Statistical Learning of Knowledge from Sequence Data

In this seminar Nick Harley will describe the theoretical background of his team’s approach and two different application contexts. The formal basis is a predictive modelling framework called IDyOM (Pearce, 2005) which finds statistical patterns in multi-dimensional sequence data. Information theoretic measures are used to segment sequences into perceptual or cognitive units which then form the basis of a knowledge representation. One application of this method is in modelling music cognition, which involves the real-time organisation of low-level stimuli, such as notes, into high-level structures, such as chords and phrases. A second application is in reinforcement learning. Partial observability, dynamic environments, hidden rewards and explainability are problems for RL models. Supervised learning incorporating knowledge about sub-tasks can mitigate these problems. Harley will describe how statistical modelling of agent behaviour sequences can be used to learn sub-task knowledge in an unsupervised manner. He will finish with a discussion of the open challenges we are addressing and invite the audience to participate in a discussion.