It uses distributed training of the smaller model , take the average weights of all of them and use that to train a bigger model. Used CNN to classify digits and Open-CV to mask digits from the picture.