Generalizing Pooling Functions in CNNs: Mixed, Gated, and Tree

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Pooling operations are fundamental in deep neural network architectures.
• Current max and average pooling methods have limitations in adapting to complex patterns.

Purpose of the Study:

• To generalize pooling operations in deep neural networks.
• To enable pooling functions to learn and adapt to variable patterns.
• To improve the performance and invariance properties of deep learning models.

Main Methods:

• Learning pooling functions by combining max and average pooling strategies.
• Developing tree-structured fusion of learned pooling filters.
• Experimental evaluation on benchmark datasets.

Main Results:

• All explored generalized pooling operations outperformed conventional max and average pooling.
• Proposed methods demonstrated enhanced invariance properties.
• Achieved state-of-the-art performance on several benchmark datasets.

Conclusions:

• Generalized pooling operations offer significant performance improvements with minimal computational cost.
• These methods are easily implementable and applicable to various deep neural network architectures.
• The study provides insights into learned pooling masks and feature response embeddings.