Implicit Neural Representation Training with Recursive Residual Refinement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Training implicit neural representations (INRs) for large or high-dimensional digital signals is computationally expensive and results in a large memory footprint, with known methods leading to image artefacts due to training separate INRs on different resolutions of signal tiles.
Innovation Solution
A method of training INRs by recursively refining neural networks using ground-truth signals of varying sizes, starting from the smallest, and combining residual networks to minimize the impact of tiling artefacts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If training is performed on the entire high-dimensional signal at full resolution, then the INR achieves high accuracy, but the computational cost and memory footprint become intractable
Solution Approach 1:
The patent divides the high-dimensional signal into multiple smaller tiles or patches, training separate INRs on each tile. This segmentation reduces the computational burden and memory requirements while maintaining the ability to reconstruct the full signal by combining tile predictions, thus resolving the contradiction between accuracy and computational cost.
Solution Approach 2:
The patent introduces a tile index dimension to the INR architecture, transforming the problem from directly modeling high-dimensional space to modeling lower-dimensional tiles with an additional discrete index dimension. This dimensional transformation enables efficient computation while preserving the ability to represent the full high-dimensional signal through indexed tile assembly.
2Power
If the signal is split into tiles and separate INRs are trained on each tile, then computational requirements are reduced, but image artefacts appear at tile boundaries
Solution Approach 1:
The patent merges multiple tile-based INR predictions into a unified output by combining the predictions from adjacent tiles. This merging process, which may involve averaging or other fusion techniques, reduces the visibility of boundary artefacts while maintaining the computational efficiency of tile-based processing.
Solution Approach 2:
The patent applies different processing or refinement strategies to different regions, particularly at tile boundaries versus interior regions. By locally adjusting the combination or refinement process at boundaries, the method reduces artefact visibility while maintaining efficiency, thus resolving the contradiction between computational cost and image quality.
3Measurement precision
If more parameters are used in the INR to improve accuracy, then the representation quality increases, but the memory footprint and training time increase disproportionately
Solution Approach 1:
The patent segments the parameter space by training separate compact INRs on different tiles rather than using one large high-parameter INR. This segmentation allows each tile's INR to have fewer parameters while the collective set of tile INRs maintains overall representation quality, thus resolving the contradiction between accuracy and parameter quantity.
Solution Approach 2:
The patent uses multiple lower-resolution tiles that collectively cover the full signal space, applying partial processing to each tile rather than attempting to process the entire high-dimensional signal at once with a single high-parameter model. This partial action approach achieves comparable representation quality with fewer total parameters.
Data Source
AI summary
A method of training an implicit neural representation (INR) of an input digital data signal comprises the steps of obtaining a set of at least two ground-truth signals derived from the input signal, each ground-truth signal being a different size of the input signal, and creating an estimate INR. training the estimate INR comprises training an estimate INR using the smallest ground-truth signal in the set, and for each of a subset of the remaining ground-truth signals in the set, from the smallest to the largest in the subset, recursively refining the immediately previous estimate INR. This comprises training a first residual INR on a comparison of the previous estimate INR and the ground-truth signal, and combining the first residual INR and previous estimate INR.