ChannelNets: Compact and Efficient Convolutional Neural Networks via Channel-Wise Convolutions

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Convolutional neural networks (CNNs) are powerful AI tools but face challenges due to large model sizes in resource-limited environments.
• Existing CNNs for mobile devices often struggle with parameter count and computational demands.

Purpose of the Study:

• To propose a novel method for compressing deep convolutional neural networks (CNNs).
• To introduce light-weight CNNs, termed ChannelNets, that are efficient for resource-constrained applications.

Main Methods:

• Introduced channel-wise convolutions to replace dense connections with sparse ones in CNNs.
• Developed ChannelNets utilizing group channel-wise convolutions, depth-wise separable channel-wise convolutions, and a novel convolutional classification layer.
• Investigated and analyzed the performance of the proposed convolutional classification layer, including versions without weight-sharing.

Main Results:

• ChannelNets achieved significant reductions in parameters and computational cost compared to existing mobile-optimized CNNs.
• The proposed convolutional classification layer effectively compresses the classification layer, a major parameter contributor.
• Experimental results on ImageNet demonstrated superior performance of ChannelNets over prior methods.

Conclusions:

• ChannelNets offer an effective approach to compress deep models for efficient deployment in resource-limited scenarios.
• The novel channel-wise convolutions and classification layers provide a promising direction for developing highly efficient CNNs.
• ChannelNets achieve a strong balance between model efficiency and predictive accuracy.