LRPRNet: Lightweight Deep Network by Low-Rank Pointwise Residual Convolution.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models are computationally intensive, limiting deployment on resource-constrained devices like smartphones.
• Graphics Processing Units (GPUs) have driven deep learning advancements, but efficiency remains a challenge for edge computing.

Purpose of the Study:

• To develop a novel lightweight deep learning module to reduce computation and memory costs.
• To enable efficient deployment of deep learning models on multimedia devices and embedded systems.

Main Methods:

• Proposed a lightweight deep learning module named LRPRNet, utilizing low-rank pointwise residual (LRPR) convolution.
• Implemented LRPR by approximating pointwise convolution with a low-rank method and using depthwise convolutions as a residual module.
• Demonstrated the general applicability of LRPR to existing architectures like MobileNetv1, ShuffleNetv2, and MixNet.

Main Results:

• LRPRNet achieved competitive performance on visual recognition tasks, including image classification and face alignment.
• Significant reductions in Floating Point Operations (Flops) and memory costs were observed compared to state-of-the-art lightweight models.
• The proposed LRPR module effectively reduced model size while maintaining accuracy.

Conclusions:

• LRPRNet offers an effective solution for deploying deep learning models on resource-limited devices.
• The LRPR module provides a general and efficient method for lightweight deep learning model design.
• This research contributes to advancing efficient deep learning for edge AI applications.