Redes Neurais Profundas com Saídas Antecipadas para Imagens com Distorção em Ambientes de Nuvem

• Roberto G. Pacheco UFRJ
• Fernanda D. V. R. Oliveira UFRJ
• Rodrigo S. Couto UFRJ

DOI: https://doi.org/10.5753/sbrc.2023.524

Abstract

Deep neural networks (DNNs) are sensitive to distorted images, reducing their accuracy. This work analyzes how solutions of early-exit DNNs (EE-DNNs) can solve this problem. EE-DNNs have side branches inserted into their middle layers to classify images earlier, at the edge, and thus avoid sending them to the cloud. Besides, multiple side branches can compose an ensemble that collectively produces a more accurate inference. The results, in terms of accuracy, show that EE-DNNs and the ensemble are as sensitive as conventional DNNs. However, given the edge usage, these approaches can reduce the inference time and the number of operations to classify images.

References

Desislavov, R., Martínez-Plumed, F. e Hernández-Orallo, J. (2021). Compute and energy consumption trends in deep learning inference. arXiv preprint arXiv:2109.05472.

Dodge, S. e Karam, L. (2016). Understanding how image quality affects deep neural networks. Em IEEE International Conference on Quality of Multimedia Experience (QoMEX), p. 1-6.

Dodge, S. F. e Karam, L. J. (2018). Quality robust mixtures of deep neural networks. IEEE Transactions on Image Processing, 27(11):5553-5562.

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

Griffin, G., Holub, A. e Perona, P. (2007). Caltech-256 object category dataset. Relatório técnico, California Institute of Technology.

Howard, A. G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M. e Adam, H. (2017). Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861.

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

Li, E., Zeng, L., Zhou, Z. e 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, Z., Lan, G., Stojkovic, J., Zhang, Y., Joe-Wong, C. e Gorlatova, M. (2020). Collabar: Edge-assisted collaborative image recognition for mobile augmented reality. Em ACM/IEEE International Conference on Information Processing in Sensor Networks (IPSN), p. 301-312.

Pacheco, R. e Couto, R. (2020). Introduzindo a qualidade de imagem como uma nova condição de particionamento de dnn na borda. Em Anais do XXXVIII Simpósio Brasileiro de Redes de Computadores e Sistemas Distribuídos, p. 43-56, Porto Alegre, RS, Brasil. SBC.

Pacheco, R., Couto, R. e Simeone, O. (2021a). Calibration-aided edge inference offloading via adaptive model partitioning of deep neural networks. Em IEEE Int. Conf. Communications (ICC), p. 1-6.

Pacheco, R. G., Bochie, K., Gilbert, M. S., Couto, R. S. e Campista, M. E. M. (2021b). Towards edge computing using early-exit convolutional neural networks. Information, 12(10):431.

Pacheco, R. G., Oliveira, F. D. e Couto, R. S. (2021c). Early-exit deep neural networks for distorted images: providing an efficient edge offloading. Em IEEE Global Communications Conference (GLOBECOM), p. 1-6.

Qendro, L., Campbell, A., Lio, P. e Mascolo, C. (2021). Early exit ensembles for uncertainty quantification. Em Roy, S., Pfohl, S., Rocheteau, E., Tadesse, G. A., Oala, L., Falck, F., Zhou, Y., Shen, L., Zamzmi, G., Mugambi, P., Zirikly, A., McDermott, M. B. A. e Alsentzer, E., editors, Proceedings of Machine Learning for Health, volume 158 of Proceedings of Machine Learning Research, p. 181-195. PMLR.

Rozsa, A., Gunther, M. e Boult, T. E. (2016). Towards robust deep neural networks with bang. arXiv preprint arXiv:1612.00138.

Secci, F. e Ceccarelli, A. (2020). On failures of RGB cameras and their effects in autonomous driving applications. Em IEEE International Symposium on Software Reliability Engineering (ISSRE), p. 13-24.

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

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