A Novel Memory-Scheduling Strategy for Large Convolutional Neural Network on Memory-Limited Devices.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning, particularly convolutional neural networks (CNNs), is widely used but memory-intensive.
• Deployment of CNNs on memory-limited platforms like mobile devices is challenging.
• Existing methods struggle with the high memory footprint of deep learning models.

Purpose of the Study:

• To present a novel memory-management strategy, mmCNN, for efficient deep learning deployment.
• To enable the use of large CNN models on resource-constrained hardware.
• To overcome the memory bottleneck in mobile and wearable deep learning applications.

Main Methods:

• Developed a new memory-management strategy named mmCNN.
• Designed to facilitate the deployment of trained large-size CNNs.
• Tested on GPU platforms with extremely small memory sizes (down to 5MB).

Main Results:

• Achieved over 98% memory savings compared to traditional CNN algorithms.
• Demonstrated over 90% memory savings compared to state-of-the-art vDNNs.
• Successfully ran feed-forward CNN processes in severely memory-limited environments.

Conclusions:

• mmCNN effectively breaks the memory bottleneck for deep learning on memory-limited devices.
• The proposed strategy significantly improves the computing scalability of lightweight applications.
• Enables wider adoption of powerful deep learning models on mobile and wearable technology.