Learning to Compress: Local Rank and Information Compression in Deep Neural Networks