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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add previous-exit prediction history to the corrector input (ablation) | EEFP confidence correction · TinyImageNet ResNet-34 | accuracy at FLOPs budgets: 25%: 50.82 -> 50.67; 50%: 63.42 -> 64.06; 75%: 66.23 -> 67.03 | Mixed |
| use top-5 instead of top-1 confidences in the corrector (ablation) | EEFP confidence correction · TinyImageNet ResNet-34 | accuracy at 50% FLOPs: 62.82 -> 64.06 | Improved |
| use top-5 instead of all 200 class confidences (ablation) | EEFP confidence correction · TinyImageNet ResNet-34 | accuracy at 50% FLOPs: 63.36 -> 64.06 | Improved |
| use top-5 instead of top-10 (ablation) | EEFP confidence correction · TinyImageNet ResNet-34 | accuracy at 50% FLOPs: 63.91 -> 64.06 | No clear effect |
| add temperature calibration to the early-exit baseline | baseline EENN · TinyImageNet ResNet-34 | accuracy at 50% FLOPs: 62.44 -> 62.98 | Improved |
| replace temperature calibration with EEFP confidence correction | calibrated EENN · TinyImageNet ResNet-34 | accuracy at 50% FLOPs: 62.98 -> 64.06 | Improved |
| replace temperature calibration with EEFP confidence correction | EfficientNet-B2 EENN · ImageNet-1k | accuracy at 50% FLOPs: 62.99 -> 63.46 | Improved |
| replace temperature calibration with EEFP confidence correction | ViT-Small EENN · ImageNet-1k | accuracy at 25% FLOPs: 39.40 -> 40.18 | Improved |
| replace temperature calibration with EEFP confidence correction | MSDNet EENN · CIFAR-100 | accuracy at 75% FLOPs: 75.46 -> 75.58 | No clear effect |
| train corrector on EEFP stopping target instead of correctness target (CCCT) (ablation) | confidence corrector top-5 · TinyImageNet ResNet-34 | accuracy at 50% FLOPs: 63.49 -> 64.06 | Improved |
| use matrix scaling instead of temperature scaling (ablation) | calibrated EENN · TinyImageNet ResNet-34 | accuracy at 50% FLOPs: 62.98 -> 54.79 | Hurt |
| apply EEFP correction instead of calibration on Gradient Equilibrium model | Gradient Equilibrium EENN · TinyImageNet ResNet-34 | accuracy at 50% FLOPs: 60.33 -> 61.31 | Improved |
Closest papers by meaning
- Fast yet Safe: Early-Exiting with Risk Control (NeurIPS 2024 · similarity 0.87)
- Jointly-Learned Exit and Inference for a Dynamic Neural Network (ICLR 2024 · similarity 0.87)
- BEEM: Boosting Performance of Early Exit DNNs using Multi-Exit Classifiers as Experts (ICLR 2025 · similarity 0.85)
- A Benchmark Study on Calibration (ICLR 2024 · similarity 0.85)
- Balancing Two Classifiers via A Simplex ETF Structure for Model Calibration (CVPR 2025 · similarity 0.85)
- Beyond Greedy Exits: Improved Early Exit Decisions for Risk Control and Reliability (NeurIPS 2025 · similarity 0.85)
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.
