Schematic overview of the machine learning pipeline. The black arrows denote the order in which properties are derived from each other. The pink boxes show the optimization objective: three CNNs are optimizing the Galaxy Zoo 2 probabilities, using a mean squared error (L2) loss. The LightGBM and the fourth CNN optimize M/L according to a ℒ₁ loss. The initial weights of this last CNN were set to the final values of the first CNN (pretraining). The boxes with the bright green outline were used as features for the LightGBM, after disregarding the ones that always were zero (dead neurons). For each of the neural layers or blocks, the output dimension is provided on the left.