Main results of correcting z_spec-dependent residuals and mode collapse with the SDSS data and the network Net_P. Upper panel: mean residuals as a function of spectroscopic redshift for the baseline method and steps 1-3 of our methods with the SDSS data and the network Net_P. The corresponding < Δ_z > − z_spec piecewise linear fits (Eq. (9)) are shown as dashed lines with the same colors. Zero residual is indicated by the horizontal dashed black line. Middle panel: sample distributions of the estimated photometric and spectroscopic redshift in different cases. Lower left panel: σ_MAD (Eq. (11)) as a function of r-band magnitude. Lower right panel: total variation distance between the z_photo and z_spec sample distributions (Eq. (10)) as a function of the slope of the < Δ_z > − z_spec piecewise linear fit (Eq. (9)) in the high-redshift interval. The shaded area shows the total variation distance with the 1 a uncertainty for the simulated z_photo sample distribution for the baseline method that is expected to have no mode collapse. Step 3\2 denotes a case in which step 3 is applied without step 2. Step 3a denotes a variant case of step 3 in which hard labels are used instead of soft labels, but the means are adjusted. Step 3b denotes a case in which soft labels are used but the means are not adjusted.