ProtoVAE: Using Prototypical Networks for Unsupervised Disentanglement

Vaishnavi S Patil, Matthew Evanusa, Joseph JaJa

Keywords: Unsupervised Learning, Disentangled Representations

TL;DR: Unsupervised Disentangled representation learning using Isometric inductive biases

Abstract: Generative modeling and self-supervised learning have in recent years made great strides towards learning from data in a completely \emph{unsupervised} way. There is still, however, an open area of investigation into guiding the neural network to learn useful or good representations. The problem of unsupervised \textit{Disentanglement} is of particular importance as it offers to learn interpretable representations, with disjoint subsets of the representation encoding different, meaningful factors of variation. Recent work has theoretically grounded the factors of variation, via the lens of group theory, as disentangled actions of the symmetry subgroups which transform only the correspond subspaces of the disentangled representation. We use this mathematical formalism instead to impose constraints on the representations learned by a unsupervised generative neural network, such that transformations of the representation correspond to the actions of a unique symmetry subgroup. To this end, we introduce a novel model, ProtoVAE, that leverages a deep metric learning Prototypical network trained via self-supervision to constrain the latent space of a Variational Autoencoder to decompose into independent subspaces. Further, we actively change or \textit{intervene} in the latent space during training to enforce each dimension of the representation to uniquely and consistently transform the data corresponding to some symmetry subgroup. We demonstrate and evaluate our proposed model on the benchmark DSprites and 3DShapes datasets and compare with other state of the art disentanglement methods via qualitative traversals in the latent space, as well as quantitative disentanglement metrics. We further qualitatively demonstrate the effectiveness of our model on the real-world datasets CelebA which consistently encodes the different factors.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Unsupervised and Self-supervised learning