Towards ℓ1 Regularization for Deep Neural Networks: Model Sparsity Versus Task Difficulty | IEEE Conference Publication | IEEE Xplore