Skip to content

Latest commit

 

History

History
27 lines (18 loc) · 1.74 KB

README.md

File metadata and controls

27 lines (18 loc) · 1.74 KB

AE_Vis

The repository contains an autoencoder model implementation in Keras, which is trained on MNIST dataset of handwritten digits.

Animations are created to demonstrate the interpolation property of the latent space, i.e. linear interpolation in latent space results in smooth changes in image space. Some of the outstanding examples are:

Also, several dimensionaluty reduction techniques, namely PCA, LDA, Isomap and tSNE, are used to transform the latent representation (32D) of images into 2D embedding.

alt text

alt text

alt text

alt text

It can be observed that out of the above methods tSNE achieves relatively better results in finding 2D embedding that exhibits quite good separation property between the 10 classes.