Neural Network Quantization Thresholds for Circuit Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Errors occur in operation results when converting a functional model of a neural network to an arithmetic operation in a neural network circuit due to differences in operation accuracy and data format.
Innovation Solution
A neural network training device and method that generates a trained parameter with a threshold value for quantization operations, based on the difference between the operation environments of the neural network circuit and the functional model, using a floating decimal point format.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a functional model of a neural network is converted to an arithmetic operation in a neural network circuit, then the neural network can be deployed in embedded devices, but errors occur in operation results due to differences in operation accuracy and data format
Solution Approach 1:
The patent applies preliminary action by generating trained parameters and determining quantization thresholds in advance during the training phase, before deployment to the neural network circuit. The training device calculates appropriate quantization thresholds based on the statistical characteristics of input data and the specific operation environment, ensuring that the conversion from functional model to arithmetic operation maintains accuracy. This pre-processing of parameters eliminates errors that would otherwise occur during actual inference operations in embedded devices.
Solution Approach 2:
The patent employs parameter changes by dynamically adjusting quantization thresholds according to different operation environments. The training device modifies the quantization parameters (thresholds) based on the statistical properties of input data and the specific hardware constraints of the target neural network circuit. By changing these parameters during training rather than using fixed thresholds, the system adapts the functional model to match the arithmetic operation environment, thereby maintaining operation result accuracy while enabling deployment in resource-constrained embedded devices.
2Ease of manufacture
If quantization operation is performed with fixed threshold values, then the conversion process is simple, but errors occur between operation results of functional model and neural network circuit
Solution Approach 1:
The patent applies self-service by enabling the training device to automatically determine optimal quantization thresholds without requiring manual intervention or complex external calibration processes. The system uses the statistical characteristics of the training data itself to calculate appropriate thresholds, making the conversion process from functional model to arithmetic operation both simple and accurate. This automatic self-adjustment eliminates the need for complex manual tuning while ensuring operation result consistency between the functional model and the deployed neural network circuit.
Data Source
AI summary
A neural network training device is a device that trains a neural network performing an inference operation in a neural network circuit. The neural network training device includes a training unit configured to generate a trained parameter including a threshold value which is used in a quantization operation using a functional model of the neural network performing a convolutional operation and the quantization operation based on a floating decimal point format. The training unit generates the threshold value on the basis of a difference between an operation environment of the neural network circuit and an operation environment of the functional model.