Learned Compander for Region-Specific Data Compression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current data compression techniques, such as those using neural networks, compress entire data regions uniformly, leading to suboptimal compression ratios and quality degradation in regions of interest for AI applications, where higher compression ratios are needed without compromising data quality.
Innovation Solution
A data processing system that uses a learned compander to compress and expand data, with an estimation unit identifying regions of interest and a learning unit adjusting the compression ratio based on these regions, allowing for higher compression ratios in non-AI regions while maintaining data quality for AI applications.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If uniform compression is applied to entire data regions, then the compression process is simple, but the compression ratio deteriorates in AI regions of interest
Solution Approach 1:
The data is divided into multiple regions based on AI model predictions, with each region compressed at different ratios. The estimation unit identifies AI regions of interest, and the compander applies region-specific compression strategies, transforming uniform compression into segmented differential compression to resolve the contradiction between process simplicity and compression efficiency.
Solution Approach 2:
Different compression ratios are applied to different regions according to their importance for AI processing. AI regions of interest receive lower compression ratios to preserve quality, while non-AI regions receive higher compression ratios to maximize space savings, achieving local optimization that resolves the uniform compression dilemma.
2Quantity of substance
If higher compression ratio is applied to reduce data volume, then data volume reduction is improved, but data quality for AI applications deteriorates
Solution Approach 1:
The system applies differential compression where AI regions of interest maintain higher quality with lower compression ratios, while non-AI regions achieve higher compression ratios for maximum volume reduction. This local quality differentiation allows the system to simultaneously minimize data volume and preserve AI-required data quality.
Solution Approach 2:
The compression strategy dynamically adjusts compression ratios based on real-time estimation of AI regions of interest. The compander learns from the evaluation function and adapts compression parameters per region, enabling dynamic optimization between data volume reduction and quality preservation rather than using fixed uniform compression.
3Loss of information
If region-specific compression is implemented, then compression ratio is improved, but system complexity increases
Solution Approach 1:
A single learned compander model performs multiple functions: it estimates AI regions of interest, determines optimal compression ratios for each region, and executes differential compression. This multi-functional design achieves region-specific compression benefits while avoiding the complexity of separate estimation and compression systems.
Solution Approach 2:
The compander is trained using learning data and an evaluation function that automatically identifies AI regions and determines optimal compression strategies. The system self-adjusts compression parameters based on learned patterns, eliminating the need for external manual configuration and reducing operational complexity while maintaining high compression ratios.
Data Source
AI summary
Provided is a data processing system which includes a processor and a storage device, and inputs/outputs data using a learned compander that compresses and expands data, wherein the data processing system comprises an estimation unit which uses learning data and estimates a region of interest to a data model, and a learning unit which causes the compander to learn according to an evaluation function in which each region was weighted based on the region of interest, and a result of the compander compressing and expanding the learning data.