AI Paper Atlas

Research map / Large language models / Instruction and fine-tuning

Non-transferable Pruning

Ruyi Ding, Lili Su, A. Adam Ding, Yunsi Fei

ECCV 2024 · Large language models › Instruction and fine-tuning · approach: Safety alignment after fine-tuning · 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
replace NTL/CUTI with ADMM-based pruning plus Fisher space regularization (NTP)VGG-11 digits transfer · Source USPS -> target MNISTSLC-AUC: 0.42 -> -0.47 (vs NTL), 0.89 -> -0.47 (vs CUTI)Improved
use NTP on highly similar RGB digit datasets SYN and SVHNNTP VGG-11 · Source SYN -> target SVHNSLC-AUC: 1.52 -> 0.51 (NTL vs NTP)Mixed
use NTP versus NTL and CUTI under full fine-tuning across learning ratesNTP VGG-11 · MNIST -> USPS, full fine-tuning, 10% datatarget accuracy: 0.96 -> 0.21 (CUTI vs NTP, lr 1e-5)Improved
use NTP versus CUTI under linear probing at lr 1e-3NTP VGG-11 · MNIST -> USPS, linear probingtarget accuracy: 0.18 -> 0.28 (CUTI vs NTP)Hurt
apply NTP instead of keeping the original pretrained model on source accuracyNTP VGG-11 · Source SVHN accuracysource accuracy change: 4.10% down (NTP) vs 0.17% up (CUTI)Hurt
add NTP to self-supervised modelsSimCLR, MoCo-v1, MoCo-v2 ResNet-50 · STL-10 transferability (LogME)LogME: 6.19 -> -1.11 (MoCo-v2)Improved
apply NTP to self-supervised versus supervised modelsMoCo-v1 and MoCo-v2 · CIFAR-10 -> STL-10 fine-tuningnon-transferability after fine-tuningMixed
increase sparsity of the NTP-pruned model (ablation)NTP ResNet-18 and VGG-11 · CIFAR-10 -> STL-10SLC-AUC / transfer performanceImproved
increase ADMM weight-update epochs from 1 to 20 (ablation)NTP ResNet-18 · CIFAR-10 -> STL-10SLC-AUCNo clear effect
increase the amount of training data used in ADMM pruning (ablation)NTP ResNet-18 · CIFAR-10 -> STL-10non-transferabilityImproved
apply one-shot magnitude pruning at high sparsityUSPS-pretrained ResNet-18 · USPS -> various targetsSLC-AUCMixed

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