Accelerating On-Device Learning with Layer-Wise Processor Selection Method on Unified Memory.

Area of Science:

• Artificial Intelligence
• Mobile Computing
• Computer Engineering

Background:

• Deep learning models are increasingly deployed on mobile devices with limited computing power.
• Device heterogeneity, stemming from variations in sensors and processors, causes performance degradation in mobile deep learning.
• Training device-specific network models is crucial to overcome performance limitations.

Purpose of the Study:

• To propose an efficient on-device learning acceleration method to mitigate performance degradation caused by device heterogeneity.
• To reduce latency in deep learning model training on mobile devices.
• To alleviate performance disparities across different mobile devices.

Main Methods:

• Utilizing unified memory to minimize data transfer latency during network model training.
• Implementing a layer-wise processor selection strategy to account for latency variations in forward and backpropagation steps.
• Conducting experiments on an ODROID-XU4 using the ResNet-18 model.

Main Results:

• The proposed method achieved latency reductions of up to 28.4% compared to the central processing unit (CPU) and 21.8% compared to the graphics processing unit (GPU).
• Experiments demonstrated that the method effectively alleviates device heterogeneity in on-device learning.
• Average power consumption was measured across various batch sizes, confirming the method's efficiency.

Conclusions:

• The proposed on-device learning acceleration method successfully reduces latency and mitigates device heterogeneity.
• Efficient memory utilization and layer-wise processor selection are key to optimizing deep learning performance on mobile devices.
• This approach enables more robust and efficient deployment of deep learning models on diverse mobile hardware.