AI Paper Atlas

Research map / Architectures and efficiency / Pruning and efficient tuning

Sensitivity-Aware Visual Parameter-Efficient Fine-Tuning

Haoyu He, Jianfei Cai, Jing Zhang, Dacheng Tao, Bohan Zhuang

ICCV 2023 · Architectures and efficiency › Pruning and efficient tuning · approach: Adapter-based parameter-efficient tuning · 82 citations (OpenAlex)

Read the paper

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
allocate trainable parameters by sensitivity (task-specific positions) with Adapter-8 on sensitive matrices (ablation)Adapter-8 · VTAB-1k mean (supervised ViT-B/16)top-1 accuracy (%): 73.9 -> 75.5Improved
add unstructured tuning of sensitive connections to structured tuning (ablation)SPT-Adapter · VTAB-1k mean (supervised ViT-B/16)top-1 accuracy (%): 75.5 -> 75.8Improved
add structured tuning to unstructured tuning (ablation)SPT-Adapter · VTAB-1k Structuredtop-1 accuracy (%): 59.6 -> 60.8Improved
apply sensitivity criterion to pick matrices for Deep Prompt, Adapter-32, LoRA-16 (ablation)Deep Prompt, Adapter-32, LoRA-16 · VTAB-1k (supervised ViT-B/16)mean top-1 accuracy (%): +1.1%, +1.6%, +0.8%Improved
reduce number of samples C for sensitivity from 800 to 400 (ablation)SPT-LoRA · VTAB-1k mean (supervised ViT-B/16)top-1 accuracy (%): 76.4 -> 76.4No clear effect
apply SPT to self-supervised MAE backboneAdapter / LoRA · VTAB-1k mean (MAE ViT-B/16)top-1 accuracy (%): 64.8 -> 69.8Improved
use unstructured sparse-gradient tuning in SPT-LoRA (ablation)SPT-LoRA · VTAB-1k (ViT-B/16)fine-tuning memory (GB): 8.3 -> 9.8Hurt
apply SPT to Adapter on FGVCAdapter-8 · FGVC mean (supervised ViT-B/16)top-1 accuracy (%): 85.5 -> 89.5Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy