Knowledge that carries over
Objective
Carry what is learned at one place, time or set of conditions over to measurements elsewhere, wherever the same effects apply.
Scoped Current focus How stages are defined
Building the data needed to test whether what is learned under one processing history carries over to another.
A model can reproduce the data it was learned from and still fail on the next acquisition. Transfer has to be shown on sites, periods and conditions the model has never seen, and its limits have to be known.
This is what allows calibration knowledge to accumulate. Each new measurement becomes easier to interpret because of the ones before it, rather than starting from nothing.
We learn from one set of observations, then predict effects and uncertainty for other acquisitions, other sites and other combinations of conditions that took no part in the learning. The test is a physical prediction on independent data, not a random split of neighbouring pixels.
Predictions on unseen sites and periods improve on starting from nothing, and stated confidence reflects where the learned knowledge applies and where it does not.
Established work this area builds on. Results from the programme are published separately, with their data and limits.
- Learning about physical parameters: the importance of model discrepancy (2014) J. Brynjarsdottir and A. O'Hagan Why a model can predict its own training data well and still transfer badly.
Questions about this work?
If you work on this problem, or hold reference data that bears on it, contact us.