James Wilson, (University of Edinburgh) Integral Regressor Networks with Applications to Bayesian Optimization 11am Wednesday 13 April, F24 Hicks Building
Abstract: Application of Bayesian methodology to problems of interest is frequently encumbered by a need to evaluate intractable integrals. In the context of Bayesian Optimization, such integrals must be assessed thousands or, even, millions of times during a single run of the optimizer. Coupled with the computational complexity of existing approximation techniques, demand for repeated evaluation of these integrals leads to a significant bottleneck in the optimization pipeline. In addressing this issue, we propose to pay upfront by training models to predict integral values as a function of their defining values such that, at test time, we may cheaply obtain high-quality estimates for integrals of the same form.