This research is a trial to use deep learning for crime prediction in Japan. Lately many kinds of public data have been releasing as open data. With these data, we can freely try deep learning which needs massive training data. Since crime victimization location is not random but has spatial autocorrelation, we try to predict crime victimization with autocorrelation as two kinds of problems, regression and classification. As a result, adjusted coefficient of determination reached 0.82 as regression, and we get probabilities of crime as classification.