Image Feature Extraction With Hybrid Quantization Bias Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network models face challenges in balancing accuracy and efficiency due to the high storage and computational demands of floating-point data types, particularly with asymmetric linear quantization introducing zero point bias and increasing computation overhead.
Innovation Solution
A hybrid quantization method is employed, using asymmetric linear quantization for positive parameter values and symmetric linear quantization for non-positive values, along with correction processes to ensure consistent quantization coefficients and reduce zero point bias.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If asymmetric linear quantization is used to improve accuracy, then quantization accuracy is improved, but zero point bias is introduced which increases computation overhead
Solution Approach 1:
The patent segments the quantization process by dividing parameters into two categories: those containing negative values and those containing only non-negative values. Different quantization strategies are applied to each segment - asymmetric linear quantization for parameters with negative values and symmetric linear quantization for parameters without negative values. This segmentation allows the system to avoid zero point bias in appropriate cases while maintaining high accuracy where needed.
Solution Approach 2:
The patent applies different quantization quality characteristics to different parts of the parameter space. Specifically, asymmetric linear quantization (higher accuracy) is applied locally to parameters requiring it, while symmetric linear quantization (lower computation overhead) is applied to parameters where it suffices. This local differentiation resolves the contradiction by optimizing each parameter's quantization approach based on its specific characteristics.
2Measurement precision
If the number of network layers and channels is increased to improve model capability, then model accuracy is improved, but storage space and computing resources are increased
Solution Approach 1:
The patent changes the data type parameter of neural network parameters from floating-point to quantized integer representations. By applying linear quantization transformation, the precision bits are reduced (e.g., from 32-bit floating point to 8-bit or 16-bit integers), directly reducing storage space requirements while maintaining acceptable model accuracy through carefully designed quantization strategies.
3Measurement precision
If floating-point data type is used to ensure high accuracy, then computation accuracy is improved, but storage space and computing resources are increased
Solution Approach 1:
The patent changes the data representation parameter from floating-point format to quantized integer format in neural network computations. This parameter change reduces the computational complexity of arithmetic operations, decreases memory bandwidth requirements, and reduces energy consumption while maintaining model accuracy through proper quantization design and calibration.
Data Source
AI summary
Provided are a method, apparatus and device for extracting image features, and a storage medium. The method includes: obtaining parameters to be quantized of a network layer in a neural network model; determining whether values of the parameters to be quantized are all positive numbers; when the values of the parameters to be quantized are all positive numbers, executing, based on an asymmetric linear quantization logic, a quantization operation on the parameters to be quantized; when the values of the parameters to be quantized are not all positive numbers, executing, based on a symmetric linear quantization logic, a quantization operation on the parameters to be quantized; and extracting features of an input image by using the neural network model for which the quantization operation has been executed.