Adaptive Data Compression for Label-Free ML Classification

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing data compression techniques for classification models require deep understanding of the model and data labels, leading to inefficiencies and increased errors, especially in dynamic data environments where data distributions change over time.

Innovation Solution

An adaptive compression scheme that dynamically adjusts to data characteristics without the need for data labels, using a method based on Kullback-Leibler divergence to maintain model accuracy and optimize compression efficiency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Loss of energy

If data compression is applied to reduce bandwidth consumption, then bandwidth efficiency is improved, but model inference accuracy deteriorates due to increased error and quality drop

Engineering Contradiction:

Improvebandwidth consumptionVSAvoidmodel inference accuracy

Core Design Contradiction:

Loss of energyVSReliability

Solution Approach 1:

The patent implements dynamic compression by adjusting the compression rate based on the confidence scores from the classification model. When the model is highly confident about a data point's classification, higher compression rates are applied. When confidence is low, lower compression rates are used to preserve accuracy. This dynamic adaptation resolves the contradiction by making compression intensity dependent on the specific data point's importance.

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The system changes the compression parameter (compression rate) based on the classification confidence score. By mapping confidence scores to specific compression rates, the system optimizes the balance between bandwidth savings and accuracy preservation. High-confidence predictions use aggressive compression while low-confidence predictions use milder compression or remain uncompressed.

Inventive Principle:

Principle #35Parameter changes

2Productivity

If empirical techniques are used to determine compression rate, then compression efficiency can be optimized, but the process becomes complex and time-consuming requiring multiple iterations and data labels

Engineering Contradiction:

Improvecompression efficiencyVSAvoidcompression determination process

Core Design Contradiction:

ProductivityVSDevice complexity

Solution Approach 1:

The classification model itself provides the information needed to determine compression rates through its confidence scores. The system leverages the model's own output (confidence levels) to automatically control the compression process, eliminating the need for external empirical techniques, data labels, or manual tuning. The model essentially serves its dual purpose of classification and compression guidance.

Inventive Principle:

Principle #25Self-service

Solution Approach 2:

The system uses the confidence scores from the classification model as feedback to control the compression rate. This feedback loop allows the system to automatically adjust compression based on the model's performance, achieving optimal compression efficiency without complex external evaluation processes or iterative tuning with labeled data.

Inventive Principle:

Principle #23Feedback

3Ease of operation

If fixed compression parameters are used, then the compression process is simple, but the system cannot adapt to changing data distributions over time

Engineering Contradiction:

Improvecompression process simplicityVSAvoidadaptation to data distribution changes

Core Design Contradiction:

Ease of operationVSAdaptability or versatility

Solution Approach 1:

The system transitions from fixed to dynamic compression parameters by using the classification model's confidence scores to determine compression rates in real-time. This dynamic approach automatically adapts to changing data distributions because the confidence scores reflect the current data characteristics and model performance, maintaining both simplicity and adaptability.

Inventive Principle:

Principle #15Dynamics

Data Source

PatentUS20260031832A1Adaptive lossy compression for black-box classification models with label-less data

Publication Date: 2026.01.29 DELL PROD LP

AI summary

A method for an adaptive compression scheme that dynamically adjusts to data characteristics, maintaining model classification accuracy while optimizing compression efficiency, the method including receiving, from an edge device, a sample of compressed data and a sample of raw data that has not been compressed, and the sample of compressed data and the sample of raw data are unlabeled, decompressing the compressed data to obtain decompressed data, and classifying, with an ML (machine learning) model, the decompressed data, using the ML model and the raw data to update a compression quality parameter, and transmitting the compression quality parameter to the edge device, and the compression quality parameter is usable by the edge device to control compression of a subsequent sample of compressed data.