ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness

DSpace Repositorium (Manakin basiert)

ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness

Autor(en): Geirhos, Robert; Rubisch, Patricia; Michaelis, Claudio; Bethge, Matthias; Wichmann, Felix A.; Brendel, Wieland
Tübinger Autor(en):
Bethge, Matthias
Wichmann, Felix A.
Geirhos, Robert
Rubisch, Patricia
Michaelis, Claudio
Brendel, Wieland
Erschienen in: arXiv (2018-11-29), Bd. 1811.12231
Sprache: Englisch
Referenz zum Volltext: https://arxiv.org/abs/1811.12231
DDC-Klassifikation: 004 - Informatik
Dokumentart: Preprint
Zur Langanzeige

Das Dokument erscheint in: