The network is heavily inspired by this example and uses a web worker to process the neural network training whilst periodically passing vertex position predictions back to the Three.js UI for rendering. This keeps the framerate of main WebGL render high and keeps the interactions responsive.
A sample of weights from the neural network are shown at the top of the browser window and are updated as the neural network learns the sphere vertex positions.