Enriching Disentanglement: Definitions to Metrics
Abstract
A multitude of metrics for learning and evaluating disentangled representations has been proposed. However, it remains unclear what these metrics truly quantify and how to compare them. To solve this problem, we introduce a systematic approach for transforming an equational definition into a quantitative metric via enriched category theory. We show how to replace (i) equality with metric, (ii) logical connectives with order operations, (iii) universal quantifier with aggregation, and (iv) existential quantifier with the best approximation. Using this approach, we can derive useful metrics for measuring the modularity and informativeness of a disentangled representation extractor.
Cite
Text
Zhang and Sugiyama. "Enriching Disentanglement: Definitions to Metrics." ICML 2023 Workshops: TAGML, 2023.Markdown
[Zhang and Sugiyama. "Enriching Disentanglement: Definitions to Metrics." ICML 2023 Workshops: TAGML, 2023.](https://mlanthology.org/icmlw/2023/zhang2023icmlw-enriching/)BibTeX
@inproceedings{zhang2023icmlw-enriching,
title = {{Enriching Disentanglement: Definitions to Metrics}},
author = {Zhang, Yivan and Sugiyama, Masashi},
booktitle = {ICML 2023 Workshops: TAGML},
year = {2023},
url = {https://mlanthology.org/icmlw/2023/zhang2023icmlw-enriching/}
}