Proceedings of the Workshop on Visualization for AI Explainability (VISxAI),

Towards an Interpretable Latent Space

T. Spinner, J. Körner, J. Görtler, O. Deussen

Material

Abstract

We present an intuitive comparison of Auto-Encoders (AE) with Variational Auto-Encoders (VAE) by visualizing their latent activations. In order to do this, we trained an AE and the corresponding VAE on the MNIST dataset. To give a feeling for the latent compression, we visualize the latent activations of the AE/VAE by displaying the 4 latent variables in a parallel coordinate system. We provide an introduction to the architectures of AEs/VAEs and draw a comparison between the two models.

BibTeX

@inproceedings{Spinner2018TowardsInterpretableLatent,
  author    = {T. Spinner, J. Körner, J. Görtler, O. Deussen},
  booktitle = {Proceedings of the Workshop on Visualization for AI Explainability (VISxAI)},
  comment   = {noDOI},
  title     = {Towards an Interpretable Latent Space},
  url       = {https://spinthil.github.io/towards-an-interpretable-latent-space/},
  year      = {2018}
}