Katerina FragkiadakiAssistant Professor

I am interested in building machines that understand the stories that videos portray, and, inversely, in using videos to teach machines about the world. The pen-ultimate goal is to build a machine that understands movie plots, and the ultimate goal is to build a machine that would want to watch Bergman over this.
Teaching machines to appreciate Bergman is a hard problem, in fact, most people fail in doing so. Thus, it is a very long term goal of our research. Meanwhile, we are working on teaching machines basic common sense. Our recent works use recurrent nets with 3D representation bottlenecks that disentangle camera motion from appearance, while being optimized end-to-end for a final task, such as view prediction or 3D object detection. In this way, our models learn object permanence, size constancy, occlusions and dis-occlusion relationships, useful for 3D object detection, affordance learning and language grounding.