1 of 1

Regularized Ensemble Scene Representation Network

With the BER ImPACTS SciDAC5 Partnership program

Scientific Achievement

Scene Representation Networks, despite being compact scientific data representations, lack the ability to depict its prediction quality after training. We propose Regularized Ensemble SRN (RE-SRN) for quality-aware predictions and visualizations. RE-SRN outperforms alternative ensemble and Bayesian methods in uncertainty and data reconstruction evaluations

Significance and Impact

RE-SRN endows SRNs with reliable prediction variance estimation to guide a quality-aware reconstruction and visualization with variance visualizations
Advanced uncertainty-aware visualization methods, which are not applicable to conventional SRNs, can be adapted to RE-SRN to integrate the uncertainty in predictions directly into visualization

Figure: RE-SRN learns to represent a scientific dataset with a multi-decoder architecture that predicts multiple data values for any given coordinate in the spatial domain, from which a prediction variance can be computed and visualized along with the predicted data to indicate prediction confidence and quality (in blue). ��In comparison with competing methods for prediction variance estimation such as Bayesian and conventional ensemble methods, RE-SRN delivers the most accurate data reconstruction and competitive variance quality closely resembling the prediction error.

Technical Approach

RE-SRN comprises a parameter-efficient architecture with multi-decoder and a shared feature grid encoder to generate uncertain predictions
Variance regularization loss for RE-SRN minimizes the dissimilarity between spatial distributions of variance and error for a reliable variance

PI: Han-Wei Shen; Ohio State University

SciDAC Partnership: ImPACTS, Improving Projections of AMOC and Collapse Through Advanced Simulations

ASCR Program: SciDAC RAPIDS2

Publication: Regularized Multi-Decoder Ensemble for an Error-Aware Scene Representation Network

Authors: Tianyu Xiong, Skylar W. Wurster, Hanqi Guo, Tom Peterka, Han-Wei Shen (In review)

LOCAL LAB POC:

TALKING POINTS:

Scene representation networks, or SRNs, are neural networks that can efficiently represent large-scale scientific data with a significant reduction in storage cost. However, since the ground truth scientific data are discarded after the training, existing SRNs lose the ability to depict if its prediction quality, despite the predictions are prone to approximation errors. Without being informed the prediction quality, inaccurate analysis and visualizations from SRNs can mislead scientific discovery and decision making.

To empower SRNs with post-training prediction quality assessment, we propose Regularized Ensemble SRN, RE-SRN in short, that can quantify a prediction variance to indicate the quality of the predictions, such that scientists can leverage the variance in visualization and data analysis to understand where in the spatial domain RE-SRN makes inaccurate data reconstructions, as illustrated in the blue parts of the figure.

In comparison with competing methods for prediction variance estimation such as Bayesian and conventional ensemble methods, RE-SRN delivers the most accurate data reconstruction and competitive variance quality closely resembling the prediction error.

RE-SRN consists of a shared feature grid encoder and multiple multi-layer perceptron decoders to achieve a high level of parameter efficiency compared with basic ensemble approaches. Furthermore, by training with the proposed variance regularization loss, RE-SRN can acquire a faithful variance that correlates well to its prediction error.

METADATA:

Name of the associated awarded project:

PI name(s): Han-Wei Shen (OSU)

Name of the program manager: Kalyan Perumalla