Implicit Neural Representation Training with Residual Refinement

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Training implicit neural representations (INRs) for large or high-dimensional digital signals is computationally expensive and often results in image artefacts due to methods like tiling and multiple resolutions, which are inefficient and impractical.

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 artefacts, reducing the need to train on all samples.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If training is performed on all samples of a large or high-dimensional signal, then accuracy is improved, but memory footprint and training time become prohibitively large

Engineering Contradiction:
ImproveaccuracyVSAvoidmemory footprint
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The training process is segmented into multiple resolutions, where the signal is processed at different levels of detail. The method trains INRs at multiple resolutions and combines them, allowing the system to handle large signals without loading all data simultaneously into memory, thus reducing memory footprint while maintaining accuracy.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The method introduces a resolution dimension to the training process. Instead of training at a single resolution, it trains at multiple resolutions and combines the results. This dimensional approach allows efficient processing of large signals by distributing the computational load across different resolution levels rather than requiring all data in memory at once.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Measurement precision

If training is performed on all samples of a large or high-dimensional signal, then accuracy is improved, but training time becomes prohibitively long

Engineering Contradiction:
ImproveaccuracyVSAvoidtraining time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The training process is divided into segments at different resolutions. By training at multiple resolutions and combining results, the method reduces total training time compared to training on all full-resolution data, while still achieving high accuracy through the progressive refinement across resolution levels.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The method uses partial action by training at selected resolutions rather than all possible resolutions. It strategically chooses which resolutions to train at, combining them to achieve sufficient accuracy without the computational cost of training on every possible resolution level, thus reducing training time.

Inventive Principle:
Principle #16Partial or excessive action

3Quantity of substance

If the signal is split into smaller tiles and separate INRs are trained on each tile, then memory footprint is reduced, but image artefacts appear at tile boundaries

Engineering Contradiction:
Improvememory footprintVSAvoidimage quality
Core Design Contradiction:
Quantity of substanceVSManufacturing precision

Solution Approach 1:

Instead of splitting spatially (which causes boundary artefacts), the method introduces a resolution dimension. It trains INRs at multiple resolutions and combines them, avoiding spatial segmentation and its associated boundary artefacts while still reducing memory requirements through progressive resolution processing.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Solution Approach 2:

The method merges INRs trained at different resolutions rather than merging spatial tiles. By combining the outputs of multi-resolution INRs, it achieves seamless results without the boundary artefacts that occur when combining spatially segmented tiles, thus maintaining image quality while reducing memory footprint.

Inventive Principle:
Principle #5Merging (Combining)

4Productivity

If multiple resolutions are used to represent different levels of detail, then training efficiency is improved, but image artefacts appear when neighbouring tiles are trained at different resolutions

Engineering Contradiction:
Improvetraining efficiencyVSAvoidimage quality
Core Design Contradiction:
ProductivityVSManufacturing precision

Solution Approach 1:

The method merges INRs trained at different resolutions in a controlled manner. By systematically combining the multi-resolution INRs, it maintains the training efficiency benefits of using multiple resolutions while eliminating the image artefacts that occur when tiles at different resolutions are simply concatenated.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The method transitions from spatial dimension segmentation to resolution dimension segmentation. This allows efficient multi-resolution training without the boundary artefacts of spatial tiling, as the combinations occur in the resolution domain rather than the spatial domain, preserving image quality.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Data Source

PatentEP4597367A1Generating an implicit neural representation
Publication Date: 2025.08.06 SONY COMP ENTERTAINMENT EURO LTD
  • EP4597367A1 patent drawingFigure 1
  • EP4597367A1 patent drawingFigure 2
  • EP4597367A1 patent drawingFigure 3

AI summary

A method of training an implicit neural representation (INR) on 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.