User: Guest  Login

Document type:
Zeitschriftenaufsatz
Author(s):
Joppich, Philipp; Dorn, Sebastian; De Candido, Oliver; Knollmüller, Jakob; Utschick, Wolfgang
Title:
Classification and Uncertainty Quantification of Corrupted Data Using Supervised Autoencoders
Abstract:
Parametric and non-parametric classifiers often have to deal with real-world data, where corruptions such as noise, occlusions, and blur are unavoidable. We present a probabilistic approach to classify strongly corrupted data and quantify uncertainty, even though the corrupted data do not have to be included to the training data. A supervised autoencoder is the underlying architecture. We used the decoding part as a generative model for realistic data and extended it by convolutions, masking, and additive Gaussian noise to describe imperfections. This constitutes a statistical inference task in terms of the optimal latent space activations of the underlying uncorrupted datum. We solved this problem approximately with Metric Gaussian Variational Inference (MGVI). The supervision of the autoencoder’ s latent space allowed us to classify corrupted data directly under uncertainty with the statistically inferred latent space activations. We show that the derived model uncertainty can be used as a statistical “ lie detector” of the classification. Independent of that, the generative model can optimally restore the corrupted datum by decoding the inferred latent space activations.
Journal title:
Physical Sciences Forum
Year:
2022
Journal volume:
5
Journal issue:
1
Fulltext / DOI:
doi:10.3390/psf2022005012
WWW:
https://www.mdpi.com/2673-9984/5/1/12
Print-ISSN:
2673-9984
 BibTeX