Neural Network Weight Coding With 3D-Pyramid Partitioning

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for neural network model compression have not effectively harmonized the use of 3D-Octree, 3D-Unitree, 3D-Tagtree, and 3D-Unitagtree structures with baseline methods, leading to inefficiencies in representing the local distribution of weight tensors in neural networks.

Innovation Solution

Implement a 3D-Pyramid coding method that harmonizes with baseline methods by using a 3D-Pyramid structure to partition weight tensors, employing a recursive quad-tree partitioning and context modeling for entropy coding, and enabling dependent quantization processes to optimize encoding and decoding efficiency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If multiple 3D-Octree, 3D-Unitree, 3D-Tagtree, and 3D-Unitagtree structures are used separately, then different local distribution characteristics can be represented, but the system complexity increases and harmonization with baseline methods is not achieved

Engineering Contradiction:
Improverepresentation of local distribution characteristicsVSAvoidsystem complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent combines multiple 3D tree structures (3D-Octree, 3D-Unitree, 3D-Tagtree, 3D-Unitagtree) into a unified 3D-Pyramid coding framework. This merging allows the system to maintain the ability to represent different local distribution characteristics while reducing overall system complexity through a single harmonized structure that can work alongside baseline methods.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The unified 3D-Pyramid coding structure serves multiple functions by accommodating different tree structure types within a single framework. This universal structure can adapt to represent various local distribution characteristics without requiring separate independent systems for each tree type, thereby reducing complexity while maintaining versatility.

Inventive Principle:
Principle #6Universality (Multi-functionality)

2Adaptability or versatility

If 3D-Octree, 3D-Unitree, 3D-Tagtree, and 3D-Unitagtree structures are used without harmonization, then specific local distribution patterns can be captured, but encoding efficiency decreases

Engineering Contradiction:
Improvecapture of local distribution patternsVSAvoidencoding efficiency
Core Design Contradiction:
Adaptability or versatilityVSProductivity

Solution Approach 1:

By merging multiple tree structure types into a unified 3D-Pyramid coding framework, the patent achieves both pattern capture capability and encoding efficiency. The unified structure eliminates the overhead of managing multiple separate structures while maintaining the ability to represent diverse local distribution patterns through appropriate selection and application of different tree types within the framework.

Inventive Principle:
Principle #5Merging (Combining)

3Device complexity

If baseline methods are used for neural network model compression, then simplicity is maintained, but the local distribution of weight tensors is not effectively represented

Engineering Contradiction:
Improvemethod simplicityVSAvoidrepresentation of local weight distribution
Core Design Contradiction:
Device complexityVSLoss of information

Solution Approach 1:

The patent applies segmentation by dividing the weight tensor representation into different regions that can be processed using appropriate tree structures. This segmentation allows baseline methods to be used for general compression while 3D tree structures are applied specifically to capture local distribution characteristics, thereby maintaining simplicity where possible while improving information representation where needed.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The unified 3D-Pyramid coding framework enables local quality improvement by allowing different tree structure types to be applied to different regions of the weight tensor based on their specific distribution characteristics. This ensures that local distribution information is effectively represented without requiring complex methods throughout the entire model.

Inventive Principle:
Principle #3Local quality

Data Source

PatentEP4049128B1Method and apparatus for unification based coding for neural network model compression
Publication Date: 2025.12.31 TENCENT AMERICA LLC
  • EP4049128B1 patent drawingFigure 1
  • EP4049128B1 patent drawingFigure 2
  • EP4049128B1 patent drawingFigure 3

AI summary

A method of unification based coding for neural network model compression is performed by at least one processor and includes receiving a layer uniform flag indicating whether a quantized weight of an input neural network is encoded using a uniform coding method, and determining whether the quantized weight is encoded using the uniform coding method, based on the received layer uniform flag. The method further includes, based on the quantized weight being determined to be encoded using the uniform coding method, encoding the quantized weight, using the uniform coding method, and based on the quantized weight being determined to not be encoded using the uniform coding method, encoding the quantized weight, using a non-uniform coding method.