Machine learning has been quitesuccessful atgenerativemodeling ofcomplexdomains such asimages, audio, and text.

Neural Information Processing Systems http://nips.cc/

Post-hoc 8] for reject, whereincexp(x, y)= c0. AR -10 (left), CIFAR -100 (middle), and ImageNet (right) inalearningtodefersetting, isallowedtodefertoa "specialist " expert.

We provide theoretical insights that support our method's alignment with standard preference alignment objectives.

Here, we narrow this gap by developing aneffectivemethod fortraining acanonical model ofcortical neural circuits, the stabilized supralinear network (SSN), that in previous work had to beconstructed manually ortrainedwithundueconstraints.