18084948. METHOD AND APPARATUS FOR GENERATING FIXED-POINT QUANTIZED NEURAL NETWORK simplified abstract (Samsung Electronics Co., Ltd.)
Contents
METHOD AND APPARATUS FOR GENERATING FIXED-POINT QUANTIZED NEURAL NETWORK
Organization Name
Inventor(s)
Junhaeng Lee of Hwaseong-si (KR)
Seungwon Lee of Hwaseong-si (KR)
Sangwon Ha of Seongnam-si (KR)
METHOD AND APPARATUS FOR GENERATING FIXED-POINT QUANTIZED NEURAL NETWORK - A simplified explanation of the abstract
This abstract first appeared for US patent application 18084948 titled 'METHOD AND APPARATUS FOR GENERATING FIXED-POINT QUANTIZED NEURAL NETWORK
Simplified Explanation
The abstract describes a method for generating a fixed-point quantized neural network. Here are the key points:
- The method involves analyzing the statistical distribution of floating-point parameter values for each channel of feature maps and kernels in a pre-trained floating-point neural network.
- Based on the statistical distribution, a fixed-point expression is determined for each channel, covering the range of floating-point parameter values statistically.
- The method also determines the fractional lengths of bias and weight for each channel among the parameters of the fixed-point expression, based on the result of a convolution operation.
- Finally, a fixed-point quantized neural network is generated, where the bias and weight for each channel have the determined fractional lengths.
Potential Applications:
- This method can be applied in various fields that utilize neural networks, such as computer vision, natural language processing, and speech recognition.
- It can be used in applications that require efficient deployment of neural networks on resource-constrained devices, such as mobile phones, IoT devices, and embedded systems.
Problems Solved:
- Neural networks typically require high computational resources due to the large number of parameters and floating-point operations involved.
- This method addresses the problem of reducing the computational complexity and memory requirements of neural networks by quantizing the parameters to fixed-point representation.
- It allows for efficient deployment of neural networks on devices with limited resources, without significant loss in accuracy.
Benefits:
- The method enables the conversion of floating-point neural networks to fixed-point representation, reducing memory usage and computational requirements.
- It allows for the deployment of neural networks on resource-constrained devices, expanding the range of applications for neural networks.
- By quantizing the parameters based on statistical distributions, the method aims to minimize the loss in accuracy compared to the original floating-point network.
Original Abstract Submitted
A method of generating a fixed-point quantized neural network includes analyzing a statistical distribution for each channel of floating-point parameter values of feature maps and a kernel for each channel from data of a pre-trained floating-point neural network, determining a fixed-point expression of each of the parameters for each channel statistically covering a distribution range of the floating-point parameter values based on the statistical distribution for each channel, determining fractional lengths of a bias and a weight for each channel among the parameters of the fixed-point expression for each channel based on a result of performing a convolution operation, and generating a fixed-point quantized neural network in which the bias and the weight for each channel have the determined fractional lengths.