Explanation strategies in humans versus current explainable artificial intelligence: Insights from image...

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Explainable AI (XAI) methods offer insights into AI decision-making, but their alignment with human explanation processes is not well understood.
• Comparing XAI outputs with human cognitive strategies is crucial for developing more intuitive and effective AI systems.
• Current XAI techniques often lack a clear connection to how humans naturally explain their reasoning.

Purpose of the Study:

• To investigate human attention strategies during image classification and explanation tasks using eye-tracking.
• To compare human attention patterns with saliency-based explanations generated by current XAI methods.
• To identify which XAI explanation types best resemble human explanation strategies.

Main Methods:

• Human participants classified images and explained their classifications.
• Eye-tracking technology monitored participants' visual attention during both tasks.
• Saliency maps from XAI methods were compared with human attention data.

Main Results:

• Humans employed more explorative attention strategies when explaining compared to classifying.
• Two distinct human explanation strategies emerged: focused scanning with conceptual explanations and explorative scanning with visual explanations.
• XAI saliency maps most closely resembled human explorative attention strategies.
• XAI explanations emphasizing observable causality through perturbation aligned better with human strategies than those focusing on internal AI features.

Conclusions:

• Human explanations integrate both visual and conceptual information, serving distinct functions.
• XAI methods that highlight observable causality demonstrate greater similarity to human explanation strategies.
• Aligning XAI with human cognitive processes can enhance user accessibility and understanding of AI.