Training Neural Networks with Fixed Sparse Masks