Adaptive Adversarial Learning for a Diverse Visual World

Automated visual recognition is in increasingly high demand. However, despite tremendous performance improvement in recent years, state-of-the-art deep visual models learned using large-scale benchmark datasets still fail to generalize to the diverse visual world. In this talk I will discuss a general purpose semi-supervised learning algorithm, domain adversarial learning, which facilitates transfer of information between different visual environments and across different semantic tasks thereby enabling recognition models to generalize to previously unseen worlds. I’ll demonstrate applications of this approach to different visual tasks, such as semantic segmentation in driving scenes and transfer between still image object recognition and video action recognition.