Tensor Encoding and Decoding via Independent Units for Edge Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing video compression technologies, particularly in collaborative intelligence architectures, face challenges in efficiently compressing and transmitting tensor data across distributed networks due to high computational complexity and memory requirements, especially when implementing convolutional neural networks (CNNs) in edge devices and cloud environments.
Innovation Solution
A method and system for encoding and decoding tensors by decoding and encoding units of information independently, including an average tensor value, decomposition components, and coefficients, arranged in two-dimensional arrays, to facilitate efficient compression and transmission across distributed systems.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If tensor data is compressed using existing video compression technologies, then transmission efficiency is improved, but computational complexity and memory requirements increase
Solution Approach 1:
The tensor is divided into multiple non-overlapping blocks, where each block is processed independently through the compression pipeline. This segmentation allows parallel processing of blocks, reducing overall computational complexity while maintaining transmission efficiency. The encoder processes each block separately, and the decoder reconstructs each block independently, enabling distributed computation across multiple processing units.
Solution Approach 2:
The mean value of each tensor block is extracted and encoded separately from the residual data. By taking out the mean component, the residual values have zero mean, which improves the effectiveness of subsequent transform coding and quantization. This extraction simplifies the compression process by handling the DC component separately, reducing the computational burden on the main compression pipeline.
2Loss of substance
If tensor data is compressed with higher compression ratios, then bit usage is optimized, but transmission reliability decreases due to varying bitrates
Solution Approach 1:
The compression algorithm dynamically adjusts quantization parameters and transform block sizes based on the statistical properties of each tensor block. Blocks with higher variance use finer quantization, while blocks with lower variance use coarser quantization. This dynamic adaptation ensures that the compression ratio is optimized for each block's characteristics, maintaining transmission reliability while optimizing overall bit usage.
Solution Approach 2:
The encoder changes multiple parameters including quantization step size, transform block dimensions, and prediction mode selection based on the local characteristics of each tensor block. These parameter changes allow the system to adapt to varying data statistics, ensuring reliable reconstruction quality across different bitrate conditions while optimizing compression efficiency for each specific block.
3Loss of substance
If complex decomposition methods are used to reduce tensor dimensionality, then data transmission volume is reduced, but encoding and decoding time increases
Solution Approach 1:
The tensor decomposition is performed on segmented blocks rather than on the entire tensor. Each block undergoes independent mean extraction and transform coding, which significantly reduces the computational complexity of the decomposition process. This block-based approach enables parallel processing and reduces encoding/decoding time while achieving the same dimensionality reduction effect as full-tensor methods.
Solution Approach 2:
The patent applies transform coding and quantization to only the necessary components of the tensor data (residuals after mean extraction), rather than processing the entire tensor with full decomposition. This partial action approach achieves sufficient dimensionality reduction for transmission while avoiding the excessive computational overhead of complete tensor decomposition, thus reducing encoding and decoding time.
Data Source
AI summary
A method of decoding a tensor from a bitstream. The method comprises decoding a first unit of information from the bitstream, decoding a second unit of information from the bitstream, decoding a third unit of information from the bitstream, and decoding a fourth unit of information from the bitstream. The method further comprises producing the tensor on a channel-wise basis by multiplying the coefficients for a respective channel with the average tensor value and the components of the decomposition of the tensor with coefficients associated with the decomposition of the tensor, wherein at least one of the first, second, third, and fourth units of information from the bitstream are independently decodable with respect to at least one of the other units of information and each unit of information is arranged in a plurality of two-dimensional arrays of samples.