AI Paper Atlas

Research map / Detection and segmentation / CNN features

MutualNet: Adaptive ConvNet via Mutual Learning from Network Width and Resolution

Taojiannan Yang, Sijie Zhu, Chen Chen, Shen Yan, Mi Zhang, Andrew Willis

ECCV 2020 · Detection and segmentation › CNN features · approach: Efficient network training and design

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
feed different input resolutions to sub-networks with width-resolution mutual learning, replacing constant 224 inputUS-Net (MobileNet v1) · ImageNet, 150 MFLOPstop-1 accuracy (%): 62.9 -> 65.6Improved
add mutual learning from width and resolutionUS-Net (MobileNet v1) · ImageNet, 1.0x-224top-1 accuracy (%): 71.7 -> 72.4Improved
add mutual learning from width and resolutionUS-Net (MobileNet v2) · ImageNet, 1.0x-224top-1 accuracy (%): 71.5 -> 72.9Improved
add multi-scale data augmentation to MobileNet v2 baseline (ablation)MobileNet v2 · ImageNet, 1.0x-224top-1 accuracy (%): 71.8 -> 72.0No clear effect
add width-resolution mutual learning instead of multi-scale data augmentation (ablation)MobileNet v2 · ImageNet, 1.0x-224top-1 accuracy (%): 72.0 -> 72.9Improved
replace EfficientNet grid-searched scaling with MutualNet scalingMobileNet v1 · ImageNet, 2.3B FLOPstop-1 accuracy (%): 75.6 -> 77.1Improved
add mutual learning (width lower bound 0.9x) as training schemeWideResNet-28-10 · CIFAR-100top-1 accuracy (%): 81.2 -> 83.8Improved
add mutual learning as training schemeWideResNet-28-10 · CIFAR-10top-1 accuracy (%): 96.1 -> 97.2Improved
add mutual learning as training schemeResNet-50 · ImageNettop-1 accuracy (%): 76.5 -> 78.6Improved
add knowledge distillation from a teacher networkResNet-50 · ImageNettop-1 accuracy (%): 76.5 -> 76.5No clear effect
raise width lower bound from 0.7x to 0.9x (ablation)MutualNet (MobileNet v2) · ImageNetaccuracy-FLOPsMixed
apply different resolutions only at inference (US-Net+) instead of training with them (ablation)US-Net · ImageNet MobileNet v1resolution selectionHurt

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