Machine learning for prediction of extreme statistics in modulation instability