Some Remarks on Identifiability of Independent Component Analysis in Restricted Function Classes
Abstract
In this short note, we comment on recent results on identifiability of independent component analysis. We point out an error in earlier works and clarify that this error cannot be fixed as the chosen approach is not sufficiently powerful to prove identifiability results. In addition, we explain the necessary ingredients to prove stronger identifiability results. Finally, we discuss and extend the flow-based technique to construct spurious solutions for independent component analysis problems and provide a counterexample to an earlier identifiability result.
Cite
Text
Buchholz. "Some Remarks on Identifiability of Independent Component Analysis in Restricted Function Classes." Transactions on Machine Learning Research, 2023.Markdown
[Buchholz. "Some Remarks on Identifiability of Independent Component Analysis in Restricted Function Classes." Transactions on Machine Learning Research, 2023.](https://mlanthology.org/tmlr/2023/buchholz2023tmlr-some/)BibTeX
@article{buchholz2023tmlr-some,
title = {{Some Remarks on Identifiability of Independent Component Analysis in Restricted Function Classes}},
author = {Buchholz, Simon},
journal = {Transactions on Machine Learning Research},
year = {2023},
url = {https://mlanthology.org/tmlr/2023/buchholz2023tmlr-some/}
}