Left panel: bias as a function of Galactic extinction for the classifier with and without integrating E(B − V) into the training (red and orange lines respectively) and for B16 (green). The CNN tends to overestimate redshifts in obscured regions (confusing galactic dust attenuation with redshift dimming), unless E(B − V) is used for training. Right panel: bias as a function of disk inclination for galaxies classified as star-forming or starburst. The CNN is virtually unbiased wrt ellipticity, unlike B16.