Early-exit network

Early-exit networks are a class of dynamic neural networks designed for efficient inference by allowing models to make confident predictions at intermediate layers, rather than processing the full network.[1]

Early-exit mechanisms are methods for deep neural networks that add intermediate classifiers, allowing inference to stop at earlier layers for inputs assessed as low uncertainty. Decisions to exit are typically based on confidence measures such as softmax-derived scores, classification margins, or entropy-based criteria, with the goal of reducing computational cost. These approaches are commonly paired with specialized training procedures and system-level optimizations to improve efficiency while preserving accuracy.[2]

The main idea behind the technology is to stop excessive calculations when a good answer can already be given with a high degree of probability, which can save both computation and time.[3][4]

Early-exit networks have also been extended with expert-based exit criteria, where intermediate classifiers are treated as multiple “experts” whose predictions and confidence scores can be aggregated to decide whether to stop computation early.[5]

Hardware implementations are also being developed.[6]

See also

References

  1. ^ "Early-Exit Deep Neural Network - A Comprehensive Survey". ACM Digital Library. doi:10.1145/3698767. Retrieved 11 February 2026.
  2. ^ "Early-Exit Mechanisms in Neural Networks". emergentmind.com. Retrieved 15 February 2026.
  3. ^ "BranchyNet: Fast Inference via Early Exiting from Deep Neural Networks". Retrieved 11 February 2026.
  4. ^ "Confidence-gated training for efficient early-exit neural networks". Retrieved 11 February 2026.
  5. ^ Bajpai, Divya Jyoti; Hanawal, Manjesh Kumar (4 October 2024). "BEEM: Boosting Performance of Early Exit DNNs using Multi-Exit Classifiers as Experts". Retrieved 16 February 2026.
  6. ^ Biggs, Benjamin; Bouganis, Christos-Savvas; Constantinides, George (May 2023). "ATHEENA: A Toolflow for Hardware Early-Exit Network Automation". 2023 IEEE 31st Annual International Symposium on Field-Programmable Custom Computing Machines (FCCM). pp. 121–132. doi:10.1109/FCCM57271.2023.00022. Retrieved 11 February 2026.


This article is issued from Wikipedia. The text is available under Creative Commons Attribution-Share Alike 4.0 unless otherwise noted. Additional terms may apply for the media files.