Jakub Kominiarczuk

Adaptive likelihood evaluation

The electronic revolution introduced novel ways of collecting data. Massive amounts of observations can be made, albeit without guarantees of precision: measurement devices may be cheap but inaccurate, responses to Internet-based questionnaires both uncertain and inherently biased. Bayesian methods -- especially Monte Carlo simulations -- still allow for performing statistical inference, but require vast amounts of data. This quickly makes the statistical models intractable, as every Monte Carlo step requires the evaluation of the likelihood at a cost proportional to the number of observations O(n). In the present talk we describe a novel method that makes the cost of evaluating the likelihood sub-linear in n, even for problems with highly non-Gaussian posteriors, assuming the observations are not high-dimensional.

Joint work with Christophe Andrieu