Revisiting convolutional design for efficient CNN architectures in edge-aware applications.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Vision Transformers (ViTs) show promise but have high computational costs.
• Convolutional Neural Networks (CNNs) are efficient for real-time and edge AI.
• Optimizing CNNs for edge deployment is crucial for resource-constrained applications.

Purpose of the Study:

• To evaluate the performance of five different convolutional operations within the ResNet-50 architecture.
• To analyze the trade-offs between accuracy, inference time, and power consumption of these operations on edge AI platforms.
• To provide insights for designing hardware-aware CNNs for edge AI.

Main Methods:

• Integrated standard 2D spatial, grouped, shuffle, depthwise separable, and shift convolutions into ResNet-50.
• Trained models on Tiny-ImageNet and CIFAR-10/100 datasets.
• Evaluated performance on Raspberry Pi 5, Coral Dev Board, and Jetson Nano, measuring accuracy, inference time, and power consumption.

Main Results:

• Depthwise separable convolutions, while theoretically efficient, showed increased memory access issues on memory-bound edge platforms.
• Shuffle and shift convolutions demonstrated superior trade-offs between accuracy, computational load, and inference speed.
• Runtime decomposition revealed platform-specific performance characteristics for each convolution type.

Conclusions:

• Shuffle and shift convolutions are recommended for optimizing CNNs on edge AI devices.
• The findings offer practical guidance for selecting and designing efficient CNN architectures for resource-limited environments.
• Hardware-aware design choices are critical for successful edge AI deployment of computer vision models.