We developed a Deep Convolutional Neural Network ( CNN ) , used as a classifier , to estimate photometric redshifts and associated probability distribution functions ( PDF ) for galaxies in the Main Galaxy Sample of the Sloan Digital Sky Survey at z < 0.4 .
Our method exploits all the information present in the images without any feature extraction .
The input data consist of 64 \times 64 pixel ugriz images centered on the spectroscopic targets , plus the galactic reddening value on the line-of-sight .
For training sets of 100k objects or more ( \geq 20 % of the database ) , we reach a dispersion \sigma _ { MAD } < 0.01 , significantly lower than the current best one obtained from another machine learning technique on the same sample .
The bias is lower than 10 ^ { -4 } , independent of photometric redshift .
The PDFs are shown to have very good predictive power .
We also find that the CNN redshifts are unbiased with respect to galaxy inclination , and that \sigma _ { MAD } decreases with the signal-to-noise ratio ( SNR ) , achieving values below 0.007 for SNR > 100 , as in the deep stacked region of Stripe 82 .
We argue that for most galaxies the precision is limited by the SNR of SDSS images rather than by the method .
The success of this experiment at low redshift opens promising perspectives for upcoming surveys .