Offloading Adaptativo em Redes Neurais com Saídas Antecipadas: Equilíbrio e Compartilhamento de Recursos Através de Multi-Armed Bandits

• Ricardo S. Silva UFF https://orcid.org/0009-0005-9427-3302
• Roberto G. Pacheco UFF https://orcid.org/0000-0002-6763-7255
• Daniel S. Menasché UFRJ https://orcid.org/0000-0002-8953-4003
• Heudson Mirandola UFRJ https://orcid.org/0000-0003-2410-5053

DOI: https://doi.org/10.5753/sbrc.2026.19866

Resumo

Redes Neurais com Saídas Antecipadas (EENNs) reduzem custos de inferência ao classificar antecipadamente entradas em ramo lateral na borda, quando a confiança atinge um limiar fixo. Caso contrário, ocorre o offloading à nuvem. Contudo, um limiar fixo não se adapta às dinâmicas de aplicações reais e a competição de múltiplos dispositivos aos recursos da nuvem. Este trabalho modela a interação entre dispositivos via jogos estocásticos e multi-armed bandits para tornar o ajuste de limiares dinâmico, considerando restrições energéticas e disponibilidade de servidores. Os resultados numéricos mostram obtenção de equilíbrios de Nash aproximados.

Referências

Albrecht, S. V., Christianos, F., and Schäfer, L. (2024). Multi-Agent Reinforcement Learning: Foundations and Modern Approaches.

Altman, E. (1999). Constrained Markov Decision Processes, volume 7 of Stochastic Modeling. Boca Raton, FL.

Bajpai, D. J. and Hanawal, M. K. (2025). BEEM: Boosting performance of early exit DNNs using multi-exit classifiers as experts. In International Conference on Learning Representations (ICLR).

Casale, G. and Roveri, M. (2023). Scheduling inputs in early exit neural networks. IEEE Transactions on Computers, 73(2):451–465.

Fang, B., Zeng, X., Zhang, F., Xu, H., and Zhang, M. (2020). Flexdnn: Input-adaptive on-device deep learning for efficient mobile vision. In IEEE/ACM Symposium on Edge Computing (SEC), pages 84–95.

Ju, W., Bao, W., Ge, L., and Yuan, D. (2021a). Dynamic early exit scheduling for deep neural network inference through contextual bandits. In International Conference on Information Knowledge Management (CIKM), pages 823–832.

Ju, W., Bao, W., Yuan, D., Ge, L., and Zhou, B. B. (2021b). Learning early exit for deep neural network inference on mobile devices through multi-armed bandits. In IEEE International Symposium on Cluster, Cloud, and Internet Computing (IEEE/ACM CCGrid), pages 11–20.

Kim, G. and Park, J. (2020). Low cost early exit decision unit design for cnn accelerator. In IEEE International SoC Design Conference (ISOCC), pages 127–128.

Krizhevsky, A., Sutskever, I., and Hinton, G. E. (2012). Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25.

Laskaridis, S., Venieris, S. I., Almeida, M., Leontiadis, I., and Lane, N. D. (2020). SPINN: synergistic progressive inference of neural networks over device and cloud. In ACM International Conference on Mobile Computing and Networking (MobiCom), pages 1–15.

Li, E., Zeng, L., Zhou, Z., and Chen, X. (2019). Edge ai: On-demand accelerating deep neural network inference via edge computing. IEEE Transactions on Wireless Communications, 19(1):447–457.

Liu, C., Liu, K., Guo, S., Xie, R., Lee, V. C., and Son, S. H. (2020a). Adaptive offloading for time-critical tasks in heterogeneous internet of vehicles. IEEE Internet of Things Journal, 7(9):7999–8011.

Liu, Y., Peng, M., Roedig, U., and Rodrigues, J. J. (2020b). Vehicular edge computing and networks: A survey. Mobile Networks and Applications, 25:1157–1168.

Mao, Y., You, C., Zhang, J., Huang, K., and Letaief, K. B. (2017). A survey on mobile edge computing: The communication, computation, and energy perspective. IEEE Communications Surveys & Tutorials, 19(4):2322–2358.

Pacheco, R. G., Bajpai, D. J., Shifrin, M., Couto, R. S., Menasché, D. S., Hanawal, M. K., and Campista, M. E. M. (2024). Ucbee: A multi armed bandit approach for early-exit in neural networks. IEEE Transactions on Network and Service Management.

Pacheco, R. G., Bajpai, D. J., Shifrin, M., Couto, R. S., Menasché, D. S., Hanawal, M. K., and Campista, M. E. M. (2025). Otimizando saídas antecipadas em redes neurais profundas: Como lidar com buffers? In Simpósio Brasileiro de Redes de Computadores e Sistemas Distribuídos (SBRC), pages 560–573.

Pacheco, R. G., Couto, R. S., and Simeone, O. (2021). Calibration-aided edge inference offloading via adaptive model partitioning of deep neural networks. In IEEE International Conference on Communications (ICC), pages 1–6.

Rahmath P., H., Srivastava, V., Chaurasia, K., Pacheco, R. G., and Couto, R. S. (2024). Early-exit deep neural network-a comprehensive survey. ACM Computing Surveys, 57(3):1–37.

Satyanarayanan, M. (2017). The emergence of edge computing. Computer, 50(1):30–39.

Shi, W., Cao, J., Zhang, Q., Li, Y., and Xu, L. (2016). Edge computing: Vision and challenges. IEEE Internet of Things Journal, 3(5):637–646.

Slivkins, A. et al. (2019). Introduction to multi-armed bandits. Foundations and Trends® in Machine Learning, 12(1-2):1–286.

Teerapittayanon, S., McDanel, B., and Kung, H.-T. (2016). Branchynet: Fast inference via early exiting from deep neural networks. In IEEE International Conference on Pattern Recognition (ICPR), pages 2464–2469.

Wang, M., Mo, J., Lin, J., Wang, Z., and Du, L. (2019a). Dynexit: A dynamic early-exit strategy for deep residual networks. In IEEE International Workshop on Signal Processing Systems (SiPS), pages 178–183.

Wang, Z., Bao, W., et al. (2019b). SEE: Scheduling early exit for mobile dnn inference during service outage. In ACM International Conference on Modeling, Analysis and Simulation of Wireless and Mobile Systems (MSWiM), pages 279–288.

Zamzam, M., El-Shabrawy, T., and Ashour, M. (2020). Game theory for computation offloading and resource allocation in edge computing: A survey. In IEEE Novel Intelligent and Leading Emerging Sciences Conference (NILES), pages 47–53.