σ_MAD (Eq. 11), the slopes of the < Az > – z_SPec piecewise linear fit (Eq. 9) and the total variation distance between the z_Ph0t0 and z_SPec sample distributions (Eq. 10) as a function of the total number of instances in the training sample for the baseline method and step 1 of our methods. The nearly balanced training subset in steps 2 and 3 is constructed using the original training set without reducing the sample size. The dotted curves show the total variation distances for the simulated z_photo sample distributions that are expected to have no mode collapse. We compare the methods using the SDSS data and the network Net_P.