With the advent of deep and wide multi-band photometric surveys, like DES, there has been a resurgence of interest in photometric redshifts as a means of estimating the distance to a range of astrophysical objects. As consequence, the use of photometric redshifts in cosmology probes such correlation functions and weak lensing tomography is increasing. Often, however these photo-zs are treated like spectroscopic observation, using one value estimate rather the full probability density function (PDF). In this talk, I will focus the discussion mainly on the new methods and approaches we have been developing to obtain robust and accurate redshift PDF by efficiently combining Bayesian template fitting techniques with the powerful tools of active machine learning, such TPZ , a new parallel machine learning photo-z estimator that provides robust PDFs along with useful ancillary information about the data that will enable more accurate predictions, a better understanding of the internal structure of the survey and a better design for follow up observations.