Learning rates and schedules are often chosen heuristically to address this trade-off.

The impressive capability of these models, however, often entails significant computational overhead during both training and inference.

The main observation behind our approach is that separation does not require optimization butcan besolvedinclosed-form prior totraining and plugged into a network.

Our approach relies on the principle ofoptimism under uncertaintyto efficiently explore. Our learning algorithms optimizetheiractions withrespect toamodel based ontheempirical statistics, while optimistically overestimating rewards and underestimating the resource consumption (i.e., overestimating the distance from the constraint).

Many recent approaches to learn submodular functions suffer from limited expressiveness.

Explanations forthisphenomenon arecontroversial: Whilemostworksattribute the artifacts to the generator, other works point to the discriminator. We take a sober look at those explanations and provide insights on what makes proposed measures against high-frequency artifacts effective.