Bonheme, Lisa and Grzes, Marek (2022) How do Variational Autoencoders Learn? Insights from Representational Similarity. [Preprint] (doi:10.48550/arXiv.2205.08399) (The full text of this publication is not currently available from this repository. You may be able to access a copy if URLs are provided) (KAR id:98783)
The full text of this publication is not currently available from this repository. You may be able to access a copy if URLs are provided. | |
Official URL: https://doi.org/10.48550/arXiv.2205.08399 |
Abstract
The ability of Variational Autoencoders (VAEs) to learn disentangled representations has made them popular for practical applications. However, their behaviour is not yet fully understood. For example, the questions of when they can provide disentangled representations, or suffer from posterior collapse are still areas of active research. Despite this, there are no layerwise comparisons of the representations learned by VAEs, which would further our understanding of these models. In this paper, we thus look into the internal behaviour of VAEs using representational similarity techniques. Specifically, using the CKA and Procrustes similarities, we found that the encoders' representations are learned long before the decoders', and this behaviour is independent of hyperparameters, learning objectives, and datasets. Moreover, the encoders' representations in all but the mean and variance layers are similar across hyperparameters and learning objectives.
Item Type: | Preprint |
---|---|
DOI/Identification number: | 10.48550/arXiv.2205.08399 |
Refereed: | No |
Other identifier: | https://arxiv.org/abs/2205.08399 |
Name of pre-print platform: | arXiv |
Subjects: | Q Science > QA Mathematics (inc Computing science) > QA 76 Software, computer programming, |
Divisions: | Divisions > Division of Computing, Engineering and Mathematical Sciences > School of Computing |
Funders: | University of Kent (https://ror.org/00xkeyj56) |
Depositing User: | Marek Grzes |
Date Deposited: | 06 Dec 2022 10:58 UTC |
Last Modified: | 05 Nov 2024 13:04 UTC |
Resource URI: | https://kar.kent.ac.uk/id/eprint/98783 (The current URI for this page, for reference purposes) |
- Export to:
- RefWorks
- EPrints3 XML
- BibTeX
- CSV
- Depositors only (login required):