Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

REV-INR: Regularized Evidential Implicit Neural Representation for Uncertainty-Aware Volume Visualization

Shanu Saklani, Tushar M. Athawale, Nairita Pal, David Pugmire, Christopher R. Johnson, Soumya Dutta

TL;DR

REV-INR introduces a Regularized Evidential INR that jointly predicts volume data values and coordinate-level epistemic and aleatoric uncertainties in a single forward pass using a Normal Inverse Gamma posterior. By coupling evidential learning with dedicated uncertainty regularizations, REV-INR delivers calibrated per-voxel uncertainties and improved isosurface reliability, outperforming MCD-INR and RMD-INR in reconstruction quality and uncertainty interpretability. Through two additional baselines and extensive visualization-focused evaluation, the work demonstrates that uncertainty-aware INRs enable trustworthy volume visualization while maintaining fast inference; however, training cost is higher and regularization choices may vary with datasets. The results suggest that uncertainty-aware INRs offer practical gains for scientific visualization, enabling analyses to be driven by model-predicted data with quantified confidence, and point to future directions in acceleration and extension to multivariate datasets.

Abstract

Applications of Implicit Neural Representations (INRs) have emerged as a promising deep learning approach for compactly representing large volumetric datasets. These models can act as surrogates for volume data, enabling efficient storage and on-demand reconstruction via model predictions. However, conventional deterministic INRs only provide value predictions without insights into the model's prediction uncertainty or the impact of inherent noisiness in the data. This limitation can lead to unreliable data interpretation and visualization due to prediction inaccuracies in the reconstructed volume. Identifying erroneous results extracted from model-predicted data may be infeasible, as raw data may be unavailable due to its large size. To address this challenge, we introduce REV-INR, Regularized Evidential Implicit Neural Representation, which learns to predict data values accurately along with the associated coordinate-level data uncertainty and model uncertainty using only a single forward pass of the trained REV-INR during inference. By comprehensively comparing and contrasting REV-INR with existing well-established deep uncertainty estimation methods, we show that REV-INR achieves the best volume reconstruction quality with robust data (aleatoric) and model (epistemic) uncertainty estimates using the fastest inference time. Consequently, we demonstrate that REV-INR facilitates assessment of the reliability and trustworthiness of the extracted isosurfaces and volume visualization results, enabling analyses to be solely driven by model-predicted data.

REV-INR: Regularized Evidential Implicit Neural Representation for Uncertainty-Aware Volume Visualization

TL;DR

REV-INR introduces a Regularized Evidential INR that jointly predicts volume data values and coordinate-level epistemic and aleatoric uncertainties in a single forward pass using a Normal Inverse Gamma posterior. By coupling evidential learning with dedicated uncertainty regularizations, REV-INR delivers calibrated per-voxel uncertainties and improved isosurface reliability, outperforming MCD-INR and RMD-INR in reconstruction quality and uncertainty interpretability. Through two additional baselines and extensive visualization-focused evaluation, the work demonstrates that uncertainty-aware INRs enable trustworthy volume visualization while maintaining fast inference; however, training cost is higher and regularization choices may vary with datasets. The results suggest that uncertainty-aware INRs offer practical gains for scientific visualization, enabling analyses to be driven by model-predicted data with quantified confidence, and point to future directions in acceleration and extension to multivariate datasets.

Abstract

Applications of Implicit Neural Representations (INRs) have emerged as a promising deep learning approach for compactly representing large volumetric datasets. These models can act as surrogates for volume data, enabling efficient storage and on-demand reconstruction via model predictions. However, conventional deterministic INRs only provide value predictions without insights into the model's prediction uncertainty or the impact of inherent noisiness in the data. This limitation can lead to unreliable data interpretation and visualization due to prediction inaccuracies in the reconstructed volume. Identifying erroneous results extracted from model-predicted data may be infeasible, as raw data may be unavailable due to its large size. To address this challenge, we introduce REV-INR, Regularized Evidential Implicit Neural Representation, which learns to predict data values accurately along with the associated coordinate-level data uncertainty and model uncertainty using only a single forward pass of the trained REV-INR during inference. By comprehensively comparing and contrasting REV-INR with existing well-established deep uncertainty estimation methods, we show that REV-INR achieves the best volume reconstruction quality with robust data (aleatoric) and model (epistemic) uncertainty estimates using the fastest inference time. Consequently, we demonstrate that REV-INR facilitates assessment of the reliability and trustworthiness of the extracted isosurfaces and volume visualization results, enabling analyses to be solely driven by model-predicted data.
Paper Structure (30 sections, 13 equations, 10 figures, 2 tables)

This paper contains 30 sections, 13 equations, 10 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Techniques of Uncertain Volume Visualization
  4. Deep Learning for Scalar Field Visualization
  5. Uncertainty in Neural Networks
  6. Uncertainty in INR-based Volume Modeling
  7. Uncertainty-Aware INRs for Reliable Volume Data Modeling and Visualization
  8. Implicit Neural Representation (INR)
  9. Proposed Regularized Evidential INR (REV-INR)
  10. Value Prediction and Uncertainty Estimation
  11. Loss Functions
  12. Regularization of REV-INR
  13. Training REV-INR with Combined Loss Functions
  14. Monte Carlo Dropout INR (MCD-INR) Enhanced with Aleatoric Uncertainty Estimation Capability
  15. Value Prediction and Uncertainty Estimation
  16. ...and 15 more sections

Figures (10)

  • Figure 1: Schematic architecture of the proposed REV-INR, MCD-INR, and RMD-INR. The proposed REV-INR is trained to predict the parameters of an evidential distribution, modeling a higher-order probability distribution.
  • Figure 2: Visualization of the Teardrop Dataset for isovalue $159.9798$. The middle segment shows the mean isosurfaces generated by REV-INR, RMD-INR, and MCD-INR, respectively, where differences from the ground truth are highlighted (dashed circles). The right segment shows the corresponding uncertain isosurfaces using the Level Crossing Probability (LCP). We observe that REV-INR produce the most accurate LCP visualization.
  • Figure 3: Visualization of the Vortex Dataset for the isovalue 5.8. The middle column shows the mean isosurfaces generated by REV-INR, RMD-INR, and MCD-INR, respectively, where differences from the ground truth are highlighted(dashed circles). The right column shows the corresponding uncertain isosurfaces using the Level Crossing Probability (LCP), showing the most probable isosurface. We observe that REV-INR produce the most accurate LCP visualization. The isosurface generated by Det-INR also fails to preserve the thin connection.
  • Figure 4: Visualization of the Combustion Dataset. The ground truth scalar field is compared with REV-INR, RMD-INR, and MCD-INR. REV-INR better preserves complex flame structures, as highlighted in the zoomed-in regions marked with black boxes.
  • Figure 5: Visualization of the Foot Dataset. The ground truth scalar field (left) is compared with reconstructions from REV-INR, RMD-INR, and MCD-INR. The highlighted regions around the joints of bone segments indicate that REV-INR achieves the closest match to the ground truth compared to the other methods.
  • ...and 5 more figures