CX-ToM: Counterfactual explanations with theory-of-mind for enhancing human trust in image recognition models

Area of Science:

• Artificial Intelligence
• Machine Learning
• Human-Computer Interaction

Background:

• Current explainable AI (XAI) methods often provide single-shot explanations.
• Attention-based explanations (heat maps) are insufficient for building user trust in deep learning models.
• Explaining decisions of deep convolutional neural networks (CNNs) remains a challenge.

Purpose of the Study:

• To propose CX-ToM (counterfactual explanations with theory-of-mind), a new XAI framework.
• To frame AI explanation as an iterative communication process (dialogue).
• To improve human trust in CNN models through novel explanation techniques.

Main Methods:

• Utilizing Theory of Mind (ToM) to model user and machine perspectives.
• Generating explanations through an iterative dialogue process.
• Employing counterfactual explanations called 'fault-lines' to identify minimal feature changes for altering predictions.

Main Results:

• CX-ToM generates explanations via a dialogue, mediating differences between human and machine understanding.
• Fault-lines identify minimal semantic features (e.g., stripes) to change a CNN's classification.
• CX-ToM significantly outperforms existing state-of-the-art XAI models in experiments.

Conclusions:

• Iterative, dialogue-based explanations using Theory of Mind are more effective than single-shot methods.
• Counterfactual 'fault-lines' provide more insightful explanations than attention maps for CNNs.
• CX-ToM enhances human trust and understanding of AI decisions.