New SoftMoR technique enables deeper, more efficient Vision Transformers

Researchers have introduced Soft Mixture-of-Recursions (SoftMoR), a novel approach to enhance Vision Transformers (ViTs) by enabling them to leverage intermediate representations from all recursion steps. This method, instantiated as the Soft Recursive Vision Transformer (SR-ViT), allows for the creation of deeper and more powerful ViTs with minimal parameter increases. Experiments on ImageNet-1K showed that increasing recursion depth in SR-ViT-S from one to four improved top-1 accuracy from 79.83% to 82.48%, while using significantly fewer parameters than larger models like DeiT-B. AI

IMPACT This research offers a parameter-efficient method for developing more capable Vision Transformers, potentially impacting computer vision applications.

RANK_REASON The cluster contains an academic paper detailing a new method for improving Vision Transformers. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• DeiT-B
• Recursive Vision Transformers
• Soft Mixture-of-Recursions
• SoftMoR
• SR-ViT
• SR-ViT-S

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →