1 of 1

Guaranteeing Error Bounds in Neural Network Based Autoencoders�With HBPS FES SciDAC and Data Reduction projects and Sirius-2 project

Scientific Achievement

We developed a neural network formulation for autoencoders than can be used for data reduction of large-scale scientific applications with guaranteed uncertainties, and we applied this to XGC fusion edge code.

Significance and Impact

Approach provides a Guaranteed Autoencoder (GAE) that guarantees error bounds on all instances.
Approach provides novel post-processing algorithms for both linear and nonlinear constraints that obtain high accuracy on downstream Quantities of Interest (QoI).

Technical Approach

Our approach utilizes neural networks with piece-wise linear activation units (PLUs) which can be represented as instance-specific linear operators.

Institution Logo #1

PI(s)/Facility Lead(s): Anand Rangarajan and Sanjay Ranka, University of Florida

Collaborating Institutions: Scott Klasky, ORNL

ASCR Program: RAPIDs/Data Reduction

ASCR PM: Lali Chatterjee, Lentz, Margaret, Kalyan Perumalla

The Guaranteed Autoencoder (GAE) provides both primary data (PD) and quantities of interest (QoI) error bound guarantees. The autoencoder and/or the constraint satisfaction network can be converted into instance specific linear operators (by recording the paths of the PD through the network). Using a linear operator formulation, we can reduce errors on both PD and QoI (the latter by several orders of magnitude for 4 XGC QoI). Results are for XGC ITER Simulation with 162 GB of data per time step.

LOCAL LAB POC: Sanjay Ranka, University of Florida, Gainesville, 352 514 4213

TALKING POINTS:

We have developed a new Guaranteed Autoencoder (GAE) that provides error guarantees on all instances rather than only at an aggregate basis.
Our approach is based on a key observation that neural networks with piece-wise linear activation units (PLUs) can be represented as instance-specific linear operators.
Experimental results on scientific and image datasets show that the GAE can successfully satisfy user-provided error bounds for primary data and quantities of Interest at an instance level while achieving high compression.
The underlying methods can also work in conjunction with traditional reduction approaches such as MGARD, SZ and ZFP for data compression.

CITATIONS:

Jaemoon Lee, Anand Rangarajan, and Ranka Sanjay. Nonlinear Constraint Satisfaction for Compressive Autoencoders using Instance-specific Linear Operators. Proceedings of the 15^th International Conference on Contemporary Computing, 2023, pages 562-571.
Jaemoon Lee, Anand Rangarajan, and Ranka Sanjay. Nonlinear-by-Linear: Guaranteeing Error Bounds in Compressive Autoencoders. Proceedings of the 15^th International Conference on Contemporary Computing, 2023, pages 552-561.
Jaemoon Lee, Anand Rangarajan, and Sanjay Ranka, S. (2023, March). Constrained Autoencoders: Incorporating equality constraints in learned scientific data compression. In 2023 Data Compression Conference (DCC) (pp. 347-347). IEEE.

METADATA:

Name of the associated awarded project: Hybrid learning techniques for Scientific Data Reduction with performance guarantees

PI name(s): Sanjay Ranka and Anand Rangarajan, University of Florida, Scott Klasky, ORNL

Name of the program manager: Lali Chatterjee, Lentz, Margaret, Kalyan Perumalla