Neural Network Feature Map Encoding Decoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for encoding and decoding neural network feature maps are inefficient, particularly in reconstructing features across multiple layers of deep neural networks, which affects the performance in machine-to-machine communication applications.
Innovation Solution
A method and apparatus for encoding and decoding neural network feature maps that utilize a bottom-up or top-down structure for reconstructing features, employing upsampling, downsampling, and convolution operations based on scaling factors determined by layer sizes, to efficiently decode and reconstruct features from bitstreams.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional encoding/decoding methods are used for neural network feature maps, then the process is simple, but the encoding/decoding efficiency is poor and feature reconstruction across layers is ineffective
Solution Approach 1:
The feature map reconstruction process is segmented into multiple stages corresponding to different neural network layers. Each layer's features are reconstructed separately using bottom-up or top-down structure, with upsampling and downsampling operations applied at appropriate stages. This segmentation enables efficient processing by handling each layer's features independently while maintaining overall system organization.
Solution Approach 2:
The patent introduces a hierarchical dimension to feature reconstruction by organizing layers from bottom-up or top-down sequences. Features at different resolution levels are processed through upsampling (expanding spatial dimensions) and downsampling (reducing spatial dimensions) operations, creating a multi-dimensional reconstruction space that improves encoding/decoding efficiency while managing complexity through structured progression.
2Manufacturing precision
If features are reconstructed using bottom-up or top-down structure with upsampling and downsampling, then feature reconstruction accuracy improves, but computational complexity increases
Solution Approach 1:
The encoding process performs preliminary actions by extracting and organizing features at different layers before the decoding/reconstruction process begins. The bitstream is prepared with layer-specific feature data, and reconstruction parameters (scaling factors, upsampling/downsampling operations) are pre-determined based on the neural network architecture. This preliminary preparation enables accurate feature reconstruction during decoding while reducing real-time computational complexity.
Solution Approach 2:
The patent utilizes parameter changes through scaling factors that are determined based on the ratio between feature sizes of different layers. By adjusting these scaling parameters during upsampling and downsampling operations, the system achieves accurate feature reconstruction across layers. The scaling factors are pre-calculated from layer size ratios, allowing efficient application during decoding without complex real-time computations.
3Reliability
If multiple layers of neural network features are processed, then comprehensive feature representation is achieved, but processing time increases
Solution Approach 1:
The bottom-up or top-down reconstruction structure ensures continuity of useful action by systematically processing features from one layer to the next without interruption. Each layer's features are reconstructed sequentially, with upsampling and downsampling operations continuously applied to maintain feature flow across layers. This continuous processing approach, combined with pre-determined scaling factors, enables complete multi-layer feature representation while minimizing processing time through efficient sequential execution.
Data Source
AI summary
A neural network feature decoding method and apparatus according to the present disclosure receives a bitstream including an encoded feature, decodes a feature from a bitstream, and reconstructs features corresponding to a plurality of layers of a neural network based on a decoded feature.