Upcoming wide-field surveys, such as LSST and Euclid, will require precise and reliable photometric redshifts (photo-z) for fulfilling their goals. Currently, there are several major methods for obtaining photo-z, but all of them rely on spectroscopic samples, that are usually much more scarce than photometric data. The difference between spectroscopic and photometric datasets, as well as contamination of spectr_z catalog, is a possible source of biases. Such biases can significantly affect precise cosmological and astrophysical models. In this work, we investigate how unsupervised machine learning methods, in particular, Self-Organizing Maps, can help us to improve the reliability and precision of photo-z catalogs. On the example of COSMOS2015 catalog, we demonstrate that the use of such methods improves standard deviation of residuals of photo-z by the factor of two (from 0.04 to 0.02) and discuss perspectives for creating quality flag system for photometric redshifts.