An illustrative correcting map sends nearby observations onto a curve. The autoencoder sketch highlights two controls on dimensionality: its code space and the rank of its Jacobian. The paper also penalizes curvature of the learned map. Research source
About this paper
Uses Jacobian rank constraints to encourage autoencoders to learn low-dimensional manifolds.
Manifold learning is formulated as finding a correcting map onto a hidden manifold. A Jacobian penalty and a curvature term supplement the architectural bottleneck, with experiments on synthetic and real datasets.
Citation
BibTeX
@article{takhanov2023autoencodersjacobianrank,
title = {Autoencoders for a manifold learning problem with a Jacobian rank constraint},
author = {Takhanov, Rustem and Abylkairov, Y. Sultan and Tezekbayev, Maxat},
year = {2023},
journal = {Pattern Recognition},
eprint = {2306.14194},
archivePrefix = {arXiv},
url = {https://arxiv.org/pdf/2306.14194}
}