In order to answer the open questions of modern cosmology and galaxy evolution theory, robust algorithms for calculating photometric redshifts (photo-z) for very large samples of galaxies are needed. Correct estimation of the various photo-z algorithms' performance requires attention to both the performance metrics and the data used for the estimation. In this work, we use the supervised machine learning algorithm MLPQNA (Multi-Layer Perceptron with Quasi-Newton Algorithm) to calculate photometric redshifts for the galaxies in the COSMOS2015 catalogue and the unsupervised Self-Organizing Maps (SOM) to determine the reliability of the resulting estimates. We find that for z_spec_<1.2, MLPQNA photo-z predictions are on the same level of quality as spectral energy distribution fitting photo-z. We show that the SOM successfully detects unreliable zspec that cause biases in the estimation of the photo-z algorithms' performance. Additionally, we use SOM to select the objects with reliable photo-z predictions. Our cleaning procedures allow us to extract the subset of objects for which the quality of the final photo-z catalogues is improved by a factor of 2, compared to the overall statistics.