Explainable deep learning improves human mental models of self-driving cars

The study introduces the Concept-Wrapper Network (CW-Net), a novel approach for enhancing the interpretability of deep learning-based planners in self-driving cars by grounding their reasoning in human-understandable concepts. Implemented in a real-world self-driving vehicle, CW-Net improved human drivers' mental models, enabling better predictions of vehicle behavior in unexpected scenarios. This advancement in explainable AI not only enhances safety in autonomous driving but also has potential applications in other safety-critical systems and various AI architectures.