Neural Network Model Compression via Segmented Access Units
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for compressing and coding video signals and neural networks are inefficient, particularly in terms of coding efficiency and performance during training and inference processes.
Innovation Solution
A neural network-based signal processing method that reconstructs neural network models from bitstreams containing access units with model, layer, and parameter information, including quantization data, to improve compression and reconstruction efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of energy
If neural network models are compressed for transmission and storage, then transmission bandwidth and storage space are reduced, but reconstruction accuracy and inference performance may deteriorate
Solution Approach 1:
The neural network model is segmented into multiple access units (AUs) with different priority levels. High-priority AUs contain critical model parameters essential for accurate reconstruction, while low-priority AUs contain less critical data. This segmentation allows the system to transmit only necessary portions at different quality levels, balancing bandwidth usage with reconstruction accuracy.
Solution Approach 2:
The patent applies quantization to transform precise model parameters into discrete levels, reducing the precision of parameter representation. By controlling the quantization precision dynamically based on priority levels and available bandwidth, the system achieves compression while maintaining acceptable reconstruction accuracy for critical parameters.
2Manufacturing precision
If quantization precision is increased for better model reconstruction, then coding efficiency decreases and larger bitstream size is required
Solution Approach 1:
Different quantization precisions are applied to different access units based on their priority levels. High-priority AUs use higher quantization precision to ensure accurate reconstruction of critical model components, while low-priority AUs use lower precision to reduce bitstream size. This local differentiation optimizes the overall balance between reconstruction quality and compression ratio.
3Reliability
If all neural network access units are transmitted to ensure complete model reconstruction, then transmission time and bandwidth consumption increase
Solution Approach 1:
The system performs preliminary classification of access units into priority groups before transmission. Critical AUs are identified and marked for mandatory transmission, while non-critical AUs are marked as optional. This preliminary action enables the receiver to reconstruct functional models using only the essential AUs, reducing transmission time while maintaining sufficient reliability for practical applications.
Data Source
AI summary
A neural-network-based signal processing method and apparatus according to the present invention may: receive a bitstream including information about a neural network model, wherein the bitstream includes at least one neural network access unit; obtain information about the at least one neural network access unit from the bitstream; and reconstruct the neural network model on the basis of the information about the at least one neural network access unit.