On Sparse And Efficient Deep Learning