Learning sparse deep neural networks using efficient structured projections on convex constraints for green AI
Michel Barlaud, Frédéric Guyard
Abstract
Deep neural networks (DNN) have been applied recently to different domains and perform better than classical state-of-the-art methods. However the high level of performances of DNNs is most often obtained with networks containing millions of parameters and for which training requires substantial computational power. To deal with this computational issue proximal regularization methods have been proposed in the literature but they are time consuming. In this paper, we propose instead a constrained approach. We provide the general framework for this new projection gradient method. Our algorithm iterates a gradient step and a projection on convex constraints. We studied algorithms for different constraints: the classical ℓ <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sub> unstructured constraint and structured constraints such as the £2,1 constraint (Group LASSO). We propose a new ℓ <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1,1</sub> structured constraint for which we provide a new projection algorithm. Finally, we used the recent “Lottery optimizer” replacing the threshold by our ℓ <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1,1</sub> projection. We demonstrate the effectiveness of this method with three popular datasets (MNIST, Fashion MNIST and CIFAR). Experiments with these datasets show that our projection method using this new ℓ <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1,1</sub> structured constraint provides the best decrease in memory and computational power.