Towards Generalized Entropic Sparsification for Convolutional Neural Networks

arXiv:2404.04734v1 Announce Type: new
Abstract: Convolutional neural networks (CNNs) are reported to be overparametrized. The search for optimal (minimal) and sufficient architecture is an NP-hard problem as the hyperparameter space for possible network configurations is vast. Here, we introduce a layer-by-layer data-driven pruning method based on the mathematical idea aiming at a computationally-scalable entropic relaxation of the pruning problem. The sparse subnetwork is found from the pre-trained (full) CNN using the network entropy minimization as a sparsity constraint. This allows …

abstract architecture arxiv cnns convolutional neural networks cs.cv data data-driven generalized hyperparameter layer network networks neural networks np-hard pruning search space type vast