Learning feature relationships in CNN model via relational embedding convolution layer.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Cognitive Science

Background:

• Convolutional Neural Networks (CNNs) excel at hierarchical feature extraction but neglect feature relationships, limiting interpretability and domain generalization.
• Existing methods to incorporate feature relationships often require external prior knowledge or auxiliary modules, increasing computational and storage costs.
• Human cognition effectively utilizes hierarchical visual attribute relationships, a capability lacking in standard CNNs.

Purpose of the Study:

• To enable CNNs to learn relationships among hierarchical deep features without requiring prior knowledge or increasing computational/storage resources.
• To enhance fundamental CNN performance in areas like interpretability, domain generalization, and robustness.
• To address key challenges in feature relation learning: quantifying connection intensity, identifying useless connections, and updating relation graphs.

Main Methods:

• Defined the task of learning relationships among hierarchical features in CNNs.
• Proposed the Relational Embedding Convolution (RE-Conv) layer for representing feature relationships within convolutional layers.
• Introduced a 'use & disuse' strategy to manage connection intensity, useless connections, and relation graph updates.

Main Results:

• Demonstrated significant improvements in interpretability, domain generalization, noise robustness, and inference efficiency.
• Outperformed state-of-the-art methods in domain generalization tasks.
• Achieved comparable precision to standard CNNs while reducing floating point operations (FLOPs) by approximately 50%.

Conclusions:

• The proposed feature relation learning scheme effectively enhances CNN performance without additional resource consumption.
• The method offers a practical and flexible solution for improving CNNs, particularly in domain generalization.
• Seamless integration with existing methods allows for further performance gains.