Link to the presentation

Optimal transport maps for distribution preserving operations on latent spaces of Generative Models

Reproducing experiments

• Prepare GAN models(Faces, MNIST, Rooms) for different generator priors(Normal, Uniform)
• Implement interpolation functions from the paper:
  • 2-point interpolation
  • n-point interpolation
  • vicinity sampling
  • analogies
• Conduct experiments and compare results

Experiments with VAE

• Prepare VAE models(Faces, MNIST, Rooms)
• Implement functions from the paper:
  • 2-point interpolation
  • n-point interpolation
  • vicinity sampling
  • analogies
• Implement new interpolation for latent space of VAE
• Conduct experiments and compare results

Working with missing values

The idea is to check whether it’s possible to combine 2 pictures with missing parts into 1 good image by first mapping into latent space, performing interpolation and then mapping back using decoder(Using different interpolation techniques). So, for this we will need encoder-decoder architecture. We are planning to use VAE for the moment.

Additional fun stuff(if we have time)

Draw 2 dimensional map of the dataset using vicinity sampling