AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Rethinking Calibration for Early-Exit Neural Networks

Piotr Kubaty, Filip Szatkowski, Grzegorz Choczyński, Eric Nalisnick, Bartosz Wójcik

ICML 2026 · Architectures and efficiency › RNNs and normalization · approach: Efficient network compression and training · 0 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 previous-exit prediction history to the corrector input (ablation)EEFP confidence correction · TinyImageNet ResNet-34accuracy at FLOPs budgets: 25%: 50.82 -> 50.67; 50%: 63.42 -> 64.06; 75%: 66.23 -> 67.03Mixed
use top-5 instead of top-1 confidences in the corrector (ablation)EEFP confidence correction · TinyImageNet ResNet-34accuracy at 50% FLOPs: 62.82 -> 64.06Improved
use top-5 instead of all 200 class confidences (ablation)EEFP confidence correction · TinyImageNet ResNet-34accuracy at 50% FLOPs: 63.36 -> 64.06Improved
use top-5 instead of top-10 (ablation)EEFP confidence correction · TinyImageNet ResNet-34accuracy at 50% FLOPs: 63.91 -> 64.06No clear effect
add temperature calibration to the early-exit baselinebaseline EENN · TinyImageNet ResNet-34accuracy at 50% FLOPs: 62.44 -> 62.98Improved
replace temperature calibration with EEFP confidence correctioncalibrated EENN · TinyImageNet ResNet-34accuracy at 50% FLOPs: 62.98 -> 64.06Improved
replace temperature calibration with EEFP confidence correctionEfficientNet-B2 EENN · ImageNet-1kaccuracy at 50% FLOPs: 62.99 -> 63.46Improved
replace temperature calibration with EEFP confidence correctionViT-Small EENN · ImageNet-1kaccuracy at 25% FLOPs: 39.40 -> 40.18Improved
replace temperature calibration with EEFP confidence correctionMSDNet EENN · CIFAR-100accuracy at 75% FLOPs: 75.46 -> 75.58No clear effect
train corrector on EEFP stopping target instead of correctness target (CCCT) (ablation)confidence corrector top-5 · TinyImageNet ResNet-34accuracy at 50% FLOPs: 63.49 -> 64.06Improved
use matrix scaling instead of temperature scaling (ablation)calibrated EENN · TinyImageNet ResNet-34accuracy at 50% FLOPs: 62.98 -> 54.79Hurt
apply EEFP correction instead of calibration on Gradient Equilibrium modelGradient Equilibrium EENN · TinyImageNet ResNet-34accuracy at 50% FLOPs: 60.33 -> 61.31Improved

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