LAUSR

Deep neural networks exploiting million parameters are currently the norm. This is a potential issue because of the great number of computations needed for training, and the possible loss of generalization performance of overparameterized networks. We propose in this paper a method for learning sparse neural topologies via a regularization approach that identifies nonrelevant weights in any type of layer (i.e., convolutional, fully connected, attention and embedding ones) and selectively shrinks their norm while performing a standard back-propagation update for relevant layers. This technique, which is an improvement of classical weight decay, is based on the definition of a regularization term that can be added to any loss function regardless of its form, resulting in a unified general framework exploitable in many different contexts. The actual elimination of parameters identified as irrelevant is handled by an iterative pruning algorithm. To explore the possibility of an interdisciplinary use of our proposed technique, we test it on six different image classification and natural language generation tasks, among which four are based on real datasets. We reach state-of-the-art performance in one out of four imaging tasks while obtaining results better than competitors for the others and one out of two of the considered language generation tasks, both in terms of compression and metrics.