Ontheotherhand,wefindthatstrong regularization is not necessary for learning high-quality feature representations.

Thus a gap has arisen between theory and practice.

Pre-trained DNN!!-1Wavelet transform!X"#DNN attributions for wavelet coefficients"$

In 1weintroduce ourmotivations, ourmain results and discuss the related work.

Specifically,at each iteration the black-box function is evaluated at one or more input locations that maximizes anacquisition function onthemodel.

The above problem requires the knowledge of the distributionπ which is typically unknown in practice.

The WBP problem is related to the optimal transport (OT) problem, in that both are based on the Wassersteindistance, buttheWBP issignificantly harder.