ANN Quantization with Compensation Bias for Error Correction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Artificial neural networks (ANNs) face challenges in maintaining accuracy during quantization due to errors introduced in the process, leading to reduced performance and increased computational complexity.
Innovation Solution
A method is introduced where the expected value of the quantization error is generated as a compensation bias, which is then reflected in the operation results through a quantized ANN, thereby improving performance and reducing complexity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If quantization is applied to ANN parameters to reduce computational complexity, then device complexity and energy consumption are reduced, but manufacturing precision and reliability deteriorate due to quantization errors
Solution Approach 1:
The patent applies preliminary action by pre-calculating the expected quantization error (compensation bias) before the actual quantized operation. The compensation bias is computed offline based on the quantization step size and statistical properties of the data, then added to the quantized operation results to correct systematic errors. This approach resolves the contradiction by preparing the correction in advance, allowing low-precision quantized operations to achieve high-precision results without increasing runtime computational complexity.
2Use of energy by moving object
If quantization is applied to ANN parameters to reduce computational complexity, then energy consumption is reduced, but reliability deteriorates due to accumulated quantization errors
Solution Approach 1:
The patent implements feedback by computing the compensation bias based on the statistical characteristics of the quantization error. The system analyzes the distribution and magnitude of quantization errors during training or calibration, then feeds this information back into the inference process by adding the computed compensation bias to correct the errors. This feedback mechanism ensures that quantization errors do not accumulate and degrade reliability, while maintaining the low energy consumption benefits of quantized operations.
3Ease of operation
If standard quantization is used to simplify the operation process, then ease of operation improves, but manufacturing precision deteriorates due to uncorrected quantization errors
Solution Approach 1:
The patent introduces an intermediary element (compensation bias) that mediates between the simple quantized operation and the precise original operation. The compensation bias acts as a correction term that is easily computed and added to the quantized results, bridging the gap between simple low-precision operations and high-precision requirements. This intermediary approach maintains operational simplicity while significantly improving calculation precision without requiring complex correction procedures.
Data Source
AI summary
The computing system includes a neural network system configured to drive an artificial neural network (ANN); and a quantization system configured to quantize the ANN, wherein the quantization system is further configured to quantize parameters of the ANN to generate the quantized parameters of the ANN, generate a quantization error of the parameters of the ANN based on the parameters of the ANN and the quantized parameters, generate a compensation bias based on the quantized parameters and the quantization error of the parameters of the ANN, and transmit the generated quantized parameters and the compensation bias to the neural network system.