The work discussed here represents the natural evolution of a previous attempt described in these pages and presented in the 2002 and 2003 papers.
The final result, namely the redshifts for a large subsample of the galaxies present in the SDSS are downloadable here. This work was part of the Ph.D. Thesis of Raffaele D'Abrusco and has been published in Ap.J (2007).

The main idea behind the work is to exploit the huge data wealth of the SDSS to train a supervised neural network to recognize photometric redshifts. The details of the work can be found in this paper. In short the procedure can be summarized as it follows:

• The training, validation and test sets are built using the SDSS spectroscopic subsample. This sample is almost complete at m(R)<17.7, while for fainter magnitudes it includes mainly Luminous Red Galaxies or LRG's.

• A first MLP is trained at recognizing nearby (z<0.25) objects from distant (0.25<z<0.5) ones.

• Then two networks are trained in the two different redshift ranges and the optimal architecture is found by varying the NN parameters

• The resulting redshifts show a trend which is corrected by applying an interpolative correction.

• Once the three NN have been trained the photometric data are processed for the whole galaxy sample and the photometric redshifts are derived.

The whole procedure outlined above is repeated indipendently for all objects in the MAIN GALAXY sample of the SDSS and for the LRG's only. The resulting catalogues can be downloaded here.

The main results can be summarized as it follows.

1. The method leads to an r.m.s. error (evaluated on the test set only) better than any other method so far appeared in the literature.

In the following figure we display for the spectroscopic datasets the trend of spectroscopic versus photometric redshifts for the Main Galaxy sample (id est all galaxies regardless they are LRG's or not). Due to the large number of points to be displayed we show them as isocontours.

Errors appear to be rather symmetric and constatnt over the whole range (see next figure)