Enhancing Interpretability in Autonomous Driving Models Through the Use of Concept Bottleneck Layers

Newer and more complex neural networks are being developed each day, these systems continue to grow more intelligent and better at doing specific tasks. Yet, ever since the introduction of deep neural networks (DNN), the black-box nature of such models makes it harder to understand or justify the actions predicted. One of the key limitations of autonomous driving systems is the absence of reasoning that users can easily interpret. This lack of clarity becomes especially problematic in high-stakes, safety-critical scenarios such as when the vehicle encounters difficult road conditions. In these scenarios, understanding how and why specific actions were taken could reassure the driver not to worry. This thesis explores the application of concept bottleneck layers (CBLs) to pre-trained autonomous driving models, enabling the extraction of the reasoning behind each predicted control signal. In addition, it presents a method to generate training data for the CBL with little manual effort. In this framework, vision grounding and segmentation models are used to extract information from each frame. These extractions are integrated to generate concept labels for the dataset used in training the original model. This semi-automatic approach enables frequent concept adjustment without requiring manual data relabeling or recollection. To illustrate the feasibility of such an approach, experiments are conducted on a transformer-based autonomous driving model with a set of extracted concepts for explanation. The results demonstrate that the proposed approach can maintain similar or better accuracy than the original model while providing a certain level of interpretability. The flexible structure of concept generation allows it to handle as many concepts as needed as well as incorporate more complex concepts. This scalability paves the way for future enhancements and refinements.