Container Image Layering for Selective Neural Model Layer Updates

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional containerization mechanisms inefficiently store machine-learned models in a single layer of the container image, requiring the entire container to be recreated and transmitted when updates are needed, leading to substantial computational and bandwidth costs.

Innovation Solution

Group model layers of machine-learned models based on criteria such as layer type and computational complexity, mapping them to multiple container image layers, allowing selective updates without recreating the container.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Device complexity

If machine-learned models are stored in a single layer of the container image, then the containerization mechanism is simple, but the computational and bandwidth costs are substantial when updates are needed

Engineering Contradiction:
Improvecontainerization mechanism complexityVSAvoidcomputational and bandwidth costs
Core Design Contradiction:
Device complexityVSLoss of energy

Solution Approach 1:

The patent divides the machine-learned model into multiple layers and stores each layer in a separate container image layer. This segmentation allows individual layers to be updated independently without requiring updates to the entire container, thereby reducing computational and bandwidth costs while maintaining a manageable containerization structure.

Inventive Principle:
Principle #1Segmentation

2Reliability

If the entire container is recreated when model updates are needed, then the model can be updated completely, but the bandwidth consumption increases substantially

Engineering Contradiction:
Improvemodel update completenessVSAvoidbandwidth consumption
Core Design Contradiction:
ReliabilityVSLoss of energy

Solution Approach 1:

The patent extracts individual model layers from the container image structure, allowing specific layers to be updated independently. Instead of extracting and replacing the entire container, only the necessary layer(s) are updated and transferred, significantly reducing bandwidth consumption while ensuring complete model updates when needed.

Inventive Principle:
Principle #2Taking out (Extraction)

3Productivity

If model layers are grouped and mapped to multiple container image layers, then update efficiency improves, but the containerization structure becomes more complex

Engineering Contradiction:
Improveupdate efficiencyVSAvoidcontainerization structure complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The patent implements segmentation by grouping model layers into multiple container image layers based on specific criteria. This segmentation enables parallel processing of layer updates and selective application of changes, improving update efficiency while the grouping strategy maintains structural organization rather than creating chaos.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces a new dimension to the containerization structure by mapping model layers to multiple container image layers along a vertical stacking dimension. This dimensional transformation allows for more granular control over model updates while maintaining the horizontal containerization framework, thus improving productivity without excessively increasing complexity.

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

Data Source

PatentEP4679324A1System and method for constructing container image layers based on neural network model layers
Publication Date: 2026.01.14 RED HAT LLC
  • EP4679324A1 patent drawingFigure 1A
  • EP4679324A1 patent drawingFigure 1B
  • EP4679324A1 patent drawingFigure 2

AI summary

A plurality of model layers of a machine-learned model can be grouped to obtain a plurality of model layer groupings based on one or more grouping criteria. For each model layer grouping of the plurality of model layer groupings, mapping information can be generated that maps the model layer grouping to a corresponding container image layer of a plurality of container image layers of a container image. Based on the mapping information, the model layer grouping can be stored to the corresponding container image layer of the plurality of container image layers of the container image.