Kernel Tensor Tree Partitioning for Neural Network Compression
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video coding systems, such as HEVC, have limitations in achieving higher coding efficiency, and there is a need for efficient compression of neural networks (NNs) used in AI applications, particularly in multimedia signal processing, to enable interoperable and interpretable neural network representation (NNR) standards.
Innovation Solution
A kernel tensor and tree partition-based neural network compression framework is employed, utilizing multi-tree and mixed-tree coding techniques to optimize NN compression, enabling efficient storage and transmission of NN models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If HEVC standard is used for video coding, then coding efficiency is improved by 50% bit-rate saving, but further coding efficiency improvement is limited
Solution Approach 1:
The patent segments the neural network model into multiple layers and further divides each layer into kernel tensors and channel data. This segmentation allows independent compression of different components using optimized algorithms, achieving higher overall compression efficiency beyond what HEVC can provide while maintaining manageable complexity through modular processing
Solution Approach 2:
The patent transforms the traditional 2D block-based HEVC coding approach into a multi-dimensional neural network-specific coding framework. By treating kernel tensors as 3D or 4D structures and applying tensor decomposition, the system exploits additional dimensionalities for more efficient compression while keeping the base HEVC framework intact
2Quantity of substance
If neural networks are compressed for storage and transmission, then storage and bandwidth requirements are reduced, but compression efficiency and interoperability are challenging
Solution Approach 1:
The patent changes the representation parameters of neural network data by separating kernel tensors from channel data and applying different compression parameters to each. Kernel tensors use tensor decomposition with specific rank parameters, while channel data uses conventional vector quantization, optimizing compression efficiency for each data type's characteristics
Solution Approach 2:
The patent creates a universal compression framework that can handle different types of neural network layers and architectures through the same basic methodology. The multi-tree coding structure and tensor decomposition approach work across various NN configurations, achieving both high compression efficiency and interoperability
3Productivity
If kernel tensor and tree partition based compression is applied, then coding efficiency is improved, but computational complexity increases
Solution Approach 1:
The patent implements dynamic tree partitioning where the partitioning structure adapts to the specific characteristics of each kernel tensor. Rather than using fixed partition sizes, the system dynamically determines optimal partition boundaries based on tensor properties, improving compression efficiency while avoiding unnecessary computational complexity through adaptive rather than exhaustive processing
Data Source
AI summary
A method of encoding or decoding a video comprising a current picture, a first reference picture, and a weight tensor associated with a trained neural network (NN) model are provided. The method includes generating any number of kernel tensors, input channels and output channels associated with the weight tensor, each kernel tensor being associated with any of: a layer type, an input signal type, and a tree partition type, and each kernel tensor including weight coefficients, generating, for each of the any number of kernel tensors, tree partitions for any of a coding tree unit (CTU), a coding unit (CU), a prediction unit (PU), and a transform unit (TU) according to respective tree partition types associated with each of the any number of kernel tensors, and generating a compressed representation of the trained NN model by compressing and coding the any number of kernel tensors.