@paul_rietschka Before they had to optimize hyperparams again after each tweak to the features, and the pipeline had several steps. This simplified a lot going to neural networks. The net didn‘t do something wildly different after all, but it paid off.