Understanding How CNNs Recognize Facial Expressions: A Case Study with LIME and CEM.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Facial expression recognition is a key AI challenge.
• Convolutional Neural Networks (CNNs) excel at image recognition but lack transparency.
• Explainable Artificial Intelligence (XAI) methods are needed to interpret CNN decisions.

Purpose of the Study:

• To compare the effectiveness of LIME and CEM explainable AI techniques for facial expression recognition.
• To determine which XAI method is more suitable for complex image data.

Main Methods:

• Applied LIME (Local Interpretable Model-agnostic Explanations) to highlight image regions contributing to classification.
• Applied CEM (Concept Embedding) to identify features sufficient for classification and those necessary for distinction.
• Compared LIME and CEM on complex facial expression images.

Main Results:

• LIME effectively highlights image areas crucial for CNN classification.
• CEM provides explanations by identifying both sufficient and absent features for classification.
• LIME is more suitable for high-dimensional image data, while CEM is effective for lower-dimensional feature sets.

Conclusions:

• The choice between LIME and CEM depends on the complexity and dimensionality of the image data.
• LIME is recommended for complex facial expression images with numerous features.
• CEM offers a more human-interpretable explanation by considering absent features.