Influence of a few features on M/L, for the training set. Due to the large number of datapoints, we use (hexagonal) bins. The opacity of each bin corresponds to the number of galaxies in the bin, in a non-linear way (ensuring that lower densities are still visible). The total luminosity denotes the total g-band luminosity in L_⊙, g, and is shown in log space. For the top left panel, the P_feature feature from CNN 1 estimates the GZ2 probability of the galaxy having a feature or disk. For the bottom left panel, the morphology is determined from the last layer of CNN 1. A galaxy is defined as irregular if the predicted probability of being irregular is larger than 20%, it is considered edge-on if it is not irregular but has a probability of being edge-on larger than 40%, and it is a feature or disk if the corresponding probability is larger than 40% (but it is not in the previous two categories). The ellipticals are the remaining datapoints.