Hierarchical Tensor Coding for Multi-Scale Feature Map Transmission
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing CNN architectures face challenges in efficiently compressing and transmitting tensor data across distributed systems, leading to increased computational complexity and power consumption, especially when implemented in edge devices with limited capabilities.
Innovation Solution
A method and system for encoding and decoding tensors using a hierarchical representation, where tensors with different spatial resolutions are processed separately to form a multi-scale feature pyramid network, allowing for efficient compression and transmission across edge devices and cloud servers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Power
If CNN processing is distributed across edge devices and cloud servers, then computational capability is improved, but tensor compression and transmission complexity increases
Solution Approach 1:
The patent segments the CNN processing into multiple stages, with different spatial resolution tensors processed separately through a hierarchical representation. This divides the complex tensor compression problem into manageable segments that can be handled by edge devices, reducing overall compression complexity while maintaining distributed computational capability.
Solution Approach 2:
The patent introduces a hierarchical dimension to tensor representation, organizing tensors by spatial resolution levels. This dimensional organization enables more efficient compression strategies at each level, reducing the complexity of tensor transmission in distributed systems while preserving computational power.
2Productivity
If tensors are compressed for transmission, then data transmission efficiency is improved, but spatial detail and accuracy are lost
Solution Approach 1:
The patent segments tensors into multiple hierarchical levels based on spatial resolution. Each level is compressed independently, allowing selective preservation of spatial details at different scales. This segmentation enables efficient compression while maintaining accuracy by preserving important spatial information at appropriate resolution levels.
Solution Approach 2:
The patent applies different compression strategies to different spatial resolution levels within the hierarchical tensor representation. Higher resolution tensors retain more spatial detail while lower resolution tensors undergo more aggressive compression, optimizing the balance between transmission efficiency and spatial accuracy for each local region of the data hierarchy.
3Measurement precision
If high-resolution tensors are processed, then spatial detail is preserved, but computational load on edge devices increases
Solution Approach 1:
The patent segments the computational workload by processing tensors at multiple hierarchical resolution levels. Edge devices handle lower resolution tensors requiring less computational power, while higher resolution processing is distributed to cloud servers with greater computational capacity, reducing the energy load on edge devices while preserving spatial detail where needed.
Solution Approach 2:
The patent adds a hierarchical resolution dimension to the processing architecture, enabling computational tasks to be distributed across different resolution levels. This dimensional organization allows edge devices to operate on compressed lower-resolution representations, reducing their computational and energy burden while maintaining access to high-resolution details through the hierarchical structure.
Data Source
AI summary
A method for decoding a plurality of tensors forming a hierarchical representation of feature maps for a single frame from a bitstream. The method comprises: decoding a first unit of information from the bitstream; decoding a second unit of information from the bitstream; and determining a first plurality of tensors, feature maps of at least one tensor of the first plurality of tensors having a different spatial resolution from feature maps of other tensor(s). The method also comprises determining a second plurality of tensors, feature maps of at least one tensor of the second plurality of tensors having a different spatial resolution from feature maps of other tensor(s). Feature maps of each tensor of the first plurality of tensors have different spatial resolution from feature maps of each tensor of the second plurality of tensors, and the tensors correspond to the hierarchical representation of feature maps for the single frame.