Boosting Convolutional Neural Networks With Middle Spectrum Grouped Convolution.

Area of Science:

• Computer Vision
• Deep Learning
• Neural Network Architectures

Background:

• Deep convolutional neural networks (DCNNs) are computationally intensive.
• Existing methods like channel pruning and grouped convolution have limitations in balancing efficiency and accuracy.
• A need exists for novel modules that optimize DCNNs.

Purpose of the Study:

• To propose and evaluate a new module, Middle Spectrum Grouped Convolution (MSGC), for enhancing DCNN efficiency.
• To explore the "middle spectrum" between channel pruning and grouped convolution for improved network design.
• To demonstrate MSGC's effectiveness in reducing computational cost while maintaining or improving accuracy.

Main Methods:

• Developed the Middle Spectrum Grouped Convolution (MSGC) module.
• Integrated MSGC into various DCNN backbones (e.g., ResNet, MobileNetV2).
• Evaluated MSGC on image classification (ImageNet) and object detection (MS COCO) datasets.

Main Results:

• MSGC reduced multiply-accumulates (MACs) by 50% for ResNet-18/50, increasing Top-1 accuracy by over 1%.
• MSGC achieved a 35% MACs reduction for MobileNetV2 with improved Top-1 accuracy.
• Similar performance gains were observed in object detection tasks on the MS COCO dataset.

Conclusions:

• MSGC effectively reduces computational complexity in DCNNs.
• The proposed module enhances predictive accuracy across various computer vision tasks.
• MSGC offers a powerful and interpretable approach to designing efficient neural networks.