Different approaches to the MNIST dataset

Please download the train.csv and test.csv file from Kaggle Digit Recognizer Competition: https://www.kaggle.com/c/digit-recognizer/data Place the two files in the MNIST directory.

Custon Neural Network consists of a handmade backpropagation algorithm tested on the MNIST dataset. The dataset used is a subset of the original dataset dowloaded from the sklearn.datasets module.

Keras_MNIST.py is simple Fully-Connencted NN approach to the problem. This approach fetches an accuracy of 92% which is way down the State of the art results, which happen to be 100%