Search code examples
gpflow

GPflow uncertain inputs

Does GPflow support uncertain inputs during the prediction step? What about uncertain inputs during the training step? I looked around and it seems there is no information on that.

Thanks.

Solution

  • From https://github.com/GPflow/GPflow/issues/893#issuecomment-447157109:

    Currently, there is a good deal of back-end code that makes this easy, but the front-end functionality is not there yet. conditionals.py:uncertain_conditional() is designed for exactly this. It calculates the moment matched means and variances for Gaussian inputs, which is also what is needed for GPLVMs and variational uncertain input models. To train with this, one needs to implement a model using this conditional instead of the normal one, and similarly for testing. Hope this helps.