AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Revisit the Power of Vanilla Knowledge Distillation: from Small Scale to Large Scale

Zhiwei Hao, Jianyuan Guo, Kai Han, Han Hu, Chang Xu, Yunhe Wang

NeurIPS 2023 · Architectures and efficiency › RNNs and normalization · approach: Neural scaling laws · 1 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
stronger recipe (AdamW, 300 epochs) vs previous recipe (ablation)vanilla KD · Res50-MBNetV2 ImageNet-1Ktop-1 accuracy (%): 68.58 -> 74.23Improved
stronger recipe (AdamW, 300 epochs) vs previous recipe (ablation)vanilla KD · Res34-Res18 ImageNet-1Ktop-1 accuracy (%): 70.66 -> 73.46Improved
stronger recipe with 2400 epochs on CIFAR-100 (ablation)vanilla KD · Res56-Res20 CIFAR-100gap to best method: 1.31% to 2.17%Hurt
replace hint-based distillation by logits-basedvanilla KD · ImageNet-1K Res152 teacher A2top-1 accuracy (%): 79.55 -> 80.55Improved
replace multi-class CE by BCE for hard label (ablation)vanilla KD · ImageNet-1K 300eptop-1 accuracy (%): 80.49 -> 80.89Improved
replace KL by binary KL (BKL) for soft label (ablation)vanilla KD · ImageNet-1K 300eptop-1 accuracy (%): 80.89 -> 79.36Hurt
remove hard-label supervision (ablation)vanilla KD · ImageNet-1K 600eptop-1 accuracy (%): 81.68 -> 81.65No clear effect
vary learning rate and weight decay (ablation)vanilla KD · ImageNet-1K Res50 A2top-1 accuracy (%): 79.77 -> 80.89No clear effect
increase training epochs from 300 to 4800 (ablation)vanilla KD · Res50 studenttop-1 accuracy (%): 80.96 -> 83.08Improved
increase epochs from 300 to 4800 (ablation)vanilla KD · ViT-S studenttop-1 accuracy (%): 81.38 -> 84.33Improved
use larger BEiTv2-L teacher instead of BEiTv2-B (ablation)vanilla KD · Res50, 1200eptop-1 accuracy (%): 82.35 -> 82.27Hurt
teacher pre-trained on IN-21K instead of IN-1K (ablation)vanilla KD · Res50 studenttop-1 accuracy (%): 81.97 -> 82.35Improved

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