A multitron implementation using numpy that achieves roughly 85% to 88% accuracy.

You will need the littlemnist.pkl file (can be found here) which has a structure such as follows:

[ train, validation, test ]

Where train, validation, test are each defined:

[ data items, classification ]

data items is an array of lenght n where each element is a feature vector of length 28*28 = 784. Likewise, classification, is an array of length n with the proper classification for the feature vector in data items for each i < n.

• High start learning rate with decay related to validation set error rate
• Add bias in a more memory efficient way