Incorporating Domain Knowledge in VAE Learning via Exponential Dissimilarity-Dispersion Family

Ren Togo, Nao Nakagawa, Takahiro Ogawa, Miki Haseyama

Primary Area: representation learning for computer vision, audio, language, and other modalities

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.

Keywords: variational autoencoders, deep generative models, similarity indices, domain knowledge, dispersion parameter

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: The paper introduces a new distribution family called EDDF to improve the performance of VAEs in generative modeling for high-dimensional natural data, and validates its effectiveness in the vision domain.

Abstract: Variational autoencoder (VAE) is a prominent generative model that has been actively applied to various unsupervised learning tasks such as representation learning. Despite its representational capability, VAEs with the commonly adopted Gaussian settings typically suffer from performance degradation in generative modeling for high-dimensional natural data, which is often caused by their excessively limited model family. In this paper, we introduce the exponential dissimilarity-dispersion family (EDDF), a novel distribution family that includes a dissimilarity function and a global dispersion parameter. A decoder with this distribution family induces arbitrary dissimilarity functions as the reconstruction loss of the evidence lower bound (ELBO) objective, where the model leverages domain knowledge through this dissimilarity function. For VAEs with EDDF decoders, we also propose an ELBO optimization method that implicitly approximates the stochastic gradient of the normalizing constant using log-expected dissimilarity. Empirical evaluations of the generative performance show the effectiveness of our model family and proposed method in the vision domain, indicating that the effect of dissimilarity determines the criteria of representational informativeness.

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.

Submission Number: 890