Reference is now made again to FIG. 8. For at least one of the plurality of set-operator prediction models, for example set-operator prediction model 510A, in 820 processor 590 provides to the at least one set-operator prediction model the first group of training features as the second input and the second group of training features as the first input, to produce a second model group of features. Optionally, processor 590 applies a mean square error method to the first model group of features and the second model group of features to produce a symmetric reconstruction error score. Optionally processor 590 uses the symmetric reconstruction error score when computing the loss score, for example by adding the symmetric reconstruction error score to the sum of the plurality of model loss scores.

There is a known risk in training a neural network of reaching a state where the neural network generates a common output for a plurality of input samples. This condition is sometimes called mode collapse. To reduce a risk of mode collapse when training the plurality of set-operator prediction models, processor 590 optionally computes one or more mode-collapse reconstruction errors, each indicative of a difference between an original group of features and a reconstructed group of features computed by applying a union operator to an intersection of the original group of features and another group of features and a subtraction of the other group of features from the original group of features. When A depicts the original group of features, B depicts the other group of features, ∪ depicts a union operator, ∩ depicts an intersection operator, and/depicts a subtraction operator, a mode-collapse reconstruction error score is indicative of a difference between A and ((A∩B)∪(A/B)).