This is a work in progress.

The goal of this project is to build a tool for understanding how neural networks work without having to do any kind of symbol manipulation. I feel that I’m able to much more deeply understand the way something works if I have access to a visual isomorphism that I can play with. Inspired a little bit by Bret Victor.

This project is a direct descendant of a demo in Karpathy’s ConvnetJS library.


  • come up with a good way to visualize the gradient/training steps
  • so much refactor
  • flesh out the explanation/tutorial
    • combine symbolic and graphical representations?
    • talk about biological motivation?
    • talk about philosophical motivation?