Kaggle digits preprocessing

Some trics for kaggle mnist data recognition competition

Dependencies:

• python3
• tensorflow
• keras
• marplotlib
• scipy

Architechture

Input: (60000, 784, 1) (60000 images shape = 28x28)

Hidden: Iteratively changing hidden layer size

Output: One hot vector with 10 labels (0..9)

Idea:

Rotate digit to vertical and stretch

For human it makes digits less recognisible, but for NN digits start to look more similar

Result:

~5% improovement compare to same architechture, without preprocessing

!important: no improovement in case of usage CNN