Mean residuals as a function of spectroscopic redshifts (gray) and CNN redshifts (red). The dashed lines show the corresponding results for B16. The histograms are the respective redshift distributions. The CNN redshifts tend to be slightly over(under)-estimated below(above) the most populated redshift bins of the training sample, that is, they are biased toward the region of highest training. The effect is larger for B16, however no bias is found as a function of photometric redshift in either case.