English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Deep neural networks often fail to generalize outside of their trainingdistribution, in particular when only a single data domain is available duringtraining. While test-time adaptation has yielded encouraging results in thissetting, we argue that, to reach further improvements, these approaches shouldbe combined with training procedure modifications aiming to learn a morediverse set of patterns. Indeed, test-time adaptation methods usually have torely on a limited representation because of the shortcut learning phenomenon:only a subset of the available predictive patterns is learned with standardtraining. In this paper, we first show that the combined use of existingtraining-time strategies, and test-time batch normalization, a simpleadaptation method, does not always improve upon the test-time adaptation aloneon the PACS benchmark. Furthermore, experiments on Office-Home show that veryfew training-time methods improve upon standard training, with or withouttest-time batch normalization. We therefore propose a novel approach using apair of classifiers and a shortcut patterns avoidance loss that mitigates theshortcut learning behavior by reducing the generalization ability of thesecondary classifier, using the additional shortcut patterns avoidance lossthat encourages the learning of samples specific patterns. The primaryclassifier is trained normally, resulting in the learning of both the naturaland the more complex, less generalizable, features. Our experiments show thatour method improves upon the state-of-the-art results on both benchmarks andbenefits the most to test-time batch normalization.

Learning Less Generalizable Patterns with an Asymmetrically Trained Double Classifier for Better Test-Time Adaptation