Paradigms of the baseline method and our methods for correcting z_spec-dependent residuals and mode collapse. The baseline method does not explicitly specify a representation and produces single-channel output estimates that only require one-hot labels for training. In contrast, our methods split representation learning and classification into three consecutive steps and produces multichannel output estimates. The labels for training the model match the dimensions of the output estimates, and they are softened and readjusted in step 3 according to the biases. The dashed arrows indicate that the part of the model prior to the representation is fixed in steps 2 and 3.