Redshift estimation is a fundamental task in astronomy. In particular, it serves as distance indicator in the cosmological field. As opposed to the expensive and time consuming redshift estimation via spectra, the use of photometry has emerged as a popular and practical alternative. In order to process the ever increasing number of sources, the use of machine learning methods has become mandatory. Manual feature extraction is the standard approach to train machine learning models. However, thanks to recent developments in GPU computing, it is now possible to perform massive feature selection, computing thousands of features combinations and selecting the best performing ones through a greedy forward selection scheme. On the other hand, the application of deep learning technologies and of convolutional neural networks allows estimating photometric redshifts directly from images, in a fully probabilistic way, precluding the need of pre-classification and feature extraction. In this framework, the process of feature extraction and selection is fully automatized. The success of several upcoming projects, such as the EUCLID mission, LOFAR, PANSTARRS and many more, requires the availability of highly affordable photometric redshifts. We believe that methods, along the ones presented here, will fulfill this requirement.