Daan Wierstra

Wednesday - 3 October 2018

Time: 4.00pm

Ground Floor Seminar Room

25 Howland Street, London, W1T 4JG

New architectures for sequential reasoning

Classical deep learning approaches should, in principle, be able to perform complex reasoning tasks. Yet in practice, this ability does not just `emerge' by using naive training methods, and said methods do struggle on many tasks involving complicated relational structures. This is in accordance with our intuition that suggests that solving many complex real world tasks necessarily involves multi-step sequential reasoning across interactions between implicitly or explicitly defined entities.

In this talk I will provide an overview of some of our recent research into new models that introduce architectural inductive biases designed to deal with sequential reasoning tasks, yielding state-of-the-art results on various data sets.