Modular Learning: Agile Development of Robust Traffic Sign Recognition

Autonomous driving is an important research domain with great impact for the future traffic communication. The trend of self-driving car will affect the design of vehicles that incorporates with numerous intelligent modules (e.g., traffic sign recognition). In contrast to a typical AI task (e.g., digit recognition), self-driving car will have countless situations in the real world. Therefore, the self-driving car system requires a long-term development process or even endless maintenance activities. We have to address the newly discovered image corruption rapidly for a robust recognition system. Hence, we proposed our agile development framework for AI system, which is inspired from agile software development paradigm. Our agile framework of AI system aims to speed up the development cycle for each newly discovered image corruption. We denote the training time of each incremental development cycle as the marginal cost of AI system development. We proposed an agile development paradigm called modular learning that incorporates with knowledge distillation to reduce the marginal cost. The acceleration of our method is up to $49\times$ while the recognition accuracy degradation is around 1%. We found that the studies for long-term AI system development are rarely addressed in the literature. We expect our preliminary and promising results can inspire more efforts in this direction.