Recent advances in deep learning optimization showed that, with some
a-posteriori information on fully-trained models, it is possible to match the
same performance by simply training a subset of their parameters. Such a
discovery has a broad impact from theory to applications, driving