We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

stat.ML

Change to browse by:

References & Citations

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Statistics > Machine Learning

Title: Quantitative Understanding of VAE as a Non-linearly Scaled Isometric Embedding

Abstract: Variational autoencoder (VAE) estimates the posterior parameters (mean and variance) of latent variables corresponding to each input data. While it is used for many tasks, the transparency of the model is still an underlying issue. This paper provides a quantitative understanding of VAE property through the differential geometric and information-theoretic interpretations of VAE. According to the Rate-distortion theory, the optimal transform coding is achieved by using an orthonormal transform with PCA basis where the transform space is isometric to the input. Considering the analogy of transform coding to VAE, we clarify theoretically and experimentally that VAE can be mapped to an implicit isometric embedding with a scale factor derived from the posterior parameter. As a result, we can estimate the data probabilities in the input space from the prior, loss metrics, and corresponding posterior parameters, and further, the quantitative importance of each latent variable can be evaluated like the eigenvalue of PCA.
Comments: 40 pages, 29 figures
Subjects: Machine Learning (stat.ML); Information Theory (cs.IT); Machine Learning (cs.LG)
ACM classes: I.2.4
Cite as: arXiv:2007.15190 [stat.ML]
  (or arXiv:2007.15190v3 [stat.ML] for this version)

Submission history

From: Akira Nakagawa [view email]
[v1] Thu, 30 Jul 2020 02:37:46 GMT (14861kb,D)
[v2] Fri, 23 Oct 2020 07:53:54 GMT (15445kb,D)
[v3] Sat, 12 Jun 2021 04:51:47 GMT (18935kb,D)

Link back to: arXiv, form interface, contact.