Feature Map Presence Indicators for Low-Complexity Neural Decoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network systems face challenges in efficiently compressing and transmitting feature map data across distributed devices while maintaining configurability and quality, due to limited transmission resources and the need for reduced data amount.
Innovation Solution
Implementing a method for encoding and decoding feature maps using presence indicators in the bitstream to determine whether to parse or skip data, along with optional constant filling or setting side information, to optimize bitstream structure and reduce complexity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If feature map data is transmitted in distributed neural network systems, then computational workload can be distributed across devices, but transmission resources are limited and data amount increases
Solution Approach 1:
The patent extracts and transmits only the essential presence indicator information from feature map data, rather than transmitting the complete feature map. This selective extraction reduces the quantity of transmitted data while preserving the ability to reconstruct or process the feature map at the receiving end, thereby resolving the contradiction between distributing computational workload and minimizing data transmission volume.
Solution Approach 2:
The patent applies different transmission strategies to different regions or channels of the feature map based on their importance. Presence indicators identify which regions contain significant information that must be transmitted, while other regions can be reconstructed or skipped. This local differentiation optimizes the balance between data transmission volume and computational distribution effectiveness.
2Reliability
If complete feature map data is transmitted to maintain quality, then processing accuracy is preserved, but transmission complexity and resource consumption increase
Solution Approach 1:
The patent extracts only the critical presence indicator metadata from the feature map, separating essential structural information from the complete data set. This extraction maintains processing accuracy by preserving the framework needed for reconstruction while dramatically reducing transmission complexity and resource consumption.
Solution Approach 2:
The patent transforms the feature map representation by encoding presence information as compact indicators rather than transmitting full pixel or channel data. This parameter transformation changes the data from a high-dimensional complete feature map to a low-dimensional presence indicator set, reducing complexity while maintaining the ability to restore processing accuracy at the destination.
3Loss of information
If all regions of the feature map are decoded, then complete information is available, but decoding complexity and time increase
Solution Approach 1:
The patent enables the decoder to skip regions marked as absent in the presence indicator, rushing through the decoding process by only processing necessary regions. This skipping mechanism maintains information completeness for important regions while significantly reducing decoding time by avoiding unnecessary processing of redundant or reconstructible regions.
Solution Approach 2:
The patent segments the feature map decoding process into independent region-based operations guided by presence indicators. Each region can be decoded independently based on its presence status, allowing parallel processing and optimization of decoding time while ensuring complete information recovery for all present regions.
4Productivity
If presence indicators are used to skip data parsing, then transmission efficiency improves and complexity reduces, but configurability for diverse quality requirements may be limited
Solution Approach 1:
The patent implements dynamic configurability where presence indicators and decoding strategies can be adjusted based on diverse quality requirements. The system can adaptively select which regions to mark as present or absent, and can modify decoding behavior according to different application needs, thereby maintaining both high transmission efficiency and flexible adaptability to various quality constraints.
Solution Approach 2:
The patent creates a universal presence indicator framework that can serve multiple functions: it enables efficient compression, supports different quality levels, accommodates various decoding strategies, and works across different neural network architectures. This multi-functional design maintains configurability for diverse requirements while preserving transmission efficiency benefits.
Data Source
AI summary
The present disclosure relates to efficient signaling of feature map information for a system employing a neural network. In particular, at the decoder side, a presence indicator is obtained based on information parsed from a bitstream. Based on the value of the obtained presence indicator, further data related to a feature map region are parsed or the parsing is bypassed. The presence indicator may be, for instance, a region presence indicator indicating whether feature map data is included in the bitstream or may be a side information presence indicator indicating whether a side information related to the feature map data is included in the bitstream. Similarly, an encoding method, as well as encoding and decoding devices, are provided. Accordingly, feature map data may be processed more efficiently, by reducing decoding complexity, and the amount of transmitted data can be reduced by applying the bypassing.