Compression strategies and space-conscious representations for deep neural networks