Assignment 1

 Question 3 Deep Learning 

Deep Belief Network

Observations:
Best reconstruction occurs at hidden unit= 250, batch size = 200, epoch=30.
Error comes down as the number of rounds of training is increased. Increase in batchsize causes increase
in error for lower epochs. Interestingly, this trend gets reversed for higher epochs. Hidden unit increase
decreases the erorr, for which again from 250 to 500 the trend is reversed.

SAE (stacked auto encoder)

Observations:
Best reconstruction occurs at learning rate = 2, epoch=80 and batch size=100.
Increasing the batchsize increase the size of the error and increasing the number of rounds of training
reduces the error. Increasing the learning rate reduces the error but increases time taken.

CNN (convoluted neural networks)

Observations:
Best reconstruction occurs at batch size=10 (extrapolated), epoch=20.
Increasing the number of rounds of training reduces the error but increasing the batchsize increases the error.

NN

Observations:
The Best Reconstruction occurs at learning rate=2, batch size=100 and epoch=50.
Higher batchsize causes higher error. Increasing the number of rounds of training reduces error.
Increasing the learning rate decreases error while increasing time taken.