Tensor Bitstream Encoding with Mean Feature Quantization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video compression technologies, particularly in collaborative intelligence architectures involving convolutional neural networks (CNNs), face challenges in efficiently compressing tensor data for transmission and storage, leading to high computational demands and resource consumption, especially in edge devices and cloud implementations.
Innovation Solution
A method and system for encoding and decoding tensors using mean feature and basis vector quantization, involving generating mean features, quantizing and inverse-quantizing these features, and encoding them into a bitstream, along with basis vector encoding and decoding, to reduce the computational burden and improve efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If existing video compression technologies are used for tensor data, then transmission and storage are supported, but computational demands and resource consumption are high
Solution Approach 1:
The tensor is segmented into multiple feature maps, which are then processed independently through the compression pipeline. Each feature map undergoes separate quantization and encoding, allowing parallel processing and reducing overall computational burden while maintaining compression efficiency
Solution Approach 2:
The patent extracts and encodes only the most significant features from each feature map using selective quantization. By identifying and retaining only the most important tensor elements for reconstruction, the system reduces computational demand while preserving essential information for accurate decompression
2Loss of substance
If tensor data is compressed for transmission, then bandwidth usage is reduced, but reconstruction accuracy may be compromised
Solution Approach 1:
The patent dynamically adjusts quantization parameters based on the importance of different feature maps and their contribution to the final output. By varying the compression aggressiveness according to parameter importance, the system achieves efficient compression while maintaining reconstruction accuracy for critical features
Solution Approach 2:
The system incorporates feedback mechanisms where the encoder analyzes the characteristics of the tensor data and adjusts compression parameters accordingly. This adaptive approach ensures that compression levels are optimized for each specific dataset, balancing transmission efficiency with reconstruction fidelity
3Loss of substance
If complex quantization methods are applied to tensors, then compression ratio is improved, but device complexity increases
Solution Approach 1:
The complex quantization process is divided into simpler stages, with each feature map processed through a standardized but simplified quantization routine. This segmentation allows the system to achieve good compression ratios through cumulative effect of multiple simple operations rather than one complex operation
Solution Approach 2:
Different quantization strategies are applied to different regions or feature maps based on their local characteristics and importance. Less important feature maps receive more aggressive quantization while critical features maintain higher precision, optimizing the balance between compression ratio and complexity
Data Source
AI summary
A system and method of encoding a tensor into a bitstream, the tensor having a channel dimension. The method comprises generating a mean feature of the tensor, the mean feature being an average of the tensor over the channel dimension; quantising the generated mean feature to generate a quantised mean feature map; and inverse-quantising a reconstruction of the quantised mean feature map to generate a second representation of the mean feature. The method further comprises determining mean feature coefficients using the second representation of the mean feature map and the channel dimension of the tensor; and encoding at least the quantised mean feature map and a quantised version of the mean feature coefficients into the bitstream.