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 Tools annual

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 Tools monthly

Global: Zendy Plus monthly

Presents the access of premium content as premium feature

Premium Content

Global: Zendy Plus annual

Global: Zendy Free Plan

Recent Self-Supervised Learning (SSL) methods are able to learn featurerepresentations that are invariant to different data augmentations, which canthen be transferred to downstream tasks of interest. However, differentdownstream tasks require different invariances for their best performance, sothe optimal choice of augmentations for SSL depends on the target task. In thispaper, we aim to learn self-supervised features that generalize well across avariety of downstream tasks (e.g., object classification, detection andinstance segmentation) without knowing any task information beforehand. We doso by Masked Augmentation Subspace Training (or MAST) to encode in the singlefeature space the priors from different data augmentations in a factorized way.Specifically, we disentangle the feature space into separate subspaces, eachinduced by a learnable mask that selects relevant feature dimensions to modelinvariance to a specific augmentation. We show the success of MAST in jointlycapturing generalizable priors from different augmentations, using both uniqueand shared features across the subspaces. We further show that MAST benefitsfrom uncertainty modeling to reweight ambiguous samples from strongaugmentations that may cause similarity mismatch in each subspace. Experimentsdemonstrate that MAST consistently improves generalization on variousdownstream tasks, while being task-agnostic and efficient during SSL. We alsoprovide interesting insights about how different augmentations are related andhow uncertainty reflects learning difficulty.

MAST: Masked Augmentation Subspace Training for Generalizable Self-Supervised Priors