Convolution Parameter Compression for Bandwidth-Limited CNN Chips
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Convolution operations in deep learning, particularly in convolutional neural networks, consume significant performance and bandwidth due to the large amount of data required for calculating and transmitting weight or filter coefficients.
Innovation Solution
A method and system that compress convolution parameters by determining their range, setting an effective bit number, calculating differential values, and quantifying them to a minimum effective bit, allowing for reduced data transmission and increased processing performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If original convolution parameters are transmitted directly to the convolution operation chip, then the accuracy of convolution operation is maintained, but the bandwidth consumption is high
Solution Approach 1:
The patent applies parameter changes by transforming the convolution parameters through differential calculation and quantization. The original high-precision parameters are converted into compressed representations with fewer bits by calculating differentials from a representative value and quantizing these differentials to a minimum effective bit, thereby reducing bandwidth consumption while maintaining operational accuracy
Solution Approach 2:
The patent uses copying by transmitting a representative value and compressed differential values instead of the original parameters. The convolution operation chip reconstructs the parameters by combining the representative value with the decompressed differentials, creating a functional copy that consumes less bandwidth while preserving the necessary computational accuracy
2Quantity of substance
If convolution parameters are compressed to reduce transmission amount, then bandwidth consumption is reduced, but the complexity of compression and decompression increases
Solution Approach 1:
The patent applies segmentation by dividing the parameter transmission into two parts: a representative value and compressed differential values. This segmentation allows the system to transmit only the essential information (differentials) separately from the base value, reducing overall transmission amount while keeping the compression algorithm relatively simple
Solution Approach 2:
The patent uses preliminary action by pre-calculating the representative value and preparing the compression algorithm before actual transmission. The differential calculation and quantization are performed in advance, so that during runtime, the system only needs to perform straightforward decomposition and reconstruction operations, reducing real-time computational complexity
3Productivity
If quantization is applied to differential values, then the transmission precision is reduced, but the data transmission efficiency is improved
Solution Approach 1:
The patent applies partial action by quantizing only the differential values to a minimum effective bit rather than quantizing all parameters uniformly. This selective quantization approach maintains sufficient precision for the convolution operation while maximizing compression efficiency, as the differentials typically require fewer bits than the original parameters
Data Source
AI summary
A method for compressing multiple original convolution parameters into a convolution operation chip includes steps of: determining a range of the original convolution parameters; setting an effective bit number for the range; setting a representative value, wherein the representative value is within the range; calculating differential values between the original convolution parameters and the representative value; quantifying the differential values to a minimum effective bit to obtain a plurality of compressed convolution parameters; and transmitting the effective bit number, the representative value and the compressed convolution parameters to the convolution operation chip.