Search: onr:"swepub:oai:DiVA.org:liu-198763" >
Enhancing Represent...
-
Ahmadian, Amirhossein,1992-Linköpings universitet,Statistik och maskininlärning,Tekniska fakulteten
(author)
Enhancing Representation Learning with Deep Classifiers in Presence of Shortcut
- Article/chapterEnglish2023
Publisher, publication year, extent ...
Numbers
-
LIBRIS-ID:oai:DiVA.org:liu-198763
-
https://urn.kb.se/resolve?urn=urn:nbn:se:liu:diva-198763URI
-
https://doi.org/10.1109/ICASSP49357.2023.10096346DOI
Supplementary language notes
-
Language:English
-
Summary in:English
Part of subdatabase
Classification
-
Subject category:ref swepub-contenttype
-
Subject category:kon swepub-publicationtype
Notes
-
A deep neural classifier trained on an upstream task can be leveraged to boost the performance of another classifier in a related downstream task through the representations learned in hidden layers. However, presence of shortcuts (easy-to-learn features) in the upstream task can considerably impair the versatility of intermediate representations and, in turn, the downstream performance. In this paper, we propose a method to improve the representations learned by deep neural image classifiers in spite of a shortcut in upstream data. In our method, the upstream classification objective is augmented with a type of adversarial training where an auxiliary network, so called lens, fools the classifier by exploiting the shortcut in reconstructing images. Empirical comparisons in self-supervised and transfer learning problems with three shortcut-biased datasets suggest the advantages of our method in terms of downstream performance and/or training time.
Subject headings and genre
Added entries (persons, corporate bodies, meetings, titles ...)
-
Lindsten, Fredrik,1984-Linköpings universitet,Reglerteknik,Statistik och maskininlärning,Tekniska fakulteten(Swepub:liu)freli29
(author)
-
Linköpings universitetStatistik och maskininlärning
(creator_code:org_t)
Related titles
-
In:Proceedings of IEEE ICASSP 2023
Internet link
To the university's database