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

Despite achieving remarkable performance on various vision-language tasks,Transformer-based pretrained vision-language models (VLMs) still suffer fromefficiency issues arising from long inputs and numerous parameters, limitingtheir real-world applications. However, the huge computation is redundant formost samples and the degree of redundancy and the respective components varysignificantly depending on tasks and input instances. In this work, we proposean adaptive acceleration method SmartTrim for VLMs, which adjusts the inferenceoverhead based on the complexity of instances. Specifically, SmartTrimincorporates lightweight trimming modules into the backbone to performtask-specific pruning on redundant inputs and parameters, without the need foradditional pre-training or data augmentation. Since visual and textualrepresentations complement each other in VLMs, we propose to leveragecross-modal interaction information to provide more critical semantic guidancefor identifying redundant parts. Meanwhile, we introduce a self-distillationstrategy that encourages the trimmed model to be consistent with thefull-capacity model, which yields further performance gains. Experimentalresults demonstrate that SmartTrim significantly reduces the computationoverhead (2-3 times) of various VLMs with comparable performance (only a 1-2%degradation) on various vision-language tasks. Compared to previousacceleration methods, SmartTrim attains a better efficiency-performancetrade-off, demonstrating great potential for application inresource-constrained scenarios.

SmartTrim: Adaptive Tokens and Parameters Pruning for Efficient Vision-Language Models