Sparse_Autoencoder_tf

Comparing performance of two NN models :-

• 1)Sparse autoencoder trained weights, fc model
• 2)Random initialized weights, fc model.

Dataset used:- MNIST

1)Sparse auto encoder trained weights:

• Sparse auto encoder is trained whose encoding layer acts as the 1st layer of the NN
• Then a fully connected NN is trained with the 1st layer as the auto encoder weights
• The auto encoder network architecture is [784,200,784]
• Fully connected NN architecture is [784,200,10]

2) Random initialized weights, fc model

• He weight initialisation is used.
• Network architecture of is this fully connected model is [784,200,10]

Using these 2 networks, classification accuracy for the unlabeled data was calculated and 1st model i.e. the model with Sparse auto encoder performed better.

Code run

• Clone project
• run 'download_mnist.sh'
• run 'sparse_ae.py'