Introduced in the late 1980s for generalization purposes, pruning has now become a staple for compressing deep neural networks. Despite many innovations in recent decades, pruning approaches still face core issues that hinder their performance or scalability. Drawing inspiration from early work in the field, and especially the use of weight decay to achieve sparsity, we introduce Selective Weight Decay (SWD), which carries out efficient, continuous pruning throughout training. Our approach, theoretically grounded on Lagrangian smoothing, is versatile and can be applied to multiple tasks, networks, and pruning structures. We show that SWD compares favorably to state-of-the-art approaches, in terms of performance-to-parameters ratio, on the CIFAR-10, Cora, and ImageNet ILSVRC2012 datasets.

Pruning is a compression method which aims to improve the efficiency of neural networks by reducing their number of parameters while maintaining a good performance, thus enhancing the performance-to-cost ratio in nontrivial ways. Of particular interest are structured pruning techniques, in which whole portions of parameters are removed altogether, resulting in easier to leverage shrunk architectures. Since its growth in popularity in the recent years, pruning gave birth to countless papers and contributions, resulting first in critical inconsistencies in the way results are compared, and then to a collective effort to establish standardized benchmarks. However, said benchmarks are based on training practices that date from several years ago and do not align with current practices. In this work, we verify how results in the recent literature of pruning hold up against networks that underwent both state-of-the-art training methods and trivial model scaling. We find that the latter clearly and utterly outperform all the literature we compared to, proving that updating standard pruning benchmarks and re-evaluating classical methods in their light is an absolute necessity. We thus introduce a new challenging baseline to compare structured pruning to: ThinResNet.

Graph Signal Processing is a promising framework to manipulate brain signals as it allows to encompass the spatial dependencies between the activity in regions of interest in the brain. In this work, we are interested in better understanding what are the graph frequencies that are the most useful to decode fMRI signals. To this end, we introduce a deep learning architecture and adapt a pruning methodology to automatically identify such frequencies. We experiment with various datasets, architectures and graphs, and show that low graph frequencies are consistently identified as the most important for fMRI decoding, with a stronger contribution for the functional graph over the structural one. We believe that this work provides novel insights on how graph-based methods can be deployed to increase fMRI decoding accuracy and interpretability.