[Help Wanted] visualizing inducing points

[andersdot]

After training a variational model with inducing points, is it possible to access the field values of the inducing points?

Cheers

[andersdot]

Or is it as simple as predicting the mean, variance at the inducing point locations?

[gpleiss]

Predicting the mean/variance at the inducing point locations is probably the best method. The variational parameters are typically stored in their whitened parameterization (i.e. K^{-1/2} m, K^{-1/2} S K^{-1/2}), so it's probably easier to access them via prediction.

[mbelalsh]

Hello there. How do you access inducing points via prediction? My understanding says that when we pass inducing points to the model with custom initialization, the model learns the inducing points. The learned-inducing points might be different from the ones that I passed initially. Right? If so, then how do I see the new learned inducing points? In the SVGP example notebook, the first 500 examples train_x[ :500, : ] are passed as inducing points to the model. After training, the learned inducing points' locations might be different. Right?

[gpleiss]

Call model.variational_strategy.inducing_points