Abstract

Machine learning with deep neural networks (nowadays commonly referred to as “deep learning”) has brought about significant benefits across a variety of domains (such as computer vision, natural language processing and game-playing) using essentially the same kind of learning algorithm. However, in essence, the bulk of what we know about deep learning right now has been established in the past century, the innovations made by large institutions are often hyperbolised by the popular media (for an excellent overview of related issues, I highly recommend the following article: http://approximatelycorrect.com/2017/03/28/the-ai-misinformation-epidemic/), and the research landscape is difficult to navigate even for experts, given the immense volume of papers consistently being released in the area over the past few years.

In this talk, I will channel the experience I’ve acquired in this area over the past few years to provide a concise, formal and (hopefully!) exciting overview of where the field is right now, how it has gotten there, and what are the likely next steps (with particular emphasis on the kinds of problems where this technique still notably struggles). This will be an applications-oriented talk, accompanied by a complementary overview of fundamental theory. Therefore, the talk is expected to be appropriate for people at all skill levels, but it is expected that it will be mostly valuable for those that only vaguely interacted with deep learning thus far. An entry-level grasp of linear algebra and probability will also be useful, but not essential.