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

Area of Science:

• Scientific Visualization
• Machine Learning
• Data Representation

Background:

• Scene Representation Networks (SRNs) are used for compact scientific data representation, but lack inference-time quality assessment.
• Assessing SRN prediction quality is crucial for trusting scientific visualizations, especially since they are lossy, black-box models.
• Current methods cannot evaluate coordinate-level errors without ground truth data, limiting their utility in scientific applications.

Purpose of the Study:

• To develop an SRN architecture capable of assessing reconstruction quality at inference time.
• To enable confidence-aware data reconstruction and visualization by quantifying prediction uncertainty.
• To improve the reliability of variance estimation in uncertain neural network architectures for scientific data.

Main Methods:

• Proposed a parameter-efficient multi-decoder SRN (MDSRN) architecture with a shared feature grid and multiple decoders.
• Introduced a novel variance regularization loss for ensemble learning to create Regularized multi-decoder SRN (RMDSRN).
• Evaluated MDSRN and RMDSRN against existing uncertain SRN methods (MCD, MFVI, DE, PV) on diverse scalar field datasets.

Main Results:

• RMDSRN achieved the most accurate data reconstruction and competitive variance-error correlation among uncertain SRNs.
• Demonstrated that coordinate-level variance can be rendered to inform reconstruction quality or integrated into uncertainty-aware volume rendering.
• Showcased the effectiveness of RMDSRN with default configurations across various datasets, requiring no customized hyperparameter tuning.

Conclusions:

• RMDSRN offers a robust solution for confidence-aware reconstruction in scientific visualization using SRNs.
• The proposed uncertainty quantification and regularization enhance the trustworthiness of visualized scientific data.
• This work paves the way for improved uncertainty-aware volume rendering and broader adoption of SRNs in scientific analysis.