Colin Samplawski
ABSTRACT
We consider deploying an object detection pipeline over a heterogeneous IoT network. We consider a setting where a camera-equipped IoT edge node communicates wirelessly with a cloud server. In many real-world domains, the bandwidth of this connection is constrained and variable, while the edge node may have insufficient compute resources to perform complex machine learning tasks such as object detection. Building on prior work in the image classification space, we propose an approach for detection that first partitions a deep neural network model at a given layer and then applies progressive transmission of intermediate convolutional filter maps. This capability can be exercised in response to dynamically varying network bandwidth. Further, we consider the application of lossy compression to the filter maps themselves, exposing a broader set of communication compression trade-offs that includes a choice for the representation size transmitted as well as the lossy compression level applied. The model is trained specifically to optimize distributed operation including the need for the cloud stage to decode and process a variable size representation. We investigate the performance of this approach in terms of a detection accuracy-communication cost trade-off. We compare to approaches that compress images and perform detection using cloud offload. Our results show that our approach achieves a significantly better detection accuracy-communication cost trade-off compared to cloud offload of JPEG compressed images.
- C. Bucilua, R. Caruana, and A. Niculescu-Mizil. Model compression. In SIGKDD, pages 535--541. ACM, 2006.Google Scholar
Digital Library
- P. Bulić, G. Kojek, and A. Biasizzo. Data transmission efficiency in bluetooth low energy versions. Sensors, 19(17): 3746, 2019.Google ScholarCross Ref
- M. Chen, A. Zheng, and K. Weinberger. Fast image tagging. In ICML, pages 1274--1282, 2013.Google ScholarDigital Library
- B.-G. Chun and P. Maniatis. Augmented smartphone applications through clone cloud execution. In Proceedings of the 12th conference on Hot topics in operating systems, HotOS'09, pages 8--8, Berkeley, CA, USA, 2009. USENIX Association.Google ScholarDigital Library
- J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In NAACL-HLT, 2019.Google Scholar
- D. Erhan, C. Szegedy, A. Toshev, and D. Anguelov. Scalable object detection using deep neural networks. In CVPR, pages 2147--2154, 2014.Google ScholarDigital Library
- M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The pascal visual object classes (voc) challenge. IJCV, 88(2): 303--338, June 2010.Google ScholarDigital Library
- X. Gao, Y. Zhao, L. Dudziak, R. Mullins, and X. Cheng-Zhong. Dynamic channel pruning: Feature boosting and suppression. In ICLR, 2019.Google Scholar
- R. Girshick. Fast r-cnn. In ICCV, pages 1440--1448, 2015.Google ScholarDigital Library
- R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In CVPR, pages 580--587, 2014.Google ScholarDigital Library
- A. Graves, N. Jaitly, and A.-r. Mohamed. Hybrid speech recognition with deep bidirectional lstm. In Automatic Speech Recognition and Understanding (ASRU), 2013 IEEE Workshop on, pages 273--278. IEEE, 2013.Google ScholarCross Ref
- A. Graves, A.-r. Mohamed, and G. Hinton. Speech recognition with deep recurrent neural networks. In Acoustics, speech and signal processing (icassp), 2013 ieee international conference on, pages 6645--6649. IEEE, 2013.Google Scholar
- K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In CVPR, pages 770--778, 2016.Google ScholarCross Ref
- Y. He, J. Lin, Z. Liu, H. Wang, L.-J. Li, and S. Han. Amc: Automl for model compression and acceleration on mobile devices. In ECCV, pages 784--800, 2018.Google ScholarCross Ref
- J. Huang, C. Samplawski, D. Ganesan, B. Marlin, and H. Kwon. Clio: Enabling automatic compilation of deep learning pipelines across iot and cloud. In MobiCom, 2020.Google ScholarDigital Library
- Y. Kang, J. Hauswald, C. Gao, A. Rovinski, T. Mudge, J. Mars, and L. Tang. Neurosurgeon: Collaborative intelligence between the cloud and mobile edge. In ACM SIGARCH, pages 615--629. ACM, 2017.Google ScholarDigital Library
- S. Kosta, A. Aucinas, P. Hui, R. Mortier, and X. Zhang. Thinkair: Dynamic resource allocation and parallel execution in the cloud for mobile code offloading. In Infocom, 2012 Proceedings IEEE, pages 945--953. IEEE, 2012.Google ScholarCross Ref
- A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. In NIPS, pages 1097--1105, 2012.Google ScholarDigital Library
- J. Lin, Y. Rao, J. Lu, and J. Zhou. Runtime Neural Pruning. In NIPS, 2017.Google ScholarDigital Library
- S. Liu, Y. Lin, Z. Zhou, K. Nan, H. Liu, and J. Du. On-demand deep model compression for mobile devices: A usage-driven model selection framework. In Conference on Mobile Systems, Applications, and Services, 2018.Google ScholarDigital Library
- W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y. Fu, and A. C. Berg. Ssd: Single shot multibox detector. In ECCV, pages 21--37, 2016.Google ScholarCross Ref
- J. Redmon, S. Divvala, R. Girshick, and A. Farhadi. You only look once: Unified, real-time object detection. In CVPR, pages 779--788, 2016.Google ScholarCross Ref
- M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L.-C. Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, pages 4510--4520, 2018.Google ScholarCross Ref
- M. Satyanarayanan. A brief history of cloud offload: A personal journey from odyssey through cyber foraging to cloudlets. GetMobile: Mobile Computing and Communications, 18(4): 19--23, 2015.Google ScholarDigital Library
- N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. JMLR, pages 1929--1958, 2014.Google Scholar
- C. Szegedy, A. Toshev, and D. Erhan. Deep neural networks for object detection. In NIPS, pages 2553--2561, 2013.Google ScholarDigital Library
- https://greenwaves-technologies.com/gap8-product/. GAP8: Ultra-low power, always-on processor for embedded artificial intelligence.Google Scholar
- P. Viola, M.Jones, et al. Robust real-time object detection. IJCV, 4(34--47): 4, 2001.Google Scholar
Index Terms
- Towards Objection Detection Under IoT Resource Constraints: Combining Partitioning, Slicing and Compression
Recommendations
Towards the Decentralised Cloud: Survey on Approaches and Challenges for Mobile, Ad hoc, and Edge Computing
Cloud computing emerged as a centralised paradigm that made “infinite” computing resources available on demand. Nevertheless, the ever-increasing computing capacities present on smart connected things and devices calls for the decentralisation of Cloud ...
Resource provisioning with budget constraints for adaptive applications in cloud environments
HPDC '10: Proceedings of the 19th ACM International Symposium on High Performance Distributed ComputingUtility computing was a vision stated more than 40 years ago. It refers to the idea that computing resources and services can be delivered, utilized, and paid for as utilities such as water or electricity. The recent emergence of cloud computing is ...
A Pattern for Fog Computing
VikingPLoP '16: Proceedings of the 10th Travelling Conference on Pattern Languages of ProgramsFog Computing is a new variety of the cloud computing paradigm that brings virtualized cloud services to the edge of the network to control the devices in the IoT. We present a pattern for fog computing which describes its architecture, including its ...
Comments