Interventional Assays for the Latent Space of Autoencoders

The encoders and decoders of autoencoders effectively project the input onto learned manifolds in the latent space and data space respectively. We propose a framework, called latent responses, for probing the learned data manifold using interventions in the latent space. Using this framework, we investigate "holes" in the representation to quantitatively ascertain to what extent the latent space of a trained VAE is consistent with the chosen prior. Furthermore, we use the identified structure to improve interpolation between latent vectors. We evaluate how our analyses improve the quality of the generated samples using the VAE on a variety of benchmark datasets.