Efficient Visualization of Implicit Neural Representations via Weight Matrix Analysis

Jennifer Zvonek, Andrew Gillette

Keywords: Implicit neural representation, pruning, visualization, adaptive mesh refinement

TL;DR: From a pre-trained implicit neural representation with no access to its training data, we analyze the weight matrices to produce a variable resolution visualization with significant memory savings.

Abstract:

An implicit neural representation (INR) is a neural network that approximates a function over space and possibly time. Memory-intensive visualization tasks, including modern 4D CT scanning methods, represent data natively as INRs. While such INRs are prized for being more memory-efficient than traditional data on a lattice, discretization to a regular grid is still required for many visualization tasks. We present an algorithm to store high-resolution voxel data only for regions with significant detail, reducing memory requirements. To identify these high-detail areas, we use an interpolative decomposition pruning method on the weight matrices of the INR. The information from pruning is used to guide adaptive mesh refinement, allowing automatic mesh generation, tailored to the underlying resolution of the function. From a pre-trained INR with no access to its training data, we produce a variable resolution visualization with significant memory savings.

Primary Area: other topics in machine learning (i.e., none of the above)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 11053