Semantic Interpretation for Convolutional Neural Networks: What Makes a Cat a Cat?

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Deep neural networks, particularly Convolutional Neural Networks (CNNs), are powerful but often function as "black boxes."
• Existing interpretability methods face limitations in extracting understandable semantic spaces.
• Understanding the internal workings of CNNs is crucial for trust and further development.

Purpose of the Study:

• To introduce a novel framework, Semantic Explainable Artificial Intelligence (S-XAI), for enhancing the interpretability of deep neural networks.
• To address the challenge of extracting understandable semantic spaces from CNNs.
• To provide statistical interpretation and propose the concept of semantic probability.

Main Methods:

• Developed a framework named Semantic Explainable Artificial Intelligence (S-XAI).
• Utilized a sample compression method based on distinctive row-centered Principal Component Analysis (PCA).
• Extracted semantic spaces using identified semantically sensitive neurons and visualization techniques.

Main Results:

• S-XAI effectively extracts understandable semantic spaces from CNNs.
• The framework provides statistical interpretation of the semantic space.
• Proposed and utilized the concept of semantic probability for analysis.

Conclusions:

• S-XAI offers an effective method for semantic interpretation of CNNs.
• The framework has broad applications, including trustworthiness assessment and semantic sample searching.
• This approach advances the field of explainable artificial intelligence by providing deeper insights into neural network behavior.