AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Harmonized Dense Knowledge Distillation Training for Multi-Exit Architectures

Xinglu Wang, Yingming Li

AAAI 2021 · Architectures and efficiency › RNNs and normalization · approach: Dataset distillation and data-efficient training · 19 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
Use dense distillation with harmonized adaptive loss weightsMSDNet · ImageNet(full), Exit-1top-5 accuracy: 80.15 -> 82.44Improved
Use dense distillation with harmonized adaptive loss weightsMSDNet · ImageNet(full), Exit-5top-5 accuracy: 91.74 -> 92.69Improved
Use dense distillation with harmonized adaptive loss weightsMSDNet · CIFAR(500), Exit-1top-5 accuracy: 86.47 -> 87.64Improved
Use dense distillation with harmonized adaptive loss weightsMSDNet · CIFAR(500), Exit-8top-5 accuracy: 93.46 -> 94.02Improved
Change the initial loss-weight configuration (ablation)HDKD · ImageNet(300), Exit-1top-5 accuracy: 78.72 -> 79.22No clear effect
Keep the final learned loss weights fixed from the start instead of evolving them during training (ablation)HDKD · ImageNet(300), Exit-5top-5 accuracy: 87.37 -> 83.33; 87.37 -> 87.94Hurt
Remove the ratio regularizer from the validation loss (ablation)HDKD · ImageNet(300), Exit-5top-5 accuracy: 87.37 -> 86.11Hurt
Use only validation cross-entropy, without the ratio regularizer (ablation)HDKD · ImageNet(300), Exit-1top-5 accuracy: 78.72 -> 79.55Improved

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