Sparse Meta-Learned Neural Networks for High-Rate Data Compression

Solution Overview

Problem

Existing implicit neural representation (INR) methods for data compression require transmitting a dense set of network parameters, which do not scale to real-world compression scenarios and large-scale data signals, and thus, do not achieve high compression rates while maintaining reconstruction quality.

Innovation Solution

A data reconstruction neural network that learns per-signal parameter value updates for a subset of network parameters, encouraging sparsity to reduce the amount of data communicated between compression and decompression systems, using a differentiable sparsity term to penalize non-zero updates.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a dense set of network parameters is transmitted for implicit neural representation compression, then reconstruction quality is maintained, but compression rate is low and scalability to large-scale data signals is poor

Engineering Contradiction:

Improvereconstruction qualityVSAvoidcompression rate

Core Design Contradiction:

Measurement precisionVSProductivity

Solution Approach 1:

The patent extracts only the essential information needed for compression by identifying and transmitting only a subset of network parameters that are most important for reconstructing the input signal. This is achieved through parameter selection mechanisms that filter out redundant parameters, allowing the system to maintain reconstruction quality while significantly reducing the amount of data transmitted, thus resolving the contradiction between compression rate and reconstruction quality.

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent applies local quality by differentiating between different network parameters and treating them differently based on their importance. Instead of uniformly compressing all parameters, the system identifies which parameters are critical for maintaining reconstruction quality and focuses compression efforts on those specific parameters, allowing for higher compression rates without sacrificing overall reconstruction quality.

Inventive Principle:

Principle #3Local quality

2Measurement precision

If a dense set of network parameters is transmitted for implicit neural representation compression, then reconstruction quality is maintained, but device complexity and data communication requirements increase

Engineering Contradiction:

Improvereconstruction qualityVSAvoiddata communication complexity

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent extracts only the essential information needed for compression by identifying and transmitting only a subset of network parameters that are most important for reconstructing the input signal. This is achieved through parameter selection mechanisms that filter out redundant parameters, allowing the system to maintain reconstruction quality while significantly reducing the amount of data transmitted, thus resolving the contradiction between compression rate and reconstruction quality.

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The patent segments the network parameters into different categories or groups based on their importance and functionality. By dividing the parameter space into essential and non-essential components, the system can selectively transmit only the essential segments, reducing communication complexity while preserving the ability to reconstruct the input signal with high quality.

Inventive Principle:

Principle #1Segmentation

Data Source

PatentUS20250356175A1Data compression and reconstruction using sparse meta-learned neural networks

Publication Date: 2025.11.20 GDM HOLDING LLC

AI summary

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for compressing and decompressing data signals using sparse, meta-learned neural networks.