Maggie Xinyue Hao

Video understanding is a fundamental ability required for artificially intelligent agents, including robots, self-driving cars, or interactive systems. Due to video's scale and temporal nature, existing techniques usually suffer from high complexity, expensive computation, and low accuracy. In this PhD, we will tackle both things at once by using attention. Learning to attend to the right parts of a video will help discard some of the least important parts, which will make processing faster. Learning to attend will likely also yield more accurate results, as it will discard information that can potentially be noisy. One of the main applications where we will test our findings is action recognition.