In 850 processor 590 optionally computes a mode-collapse reconstruction error for a second group of training features, for example group of training features 513. Optionally, processor 590 provides group of training features 513 and group of training features 512 to prediction model 510C to produce another subtraction group of features. Optionally, processor 590 provides the other subtraction group of features and the intersection group of features to prediction model 510B to produce another union group of features. Optionally, processor 590 applies a mean square error method to the other union group of features and group of training features 513 to produce another mode-collapse reconstruction error score. Optionally processor 590 uses the other mode-collapse reconstruction error score when computing the loss score, for example by adding the other mode-collapse reconstruction error score to mode-collapse reconstruction error score.

Reference is now made again to FIG. 6. Optionally, in 620 processor 590 modifies one or more model values of at least one of the plurality of set-operator prediction models to reduce another loss score. Optionally processor 590 repeats 601, 605, 608, 612, and 620 in each of a plurality of iterations. Optionally the other loss score is computed in another iteration of the plurality of iterations.