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

Recently, the efficient deployment and acceleration of powerful visiontransformers (ViTs) on resource-limited edge devices for providing multimediaservices have become attractive tasks. Although early exiting is a feasiblesolution for accelerating inference, most works focus on convolutional neuralnetworks (CNNs) and transformer models in natural language processing(NLP).Moreover, the direct application of early exiting methods to ViTs mayresult in substantial performance degradation. To tackle this challenge, wesystematically investigate the efficacy of early exiting in ViTs and point outthat the insufficient feature representations in shallow internal classifiersand the limited ability to capture target semantic information in deep internalclassifiers restrict the performance of these methods. We then propose an earlyexiting framework for general ViTs termed LGViT, which incorporatesheterogeneous exiting heads, namely, local perception head and globalaggregation head, to achieve an efficiency-accuracy trade-off. In particular,we develop a novel two-stage training scheme, including end-to-end training andself-distillation with the backbone frozen to generate early exiting ViTs,which facilitates the fusion of global and local information extracted by thetwo types of heads. We conduct extensive experiments using three popular ViTbackbones on three vision datasets. Results demonstrate that our LGViT canachieve competitive performance with approximately 1.8 $\times$ speed-up.

LGViT: Dynamic Early Exiting for Accelerating Vision Transformer