← Projects · research

Variational autoencoder for SVHN

A VAE for the Street View House Numbers dataset, with an added clustering loss to sharpen latent separation.

A variational autoencoder trained on the Street View House Numbers (SVHN) dataset to learn a latent representation of digit images. The standard VAE objective was extended with a clustering term that pulls latent codes toward class-consistent regions. The effect is cleaner separation between digit classes and a latent space that is easier to interpret.