See visualization and project details here
This project started off as an attempt to run an MNIST classification neural network in the browser with a real time visualization. See this repo for why we changed our mind.
Instead, we ran the neural nets in advance and saved snapshots of them through the learning process. This turned out to provide exactly the type of experience we were looking for without having to push the browser so hard.
This project was a great way to learn the details of simple feedforward neural networks. It was also a nice refresher for using D3.js.
There are more statistics that could be visualized such as the change in cost, fine grained analysis of the change in weights, or gradients analysis. Relu, other activation functions, and softmax are common techniques that could be added. The neural networks could be run in real time using a library such as Tensor Flow on the backend.
Convolutional neural networks also seem like a cool next thing to visualize... Maybe Three.js??
[Simple Redux Boilerplate] (https://github.com/tsaiDavid/simple-redux-boilerplate) was used to start off the app. Credit of course goes to the authors of React, Redux, D3.js and everything else in the package.json. We also highly suggest Michael Nielsen's excellent book.