Title: Spectral pruning of fully connected layers: ranking the nodes based on the eigenvalues

Abstract: Training of neural networks can be reformulated in dual space, by allowing eigenvalues and eigenvectors to act as target of the optimization. Working in this setting, we show that the eigenvalues can be used to rank the nodes' relevance, within the collection. By sorting the nodes based on their associated eigenvalues enabled us to devise effective pre- and post-processing trimming strategies to yield compact networks (in terms of the number of composing neurons) with virtually unchanged performance, as compared to their untrimmed homologues. The proposed methods are tested for different architectures, with just a single or multiple hidden layers, and against distinct classification datasets of valuable pedagogical interest.