The moments of a random variable are arguably the most popular all-purpose statistic.

Weconductaseriesofempirical studies showing that overconfidence may not hurt final calibration performance if post-hoc calibration is allowed, rather, the penalty of confident outputs will compress theroom ofpotential improvement inpost-hoc calibration phase.

Anomaly detection is widely used for identifying critical errors and suspicious behaviors, butcurrent methods lackinterpretability. Weleverage common propertiesofexisting methods andrecent advancesingenerativemodels tointroduce counterfactual explanations for anomaly detection. Givenan input, we generate its counterfactual as a diffusion-based repair that shows what a non-anomalous versionshouldhavelookedlike.

In this work, we find that better model usage can make a huge difference.

Thus we get the searched architecture and optimized parameters in a single run.