Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2025
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2508.21495 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866917539241525248 |
|---|---|
| author | Kubaty, Piotr Szatkowski, Filip Choczyński, Grzegorz Nalisnick, Eric Wójcik, Bartosz |
| author_facet | Kubaty, Piotr Szatkowski, Filip Choczyński, Grzegorz Nalisnick, Eric Wójcik, Bartosz |
| contents | Early-exit neural networks (EENNs) accelerate inference by allowing intermediate classifiers to stop computation once predictions are confident enough. Most methods rely on confidence thresholds for exiting, and consequently, improving classifier calibration is widely assumed to improve performance. In this work, we challenge this assumption and show that calibration alone is not sufficient for EENNs to exploit adaptive computation. To address this insufficiency, we introduce Early-Exit Failure Prediction (EEFP), which accounts for both prediction correctness and the cost of further computation. We also propose a lightweight, EEFP-motivated procedure to improve the intermediate classifiers, which can directly replace calibration in EENNs. Extensive experiments demonstrate that our approach achieves superior cost-accuracy trade-offs compared to calibration, and EEFP more reliably reflects overall EENN performance. Our code is available at https://github.com/gmum/rethinking-calibration-for-eenns. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2508_21495 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Rethinking Calibration for Early-Exit Neural Networks Kubaty, Piotr Szatkowski, Filip Choczyński, Grzegorz Nalisnick, Eric Wójcik, Bartosz Machine Learning Early-exit neural networks (EENNs) accelerate inference by allowing intermediate classifiers to stop computation once predictions are confident enough. Most methods rely on confidence thresholds for exiting, and consequently, improving classifier calibration is widely assumed to improve performance. In this work, we challenge this assumption and show that calibration alone is not sufficient for EENNs to exploit adaptive computation. To address this insufficiency, we introduce Early-Exit Failure Prediction (EEFP), which accounts for both prediction correctness and the cost of further computation. We also propose a lightweight, EEFP-motivated procedure to improve the intermediate classifiers, which can directly replace calibration in EENNs. Extensive experiments demonstrate that our approach achieves superior cost-accuracy trade-offs compared to calibration, and EEFP more reliably reflects overall EENN performance. Our code is available at https://github.com/gmum/rethinking-calibration-for-eenns. |
| title | Rethinking Calibration for Early-Exit Neural Networks |
| topic | Machine Learning |
| url | https://arxiv.org/abs/2508.21495 |