Self-grouping convolutional neural networks.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Group convolution enhances efficiency and reduces parameters in deep neural networks.
• Existing methods use fixed filter partitioning, limiting potential.
• Data-independent grouping fails to fully exploit group convolution benefits.

Purpose of the Study:

• Propose a novel Self-Grouping Convolutional Neural Network (SG-CNN) method.
• Develop a data-dependent filter grouping strategy for improved performance.
• Enhance computational efficiency and recognition accuracy in deep learning models.

Main Methods:

• Filters are grouped based on the similarity of their input channel importance vectors.
• Importance vectors are identified by evaluating input channel importance for each filter.
• Clustering is used to group these vectors, followed by pruning less important connections.
• Two fine-tuning schemes (local/global and global only) are employed to recover network capacity.

Main Results:

• SG-CNN adapts to various architectures like ResNet and DenseNet.
• Achieves superior performance in compression ratio, speedup, and recognition accuracy.
• Demonstrates generalization ability through transfer learning tasks including domain adaptation and object detection.

Conclusions:

• Self-grouping convolution offers a more effective approach to filter partitioning.
• SG-CNN provides significant improvements in efficiency and accuracy over existing methods.
• The method shows strong adaptability and generalization capabilities for diverse deep learning applications.