New Method Developed to Explain Decision-Making Mechanisms of Autonomous Vehicles
Researchers have announced the CW-Net system, which aims to increase human trust by translating the complex internal logic of driverless cars into understandable concepts.
Developed by researchers at MIT and Motional, a new method called Concept-Wrapper Network (CW-Net) converts the complex decision-making processes behind deep learning models of autonomous vehicles into understandable concepts, offering humans the ability to anticipate vehicle errors in advance.
Decision-Making Process in Autonomous Vehicles
Driverless cars are typically controlled by deep learning models, and these models occasionally fail in unexpected situations.
This newly developed technology provides clear explanations regarding the decisions of the underlying model, enabling human participants to better grasp the situation.
Working Principle of the CW-Net System
This method, called Concept-Wrapper Network, translates the internal logic processes of deep learning models into concepts that faithfully describe driving performance without altering it.
The system explains the decisions made through easily understandable concepts such as approaching a stationary vehicle or being close to a cyclist.
Driving Safety and Test Results
Through road tests and simulation studies conducted, it was observed that the explanations provided by this system offer great benefits to safety drivers and non-expert users.
The findings revealed that users can predict vehicle behaviors much more accurately and situational awareness is enhanced.