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

2025-02-12

AI Summary

To address the inefficiency of sample-wise training and the inability to share structural priors in implicit neural representations (INRs) for large-scale time-varying or ensemble volumetric data, this paper introduces meta-learning—specifically, Model-Agnostic Meta-Learning (MAML)—to INR pretraining for the first time, proposing a general framework for learning universal initial parameters. The method acquires strong generalizable priors via pretraining on a small set of volumetric datasets, enabling few-step adaptive encoding for unseen volumes and zero-shot cross-dataset transfer. Furthermore, feature disentanglement analysis is integrated to enhance parameter interpretability and task adaptability. Experiments demonstrate over 90% faster convergence across multiple volumetric benchmarks; while reconstruction quality for unseen volumes remains comparable, significant gains are achieved in downstream applications—including simulation parameter analysis and critical time-step identification.

Technology Category

• Machine Learning: Learning with Manifolds
• Computer Vision: Representation Learning for Vision
• Search and Optimization: Metareasoning and Metaheuristics

Application Category

• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Web Mining and Content Analysis: Large pretrained models with web data

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.

Problem

• Efficient encoding of volumetric data
• Improving INR network optimization
• Enhancing generalizability of implicit neural representation

Innovation

• Meta-learning initial parameters
• Enhances INR generalizability
• Efficient volumetric data encoding