AI Paper Atlas

Research map / Architectures and efficiency / Pruning and efficient tuning

Earning Extra Performance From Restrictive Feedbacks

Jing Li, Yuangang Pan, Yueming Lyu, Yinghua Yao, Yulei Sui, Ivor Wai-Hung Tsang

TPAMI 2023 · Architectures and efficiency › Pruning and efficient tuning · approach: Low-rank compression and merging · 2 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

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
add layer importance based query reassignment to layerwise tuning (LCPS) (ablation)LCPS · CIFAR-10-C Gaussianqueries to reach 22.0% error: ~3.0K -> ~0.2KImproved
add layer importance based query reassignment to layerwise tuning (LCPS) (ablation)LCPS · CIFAR-10-C Impulsequeries to reach 20.0% error: >8.4K -> ~3.5KImproved
use layerwise LCPS instead of whole-model PPS (ablation)CIFAR-10-C tuning · CIFAR-10-C Gaussianqueries to reach 22.0% error: >10.0K -> ~3.0KImproved
use LCPS with feedback only instead of Tentcorrupted CIFAR-10/100 · CIFAR-10-C average error (Q=1K)errorImproved
use PPS instead of Random SearchAdult income classification · Adult (Q=1K)accuracyImproved
tune with top-5 error as feedback metric instead of top-1 (ablation)LCPS · CIFAR-10-C Gaussian and Impulsetop-5 errorImproved
round feedback to 0 decimals instead of more decimals (ablation)PPS · Adultsupport accuracyHurt
increase feedback precision from 2 to 3 decimals (ablation)PPS · Adultsupport accuracyNo clear effect
increase support set ratio from 10% to 90% (ablation)PPS · Adultholdout accuracyImproved
vary sampling batch size from 2 to 80 (ablation)PPS · Adultaccuracy: optimum at b=10 support, b=20 holdoutMixed
tune with estimated full-batch gradient (LCPS) instead of standard stochastic-gradient tuningBERT variants · STS-BPearson correlationImproved

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