Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Meta-INR: Efficient Encoding of Volumetric Data via Meta-Learning Implicit Neural Representation

Maizhe Yang, Kaiyuan Tang, Chaoli Wang

TL;DR

This work addresses the inefficiency of encoding time-varying and ensemble volumetric data with implicit neural representations by introducing Meta-INR, a meta-learning-based pretraining method that yields a strong initial INR prior $ heta_m$ learned from partial observations. The approach combines meta-pretraining with a MAML-like inner/outer loop and sparse data downsampling, followed by volume-specific finetuning that rapidly adapts to each volume using $K$ gradient steps. Meta-INR demonstrates superior reconstruction quality and significant encoding speedups across diverse datasets, while enabling interpretable parameter representations useful for tasks like representative timestep selection and simulation-parameter analysis. The results suggest practical impact for scalable, high-fidelity volumetric encoding in visualization and simulation pipelines, with avenues for continual and grid-based extensions.

Abstract

Implicit neural representation (INR) has emerged as a promising solution for encoding volumetric data, offering continuous representations and seamless compatibility with the volume rendering pipeline. However, optimizing an INR network from randomly initialized parameters for each new volume is computationally inefficient, especially for large-scale time-varying or ensemble volumetric datasets where volumes share similar structural patterns but require independent training. To close this gap, we propose Meta-INR, a pretraining strategy adapted from meta-learning algorithms to learn initial INR parameters from partial observation of a volumetric dataset. Compared to training an INR from scratch, the learned initial parameters provide a strong prior that enhances INR generalizability, allowing significantly faster convergence with just a few gradient updates when adapting to a new volume and better interpretability when analyzing the parameters of the adapted INRs. We demonstrate that Meta-INR can effectively extract high-quality generalizable features that help encode unseen similar volume data across diverse datasets. Furthermore, we highlight its utility in tasks such as simulation parameter analysis and representative timestep selection. The code is available at https://github.com/spacefarers/MetaINR.

Meta-INR: Efficient Encoding of Volumetric Data via Meta-Learning Implicit Neural Representation

TL;DR

This work addresses the inefficiency of encoding time-varying and ensemble volumetric data with implicit neural representations by introducing Meta-INR, a meta-learning-based pretraining method that yields a strong initial INR prior learned from partial observations. The approach combines meta-pretraining with a MAML-like inner/outer loop and sparse data downsampling, followed by volume-specific finetuning that rapidly adapts to each volume using gradient steps. Meta-INR demonstrates superior reconstruction quality and significant encoding speedups across diverse datasets, while enabling interpretable parameter representations useful for tasks like representative timestep selection and simulation-parameter analysis. The results suggest practical impact for scalable, high-fidelity volumetric encoding in visualization and simulation pipelines, with avenues for continual and grid-based extensions.

Abstract

Implicit neural representation (INR) has emerged as a promising solution for encoding volumetric data, offering continuous representations and seamless compatibility with the volume rendering pipeline. However, optimizing an INR network from randomly initialized parameters for each new volume is computationally inefficient, especially for large-scale time-varying or ensemble volumetric datasets where volumes share similar structural patterns but require independent training. To close this gap, we propose Meta-INR, a pretraining strategy adapted from meta-learning algorithms to learn initial INR parameters from partial observation of a volumetric dataset. Compared to training an INR from scratch, the learned initial parameters provide a strong prior that enhances INR generalizability, allowing significantly faster convergence with just a few gradient updates when adapting to a new volume and better interpretability when analyzing the parameters of the adapted INRs. We demonstrate that Meta-INR can effectively extract high-quality generalizable features that help encode unseen similar volume data across diverse datasets. Furthermore, we highlight its utility in tasks such as simulation parameter analysis and representative timestep selection. The code is available at https://github.com/spacefarers/MetaINR.

Paper Structure

This paper contains 11 sections, 4 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Meta-INR
  4. Meta-Pretraining
  5. Volume-Specific Finetuning
  6. Results and discussion
  7. Datasets, Training, Baselines, and Metrics
  8. Time-Varying Data Representation
  9. Representative Timestep Selection
  10. Simulation Parameter Analysis
  11. Conclusions and future work

Figures (4)

  • Figure 1: Comparing different methods on volume rendering results. Top to bottom: half-cylinder, ionization, Tangaroa, and vortex.
  • Figure 2: Comparing different methods on isosurface rendering results. Top to bottom: half-cylinder, ionization, Tangaroa, and vortex. The chosen isovalues are reported in Table \ref{['tab:baseline-metrics']}.
  • Figure 3: t-SNE projections using parameters from different methods trained on the time-varying earthquake dataset, where selected timesteps are marked on the timeline on the left. SIREN's model parameters are not interpretable and fail to establish a correlation between timesteps. Volume rendering images are shown adjacent to selected timesteps for autoencoder and Meta-INR.
  • Figure 4: t-SNE projections of Meta-INR models trained on the ensemble Nyx dataset where each point represents a volume. Volumes with $h_0$ or $OmM_0$ matching the specified values are highlighted in light blue.