Human attention guided explainable artificial intelligence for computer vision models.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Explainable artificial intelligence (XAI) is crucial for understanding black-box AI models.
• Developing XAI methods that are both faithful to the model and plausible to users presents a significant challenge.
• Current XAI methods often lack faithfulness when applied to object detection tasks.

Purpose of the Study:

• To investigate if incorporating human attention knowledge into saliency-based XAI can improve plausibility and faithfulness.
• To develop novel XAI methods for object detection models that are guided by human attention.
• To enhance user trust and understanding of AI model decisions.

Main Methods:

• Developed FullGrad-CAM and FullGrad-CAM++ for object-specific explanations in object detection.
• Utilized human attention as an objective measure for explanation plausibility.
• Proposed human attention-guided XAI (HAG-XAI) using trainable activation functions and smoothing kernels.
• Evaluated methods on BDD-100K, MS-COCO, and ImageNet datasets.

Main Results:

• Novel XAI methods achieved higher explanation plausibility using human attention.
• Current XAI methods showed lower faithfulness compared to human attention maps in object detection.
• HAG-XAI enhanced plausibility and faithfulness simultaneously for object detection models.
• HAG-XAI improved user trust and outperformed existing state-of-the-art XAI methods for object detection.

Conclusions:

• Embedding human attention knowledge into XAI methods can significantly improve explanation plausibility and faithfulness.
• HAG-XAI offers a promising approach for developing more trustworthy and understandable AI systems, especially in computer vision.
• The proposed methods demonstrate superior performance in object detection tasks, advancing the field of explainable AI.