The thesis focuses on deep learning methods applied to discovery of gravitational lensing events in the universe. Publicly available I-band images of the known gravitational lenses were combined with simulated ones and randomly sampled cutouts of the galaxies and stars. Deep convolutional networks outperform the conventional discovery methods and achieve up to 0.9984 mean ROC AUC and 0.9895 mean F1-score on the out-of-sample 7-fold cross-validation. The models demonstrated excellent agreement with the latest list of 92 candidates published in the literature and created with combination of deep learning and manual analysis by professional astronomers.