A path-norm toolkit for modern networks: consequences, promises and challenges

arXiv:2310.01225v4 Announce Type: replace-cross
Abstract: This work introduces the first toolkit around path-norms that fully encompasses general DAG ReLU networks with biases, skip connections and any operation based on the extraction of order statistics: max pooling, GroupSort etc. This toolkit notably allows us to establish generalization bounds for modern neural networks that are not only the most widely applicable path-norm based ones, but also recover or beat the sharpest known bounds of this type. These extended path-norms further enjoy the …

abstract arxiv biases challenges consequences cs.lg dag etc extraction general math.st max modern networks norm path pooling relu statistics stat.ml stat.th toolkit type work