Neural net calibration
This paper describes a straightforward approach (Platt/Temperature scaling) to ``calibrating'' neural networks so that the output scores correspond more closely to probabilities/confidences:
Would be useful to have something like this included in Ketos one day