Extraction of an Explanatory Graph to Interpret a CNN

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Convolutional Neural Networks (CNNs) are powerful tools for image analysis.
• Filters in CNNs often encode complex mixtures of object parts, hindering interpretability.
• Existing methods lack effective ways to disentangle and utilize these learned part representations.

Purpose of the Study:

• To introduce an explanatory graph representation for revealing object parts within CNN convolutional layers.
• To develop a method for automatically disentangling object parts from filters without requiring part annotations.
• To enhance the transferability and utility of CNN features for downstream tasks like part localization.

Main Methods:

• Learned an explanatory graph by mining neural activations from CNN filter feature maps.
• Organized mined object parts as nodes in the graph, with edges representing co-activation and stable spatial relationships.
• Utilized the explanatory graph to transfer learned object part features to a part localization task.

Main Results:

• Each graph node consistently represented the same object part across different images.
• The method significantly boosted the transferability of CNN features.
• The explanatory graph approach outperformed existing methods in the task of part localization.

Conclusions:

• The proposed explanatory graph representation effectively reveals and organizes object parts encoded in CNN filters.
• This approach enhances CNN feature transferability and significantly improves part localization performance.
• The method offers a promising direction for interpretable and transferable deep learning models.