Doing BackProp from scratch(calculating derivatives and chain rule) and training it on the mnist dataset.
I successfully created a neural network that trains on the mnist dataset, but I am a terrible C++ coder. Since I am a terrible coder it processes each image in around 0.8 seconds. Although, a very fun project learned about parallel programming and how neural nets work in the background.