Study online performance under non-stationary data streams using prequential evaluation

Investigate the behavior and predictive performance of the Bayesian online natural gradient (bong) and related online variational methods in the non-stationary single-stream setting by evaluating one-step-ahead (prequential) log predictive density, and characterize how these methods adapt to distribution shifts over time.

Background

The experiments primarily assume training and test data are drawn from the same static distribution, and performance is evaluated via test-set negative log predictive density. The authors note that in a single-stream scenario with potential non-stationarity, prequential (one-step-ahead) log predictive density is the appropriate evaluation.

They explicitly defer studying this non-stationary case, making it an open direction to assess whether bong and baselines can rapidly adapt to drifting distributions and how their uncertainty estimates behave under such shifts.