Early Exit Network

Early exit networks are deep learning architectures designed to improve inference efficiency by allowing predictions to be made at multiple points within the network, bypassing later layers for easy samples. Current research focuses on optimizing training strategies, such as sample weighting and incorporating continual learning methods, to better leverage this early exit capability and address issues like task-recency bias and uneven client resource availability in federated learning settings. These networks offer significant potential for resource-constrained environments, particularly in edge AI and IoT applications, by reducing computational costs and latency while maintaining competitive accuracy. Furthermore, research is exploring efficient hardware implementations and the integration of early exit networks with ensemble methods for improved uncertainty estimation.

Papers

May 7, 2024

Federated Learning for Cooperative Inference Systems: The Case of Early Exit Networks
Caelin Kaplan, Tareq Si Salem, Angelo Rodio, Chuan Xu, Giovanni Neglia
Case Relevance Inference Cost Inference Task Local Inference Early Exit Network

March 12, 2024

Auxiliary Classifiers Improve Stability and Efficiency in Continual Learning
Filip Szatkowski, Fei Yang, Bartłomiej Twardowski, Tomasz Trzciński, Joost van de Weijer
Continual LEArning Continual Learning Anti Forgetting Complementary Benefit Early Exit Early Exit Network

June 5, 2023

Towards Anytime Classification in Early-Exit Architectures by Enforcing Conditional Monotonicity
Metod Jazbec, James Urquhart Allingham, Dan Zhang, Eric Nalisnick
Deep Model Early Exit Monotonicity Constraint Anytime Valid Sequential Inference Early Exit Network Intermediate Prediction

April 17, 2023

ATHEENA: A Toolflow for Hardware Early-Exit Network Automation
Benjamin Biggs, Christos-Savvas Bouganis, George A. Constantinides
Field Programmable Gate Array Early Exit Network Network Automation

March 14, 2023

Window-Based Early-Exit Cascades for Uncertainty Estimation: When Deep Ensembles are More Efficient than Single Models
Guoxuan Xia, Christos-Savvas Bouganis
Deep Learning Uncertainty Estimation Deep Ensemble Single Model Early Exit Network Computational Uncertainty

September 17, 2022

Learning to Weight Samples for Dynamic Early-exiting Networks
Yizeng Han, Yifan Pu, Zihang Lai, Chaofei Wang, Shiji Song, Junfen Cao, Wenhui Huang, Chao Deng, Gao Huang
LeArning Abstract Inference Efficiency Early eXit Multi Exit Early Exit Network Weight Prediction

July 8, 2022

Pruning Early Exit Networks
Alperen Görmez, Erdem Koyuncu
Deep Learning Model Edge Pruning Linear Classifier Computational Cost Early Exit Network

June 15, 2022

Resource-Constrained Edge AI with Early Exit Prediction
Rongkang Dong, Yuyi Mao, Jun Zhang
Early eXit Edge Artificial Intelligence Early Exit Network Device Edge Co Inference