Created on 2019-06-14 17:36:03 by Marie Treyer. % Done: 0

The training set is underpopulated at z<0.2 and I>20 (Fig 2), among other things. In the test sample (Primus) on the contrary, most of the galaxies at z<0.2 have I>20 (Fig 3, Fig 5). These galaxies tend to be classified at higher redshift where faint galaxies are in the training set, i.e. the CNN tends to reproduce the magnitude/zspec distribution of its training set. Weights (designed to reproduce the counts in magnitude and the z-distributions in bins of mag) were applied to remedy this bias, unfortunately it doesn't seem to make much difference at this point. The dip at z=0.4 has been filled though, but the peak at z=0.9 hasn't gone away.