In today's world of heavy duty machine learning frameworks and blackbox models lying around every corner, I have found it easy to go a long time without truly understanding the math that makes all these algorithms work. However, learning what goes on under the hood can be extremely useful when designing application models. So I decided to take a closer look at deep learning models and implement some of them from scratch.

This is an efficient implementation of a Convolutional Neural Network (CNN) including both back and forward propagation, written using only NumPy. This demo uses a model trained on the EMNIST dataset, and currently supports digits 0-9. More characters are coming soon!