After reading several books about deep learning, now I can use keras / tensorflow to train some models, but the mathmatical implementations behind the libraries are still have to follow.

Two years ago, when I played with Karpathy‘s [ConvnetJS], I was shocked by the agrighms behind that. I think now it’s time to go through his code, I started a small project to annotate ConvnetJS source code and re-write it with TypeScript. You can check my progress from ConvnetJS Source Annotation, you can also check the documentation directly.