Abstract. Statistical modelling is focused on parametrising the variations in a set of data using some form of latent structure. In sequential or active learning we then use the "ignorance" parametrised within the model to quantify the value the possible data we can observe. If we have an associated task, or a specific "quantity of interest" that we are interested in to learn from the data, this raises the question regarding what of our "ignorance" we want to reduce during the learning process. For the task of integration all functions with the same integral and in optimisation all functions with the same extremum are equivivalent in terms of the quantity of interest. In effect, only reduction of ignorance in our latent variables that are reflected in the quantity of interest are relevant. This characteristic raises a new set of interesting modelling questions, especially considering that the process of data acquisition is a direct consequence of our model.

In this talk I will describe a simple surrogate model for Bayesian Optimisation that allows us to think about the relevance of an observation for the task of finding the extremum of the function. Importantly the relevance changes during the data acquisition process, observations that where important initially can become irrelevant or due to model mismatch detrimental for inferring the quantity of interest. We will show that while including this structure leads to a small computational penalty it leads to a much more robust learning process.

Bio. Carl Henrik Ek is an Associate Professor in Machine learning at the University of Cambridge and a fellow at Pembroke College. Together with Prof. Neil Lawrence he leads the ml@cl research group in the department of computer science and technology. Carl Henrik's work is focused on building probabilistic models that allows us to learn efficiently from small amounts of data. Specifically he is interested in tasks related to real physical systems which are characterised by the existence of prior knowledge.

Time and place. Thursday dec 14th, at 14:00 (sharp) in T5 CS/Aalto (and at zoom https://aalto.zoom.us/j/62062416089)