An XAI method for convolutional neural networks in self-driving cars

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Machine learning models, particularly deep learning, often lack transparency, making their predictions difficult to interpret.
• This lack of interpretability poses significant risks in reliability-critical applications such as self-driving cars, where errors can have severe consequences.
• Explainable Artificial Intelligence (XAI) is emerging as a crucial field to address these transparency and reliability concerns.

Purpose of the Study:

• To propose and evaluate a novel XAI method for convolutional neural networks (CNNs) applied to self-driving car systems.
• To enhance the reliability and trustworthiness of AI decision-making in safety-critical autonomous driving scenarios.
• To accurately identify and visualize the image regions most influential in an AI model's classification decisions.

Main Methods:

• The proposed XAI method involves comparing output values from the last hidden layer of CNNs using original and modified input images.
• Sensitivity Analysis is employed to determine the influence of different image parts on the model's output.
• Images are segmented, and specific parts are obscured to compute their impact on the network's hidden layer representations, visualized through shaded regions.

Main Results:

• The XAI approach accurately identifies critical image regions essential for category distinction in the context of self-driving cars.
• The method effectively visualizes the decision-making process by highlighting salient features within input images.
• Experimental results validate the effectiveness of the proposed technique in explaining CNN predictions for autonomous driving applications.

Conclusions:

• The developed XAI method provides a reliable way to understand and verify the decision-making processes of AI in self-driving cars.
• This technique contributes to building more trustworthy and transparent AI systems for safety-critical applications.
• Further adoption of such XAI methods is essential for the safe and widespread deployment of autonomous vehicles.