usholanb

#############################################################################
#  Abstract: This code trains a model to predict next character based on
#  the previous ones. 
#
#  More details:
#  1. Each character is represented as a vector of size 256. It is all 
#     zeros except the index where that characters stands in ASCII table.
#  2. Code divides the text on chunks of input_size. 
#  3. The teaching labels of all the characters in a chunk are the next
#     characters, so we just shift the chuck to the right and assign the

usholanb

#####################################################################################
#  This code trains a model that predicts which number from 0 to 9 is drawn on the
#  MNIST pictures of 20x20 pixels. There are 60000 training examples and 10000
#  testing examples.
#  The methods used: Mini-batching, Weight Decay, Momentum, Dropout,
#                    Xavier's Initialization
#  Number of hidden layers: 3 ( Adjustable )
#  Sizes of mini-batches and hidden layer are easily adjustable
#  Non-Linear functions: Sigmoid, Funny Tanh
#  Linear: ReLU