Spectroscopic redshifts of quasars are important inputs for constructing many cosmological models. Redshift measurement is generally considered to be a straightforward task performed by automatic pipelines based on template matching.

Due to the millions of spectra delivered by surveys of SDSS or LAMOST telescopes, it is impossible to verify all redshift measurements of automatic pipelines by a human visual inspection. However, the pipeline results are still taken as the "ground truth" for further statistical inferences.

Nevertheless, because of the similarity of patterns of quasar emission lines in different spectral ranges, an optimal match may be found for a completely different template position, causing severe errors in the measured redshift. For example, it may easily happen that a faint emission star with a noisy spectrum is identified as a high redshift quasar and vice versa.

We show such examples discovered by the consistency check of redshift measurements of the SDSS pipeline and redshift predictions of a regression Bayesian convolutional network. The network is trained on a large amount of human-inspected redshifts and predicts redshifts together with their predictive uncertainties. Therefore, it can also identify cases where predictions are uncertain and thus require human visual inspection.