The Electronics group at Otago specialises in combining instrumentation with inference, producing sensors characterized by: "small front-end, big back-end". The big back-end is an inference calculation that interprets measurements in term of physical models, implemented via sample-based Bayesian inference. This approach produces quantified uncertainties (or errors) along with estimates, while optimality of information usage means minimal data size and small front-end sensors. Our inference algorithms have found wider applicability than just the smart sensors we build; we have used them to calibrate large-scale models of geothermal fields, detemine the age of the Indo-European languages, and to perform automated inspection on production lines.
The cost of this approach is the big calculation currently required to implement Bayesian inference. Accordingly, we put significant research effort into developing efficient algorithms for inference. Recent years have seen dramatic advances, largely by recognising that existing inference algorithms are equivalent to deterministic computational methods from the 1960's, and then adapting acceleration methods developed through the 1980's. In some cases the inferential calculation is now actually faster than the deterministic counterpart, signalling changes to the sensors we build, and more generally to computational statistics.