They must be restricted to linear models, right? So this is basically an implementation of logistic regression? Doing nonlinear operations optically is much trickier, and almost definitely active instead of passive.

EDIT: Actually they describe some nonlinear material inclusions in the paper, interesting.