Colloquium Computer Science - Prof. P. Tino University of Birmingham

Title: Machine Learning in the Space of Dynamical Systems

Abstract:

Current Machine Learning approaches cannot easily and naturally handle
temporal data in the form of noisy, sparse and irregularly sampled time
series. We suggest a new model-based framework for formulating machine
learning on such data.  The framework is based on the idea of Learning
in the Model space, where each data item (time series) is represented as
the posterior distribution over possible models, given the observations.
The framework is general, but in the talk we will show how to formulate
a classifier operating in the space of posteriors over the models.
Besides being "transparent" (the classifier decisions can be understood
through the baseline inferential model), the framework also allows for a
model space analog of feature extraction - detecting the most important
sub-models relevant for the task at hand.