A Collaborative On-Device CNN Execution Considering Model Parallelism for Latency-Critical Applications
(2023) 2023 IEEE 34th Annual International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC) (5.September.2023)
Files
ACollaborativeOn-DeviceCNNExecutionConsideringModelParallelismforLatency-CriticalApplications.pdf
Open Access - Adobe PDF
- 800.93 KB
Details
- Authors
Kilcioglu, Emre 
Stupia, Ivan
- Abstract
- The use of deep learning (DL) has become more common in applications that require high accuracy, but the high computational demand of DL makes it challenging to execute on a single resource-constrained end device (ED) due to low computing capability and limited energy. One way to execute DL is to use the mobile cloud computing concept, but it results in intolerable latency, network congestion and raises privacy concerns. In the absence of powerful servers, another solution could be to form a collaboration among multiple EDs to execute DL computation. This paper proposes a distributed collaborative on-device convolutional neural network (CNN) execution scheme using model parallelism for latency-critical applications. In the proposed scheme, the ED that owns the input data uses its communication capabilities to get additional computing capabilities from other EDs. A convex optimization problem is formulated to minimize the energy consumption of all EDs by jointly optimizing the communication and computation parameters along with the number of filters assignment in each convolutional layer in the CNN model to multiple EDs. The problem is then decomposed into two sub-problems to obtain analytical expressions for the optimizing parameters and make the optimization distributed among multiple EDs. The simulation results show the importance of optimizing the communication-computation trade-off and the advantages of collaborative computing.
- Affiliations
Citations
Kilcioglu, E., Stupia, I., & Vandendorpe, L. (2023). A Collaborative On-Device CNN Execution Considering Model Parallelism for Latency-Critical Applications. 2023 IEEE 34th Annual International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC), pp. 1-7. https://doi.org/10.1109/PIMRC56721.2023.10293767