Detachable Second-Order Pooling: Toward High-Performance First-Order Networks.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning Architectures

Background:

• Second-order pooling methods offer superior performance in visual classification compared to first-order methods.
• High computational resource demands of second-order pooling limit its practical deployment.

Purpose of the Study:

• To develop a novel network architecture that combines the benefits of second-order pooling with the efficiency of first-order networks.
• To maintain unchanged model complexity during inference while leveraging enhanced feature representations.

Main Methods:

• Introduced a detachable second-order pooling network architecture.
• Integrated auxiliary branches with second-order pooling at various stages of a convolutional neural network (CNN).
• Utilized auxiliary branches during training to improve feature learning, then detached them for inference.

Main Results:

• The proposed network achieved leading performance on CIFAR-10, CIFAR-100, and ImageNet datasets.
• Demonstrated higher accuracy compared to existing second-order pooling networks.
• Maintained the low inference complexity characteristic of first-order networks.

Conclusions:

• The detachable second-order pooling network effectively balances high accuracy with computational efficiency.
• This approach offers a practical solution for leveraging advanced pooling techniques in resource-constrained environments.
• The method shows significant potential for advancing visual classification tasks.