1 of 1

In Situ Compression Artifact Removal in Scientific Data Using Deep Transfer Learning and Experience Replay

Scientific Achievement

Developed a scalable in situ approach to train deep learning models that are capable of transfer learning to leverage knowledge from different domains and remove compression artifacts from lossy highly compressed images from scientific simulated data.

Significance and Impact

We developed a unified framework for in situ compression artifact removal in which the Depp neural networks are combined with scalable training, transfer learning, and experience replay to achieve superior accuracy and efficiency compared to standard image compression and more advanced compressed sensing methods.

Research Details

We adopt convolutional neural network-based architectures - Enhanced Deep Super-Resolution Network (EDSR) and Residual Dense Networks (RDN).
These models are initially trained offline using the simulation data from climate domain (shallow water equations on a sphere) with JPEG compressed images
We adopt the discrepancy-based deep domain adaptation approach combined with experience replay to transfer learn the knowledge from offline-trained model to an in situ setting to adapt to data from different domain (Kinetic Transport).
This approach is scaled by using data-parallel training with controlled learning rate updates, thus the CAR model will be ready as soon as the simulation is complete.

Comparison of EDSR and RDN enhanced images (using offline learning of Compression Artifact Removal model on climate data) with JPEG compressed and compressed sensing enhancement approach result.

ANL: S. Madireddy , P. Balaprakash ; BNL: J. Park, S. Yoo, ; NW: S. Lee, W. Liao,

ORNL: R. Archibald , C. Hauk, M. Laiu

Incremental batch transfer learning between datasets from two different domains

Madireddy, S., Park, J.H., Lee, S., Balaprakash, P., Yoo, S., Liao, W.K., Hauck, C.D., Laiu, M.P. and Archibald, R., 2020. In situ compression artifact removal in scientific data using deep transfer learning and experience replay. Machine Learning: Science and Technology, 2(2), p.025010

This work was supported in part by the U.S. Department of Energy, Office of Science, Office of Advanced Scientific Computing Research, Scientific Discovery through Advanced Computing (SciDAC) program through the FASTMath Institute under Lawrence Livermore National Laboratory Subcontract B626484.

Notes:

The massive amounts of data produced during simulation on high performance computers has grown

exponentially over the last decade, so is the need to store and move these large datasets. One approach

to decreasing this burden is to employ lossy compression, but this requires expensive decompression

methods (such as compressed sensing) when the end users require to work with decompressed data.

We propose an in situ deep transfer learning approach to learn models that can remove the compression

artifacts in the data while leveraging knowledge from offline-trained models (with data from different

scientific domains). The goal of the in situ approach is to train the compression artifact removal (CAR)

model as the simulation progresses, instead of waiting till the end. In addition, we propose data-parallel

training of the deep learning model, which further reduced the training time so that this training time is

masked behind the simulation runtime. This work specifically focuses on JPEG compressed image data

from computer simulations that are generated by hyperbolic partial differential equations (PDEs), which include a

majority of the simulations performed on HPC systems.

Link:

https://iopscience.iop.org/article/10.1088/2632-2153/abc326/meta