Authors: Hyelin Nam, Sihun Baek, Sujin Kook, Jihong Park, Seong-Lyun Kim, and Mehdi Bennis
This work will be published as a session of a SLDML survey paper by MIT workshop group.
To showcase the effectiveness of split learning (SL) in real implementation, this section studies a smart brewing factory use case. The scenario under study consists of camera equipped edge devices that are operated in a brewing factory and are connected to a server located 400km away from the factory. We apply SL to these devices and the server, and aim to collectively train distributed neural network models. Based on the images captured by edge devices, the task is to detect bottles and classify their different brands. We compare the performance of sequential SL[1], parallel SL[2], and SplitFed[3], in terms of accuracy and latency with object detection model. Our experiments corroborate that SL is effective in the smart factory application, while revealing several opportunities and challenges such as the sensitivity of SL to data distributions over classes.
Hardware configuration
Edge device: Multiple client devices attached at conveyor belts in Daesun factory as clients. They should protect their own data privacy by processing dataset by lower layer chunks stored in each of the computing devices.
Cloud server: A big computing server in Yonsei University stores the remaining bigger segment of the model.
Network architecture
In SL algorithm, the clients send massive smashed data to the server, and the server transmits back the gradient of the cut layer of the model on the client side. For a reliable communication channel, the server and edge devices in our testbed are conneted through the Korea advanced research network (KOREN), and we also examined with 5G network.
Dataset Acquisition
Our dataset is generated through the cameras and sensors in our testbed. It is important to collect high-quality image from fast-moving conveyor belt, therefore we use four cameras per one image and set camera parameters deliberately. There are 10 classes or brands of bottles to classify.
Reference
[1] O. Gupta and R. Raskar, “Distributed learning of deep neural network over multiple agents,” Journal of Network and Computer Applications, vol.
116, pp. 1–8, 2018.
[2] P. Joshi, C. Thapa, S. Camtepe, M. Hasanuzzaman, T. Scully, and H. Afli, “Splitfed learning without client-side synchronization:
Analyzing client-side split network portion size to overall performance,” CoRR, vol. abs/2109.09246, 2021. [Online]. Available:
https://arxiv.org/abs/2109.09246
[3] C. Thapa, M. A. P. Chamikara, S. Camtepe, and L. Sun, “Splitfed: When federated learning meets split learning,” arXiv preprint arXiv:2004.12088,
2020.
Comments