Balanced Decoupled Spatial Convolution for CNNs.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) are computationally intensive.
• Existing methods for lightweight CNN design often approximate filters.
• Decoupling convolution into spatial and channel dimensions offers a new approach.

Purpose of the Study:

• To propose a novel method for designing lightweight CNNs by decoupling convolution.
• To introduce the Balanced Decoupled Spatial Convolution (BDSC) structure.
• To reduce model size and computational cost while maintaining performance.

Main Methods:

• Decoupled standard convolution into spatial and channel information processing.
• Developed Balanced Decoupled Spatial Convolution (BDSC) for efficient spatial aggregation.
• Incorporated an adaptive spatial configuration using Rectified Linear Units (ReLU).

Main Results:

• BDSC achieves performance comparable to standard convolution with a smaller model size.
• Adaptive spatial configuration improves classification performance without additional cost.
• Experiments on CIFAR-100, CIFAR-10, and ImageNet demonstrate effectiveness.

Conclusions:

• The proposed decoupled view and BDSC structure offer an effective way to design lightweight CNNs.
• BDSC achieves a favorable trade-off between model size and classification accuracy.
• The approach shows potential for further redundancy reduction in channel-domain convolutions.