Locality-based transfer learning on compression autoencoder for efficient scientific data lossy compression

Abstract

Scientific simulation can generate petabyte-level data per run nowadays. To significantly reduce the data size while simultaneously maintaining the compression quality based on certain user requirements, error-bounded lossy compression techniques such as SZ and ZFP are now becoming popular. However, these techniques still cannot achieve a reduction ratio of more than two orders of magnitude with a low compression error. On the other hand, in deep learning, the autoencoder techniques have been widely used in data compression, especially images. As an alternative, the compression autoencoder (CAE) has recently been investigated to compress the scientific data. Although CAE provides a higher compression ratio than SZ and ZFP, it suffers from a high training overhead, which makes it almost impractical in real compression scenarios. In this paper, we propose a new locality-based transfer learning method in order to significantly increase the training speed of CAE while achieving a high compression ratio. We also adopt incremental learning to maintain a high prediction accuracy and use KL-divergence as an indicator to quickly identify whether a target domain has a low testing error. Our evaluation results show that, after using the locality-based transfer learning, the training time can be reduced by up to 1200 times, and still has a 2 to 4X compression ratio gain over the state-of-the-art scientific data lossy compressor SZ.