Adaptive Predictive Compression With Lockstep Model Updates

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing lossless compression techniques fail to leverage the predictive power of large-scale models effectively due to suboptimal model predictions when data distribution differs from training data, and they require transmission of model updates or side information, leading to inefficiencies and synchronization issues across different computing platforms.

Innovation Solution

A shared predictive model is instantiated at both encoder and decoder, synchronously updated using only decoded data, ensuring lockstep adaptation without transmitting model parameters or side information, utilizing adapter modules and deterministic update rules to maintain synchronization.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a frozen pretrained language model is used for compression, then the model provides strong static prediction capability, but the model cannot adapt to data streams with distributions different from training data

Engineering Contradiction:

Improveprediction accuracyVSAvoidadaptability to different data distributions

Core Design Contradiction:

Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent transforms the static frozen model into a dynamic adaptive system by enabling online parameter updates at the decoder and synchronized parameter updates at the encoder, allowing the model to continuously adapt to the specific data stream being compressed while maintaining strong prediction capabilities

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The decoder performs self-service by updating its model parameters locally using the decoded data without requiring retraining or external intervention, enabling autonomous adaptation to the data distribution

Inventive Principle:

Principle #25Self-service

2Productivity

If model parameters are updated online at the encoder, then compression efficiency improves, but the decoder cannot mirror the encoder state without transmitting side information

Engineering Contradiction:

Improvecompression efficiencyVSAvoidmodel state synchronization

Core Design Contradiction:

ProductivityVSLoss of information

Solution Approach 1:

The patent establishes a feedback loop where the decoder uses decoded data to update its parameters, and these updated parameters are then used to encode subsequent data, creating a synchronized adaptation process between encoder and decoder without information loss

Inventive Principle:

Principle #23Feedback

Solution Approach 2:

Instead of transmitting model updates from encoder to decoder, the patent inverts the approach by having the decoder independently update its parameters using decoded data, eliminating the need for side information transmission

Inventive Principle:

Principle #13The other way round (Inversion)

3Manufacturing precision

If learned compression systems fine-tune models on a per-instance basis, then rate-distortion performance improves, but transmission overhead increases due to side information requirements

Engineering Contradiction:

Improverate-distortion performanceVSAvoidtransmission overhead

Core Design Contradiction:

Manufacturing precisionVSQuantity of substance

Solution Approach 1:

The patent extracts only the essential parameter updates needed for adaptation and transmits them efficiently, or alternatively, eliminates the need for transmission by having the decoder independently perform the same updates using decoded data

Inventive Principle:

Principle #2Taking out (Extraction)

Solution Approach 2:

The system performs partial updates on selected model parameters rather than full model retraining, achieving sufficient adaptation for the specific data stream while minimizing computational overhead and transmission requirements

Inventive Principle:

Principle #16Partial or excessive action

4Adaptability or versatility

If neural model compression schemes are deployed across different hardware platforms, then compression capability improves, but cross-platform determinism fails due to floating-point arithmetic differences

Engineering Contradiction:

Improvecompression capabilityVSAvoidcross-platform determinism

Core Design Contradiction:

Adaptability or versatilityVSReliability

Solution Approach 1:

The patent changes the numerical parameters and computational precision settings to ensure deterministic behavior across platforms, using controlled precision arithmetic and standardized operations that produce consistent results regardless of hardware differences

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS20260081620A1Online adaptive lossless compression with lockstep-trained predictive models

Publication Date: 2026.03.19 BUSINESS MOBILE AG

AI summary

A method and system are disclosed for adaptive lossless data compression using a predictive model that is updated deterministically in lockstep at both encoder and decoder. An encoder generates symbol probability distributions from a frozen base model with a small set of updatable parameters, encodes input blocks via entropy coding, and applies deterministic parameter updates using the observed data. A decoder entropy-decodes the compressed bitstream, reconstructs the same blocks, and applies identical updates, thereby maintaining synchronization without transmission of model parameters. Optional features include periodic beacons for state verification and resynchronization, error-control metadata, and modality-specific tokenization. In preferred embodiments, parameter adaptation is confined to adapter modules within a largely fixed neural network, ensuring computational efficiency and reproducibility across platforms. Over time, the model specializes to the stream, reducing average coding rate while guaranteeing exact reconstruction of the original data.