Neural Network Quantization Parameter Adjustment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural networks with limited calculation resources face challenges in maintaining recognition accuracy due to the trade-off between quantization precision and computational efficiency, where small quantization parameters lead to output truncation and large parameters result in low resolution, causing deterioration in recognition accuracy.
Innovation Solution
An information processing apparatus and method that adjusts the output size of neural network operations based on the output distribution and quantization parameter, using a small quantization parameter for intermediate layers to suppress recognition accuracy deterioration without increasing the quantization parameter, thereby reducing the risk of output truncation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If a small quantization parameter is used for quantization, then computational efficiency is improved, but output truncation occurs leading to recognition accuracy deterioration
Solution Approach 1:
The patent dynamically adjusts the quantization parameter based on the output distribution characteristics of each layer. By calculating statistics (mean and standard deviation) of the output data and using them to adaptively set the quantization parameter, the system optimizes the balance between computational efficiency and recognition accuracy, preventing both truncation and resolution loss.
2Reliability
If a large quantization parameter is used for quantization, then recognition accuracy is maintained, but output resolution decreases leading to recognition accuracy deterioration
Solution Approach 1:
The system calculates the output distribution statistics for each layer and uses these to dynamically determine the quantization parameter. This adaptive approach ensures that each layer uses an optimal quantization parameter that maintains sufficient resolution while avoiding unnecessary precision that would reduce computational efficiency.
3Reliability
If individual layer-specific quantization parameters are used, then recognition accuracy is maintained, but device complexity and processing load increase
Solution Approach 1:
The patent implements a method where quantization parameters are adaptively determined based on output distribution statistics calculated during the forward propagation phase. This approach maintains recognition accuracy by optimizing parameters for each layer's specific characteristics while avoiding the need for complex separate optimization processes for each layer, as the parameters are derived automatically from the data distribution.
Data Source
AI summary
There is provided with an information processing apparatus. An obtaining unit obtains information indicating a size of an output as a result of a first operation in a neural network that performs the first operation using a weight coefficient for input data and a second operation of quantizing a result of the first operation, in order to obtain data of an intermediate layer. A control unit controls the first operation in the neural network to adjust the size of the output based on the information and a quantization parameter used for the quantization.