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

Global covariance pooling in convolutional neural networks has achievedimpressive improvement over the classical first-order pooling. Recent workshave shown matrix square root normalization plays a central role in achievingstate-of-the-art performance. However, existing methods depend heavily oneigendecomposition (EIG) or singular value decomposition (SVD), suffering frominefficient training due to limited support of EIG and SVD on GPU. Towardsaddressing this problem, we propose an iterative matrix square rootnormalization method for fast end-to-end training of global covariance poolingnetworks. At the core of our method is a meta-layer designed with loop-embeddeddirected graph structure. The meta-layer consists of three consecutivenonlinear structured layers, which perform pre-normalization, coupled matrixiteration and post-compensation, respectively. Our method is much faster thanEIG or SVD based ones, since it involves only matrix multiplications, suitablefor parallel implementation on GPU. Moreover, the proposed network with ResNetarchitecture can converge in much less epochs, further accelerating networktraining. On large-scale ImageNet, we achieve competitive performance superiorto existing counterparts. By finetuning our models pre-trained on ImageNet, weestablish state-of-the-art results on three challenging fine-grainedbenchmarks. The source code and network models will be available athttp://www.peihuali.org/iSQRT-COV

Towards Faster Training of Global Covariance Pooling Networks by Iterative Matrix Square Root Normalization