Standard ANN implementation using BackProp/ForwardProp with 1 hidden layer, current training setting has 300 hidden neurons. Classifier has >85% accuracy on all 2000 test examples with current training at 4000 examples.
Refer to bottom of ann.py on how to change initial parameters (regularization constant LAMBDA, neurons in hidden layer, etc.).
Total training examples available: 6000 Total test examples available: 2000
WARNING: Training takes a while and a good CPU is recommended for >1000 training examples with current hidden neurons. If you want to use the pre_trained data, change LOAD_FROM_FILE to False at the bottom of ann.py