There have certainly been some spectacular improvements in machine learning over the last couple of years, and one has to wonder, what comes next? The speaker will talk about recent breakthroughs but also focus on their intrinsic limitations in order to make some guesses about where the frontiers might lie. For example, the current paradigm of supervised learning is an important advance – but would unsupervised learning be more interesting if we could make it work? Neural nets have become much better at modelling some aspects of complex temporal data such as human language – but what about the aspects they’re ill-disposed to learn? Traditional neural networks learn fixed mappings from inputs to outputs – what if they could learn to implement the actual algorithms themselves?