Maximal Sparsity with Deep Networks?