Wednesday 12 October 2022, 6.00PM to 7:00 PM
Speaker(s): Will Smith
Abstract: Machine learning, in particular a specific branch called "deep learning", has had a transformative effect on many areas of computer science as well as science and industry more generally.
However, the vast majority of machine learning methods require very large datasets annotated with labels that are usually provided by humans, at great cost and effort. This sort of "supervised learning" is unlike how humans learn and may not be possible for tasks where labels are difficult to obtain or data is scarce. In addition, the trained model is a black box. We do not know how it works and it does not take advantage of any prior information we may have about the problem being solved.
In this talk, I will provide an accessible introduction to supervised deep learning before describing a recent idea called "self-supervised learning". Here, the data itself provides the supervision without the need for additional labels. This provides a more plausible explanation for how humans are able to learn new skills quickly and with very little supervision. It also provides a promising route for autonomous systems to learn useful tasks simply by observing the world.
I will mainly focus on applications and interesting problems in the area of computer vision, i.e. extracting information from images and videos. In particular, I will describe my own research on the problem of "inverse rendering". This seeks to invert the process of how an image is formed by light reflecting from surfaces towards a camera and has many applications in scene understanding and content capture.
Location: The Ron Cooke Hub lecture theatre, 133 Deramore Lane, York, YO10 5GE