Continuous Domain Adaptation

Imagine a self-driving car with a recognition system trained in mostly sunny weather conditions. Gradually, it starts to rain, which produces domain shift that may severely affect the efficacy of the car's recognition model. To enable the agent to adapt to such changes, we propose exploiting the natural continuity between gradually varying domains: by incrementally adapting in sequence from the source to the most similar target domain while simultaneously efficiently regularizing against prior examples, we obtain a single strong model capable of recognition within all observed domains.