A neural network quantization method and related apparatus

By constructing a lookup table for quadratic powers and a floating-point weight mapping, and encoding quantization weights, the problems of low quantization accuracy and low accelerator efficiency in neural networks are solved, realizing an efficient quantization method and accelerator design.

CN116306775BActive Publication Date: 2026-07-07HUNAN GOKE MICROELECTRONICS CO LTD
View PDF 1 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
HUNAN GOKE MICROELECTRONICS CO LTD
Filing Date
2023-02-15
Publication Date
2026-07-07

AI Technical Summary

Technical Problem

Existing neural network quantization methods suffer from low quantization accuracy and low accelerator efficiency, especially when deployed on edge devices, making it difficult to meet the requirements for efficient computing.

Method used

A lookup table is constructed using power terms. The quantization weights are encoded by mapping floating-point weights to power functions and then encoded using the lookup table address. Combined with unsigned quantization feature maps and fixed-point multiplication, the quantization accuracy and accelerator efficiency are improved.

Benefits of technology

It improves quantization accuracy and accelerator efficiency, reduces the memory footprint of weight parameters, simplifies mapping operations, and designs a more flexible and efficient accelerator.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN116306775B_ABST
    Figure CN116306775B_ABST
Patent Text Reader

Abstract

The application discloses a neural network quantization method, comprising the following steps: constructing a power-of-two term according to a preset quantization weight bit number and a power-of-two term number; constructing a power-of-two function according to the power-of-two term; establishing a mapping between a floating-point weight and the power-of-two function to obtain a quantization weight; constructing a lookup table corresponding to the power-of-two term; an element in the lookup table is a shift value in the corresponding power-of-two term; encoding the quantization weight to obtain an encoded weight according to the shift value in the address of the lookup table; quantizing a floating-point feature map into an unsigned quantization feature map according to the quantization weight and a first preset formula; and calculating a fixed-point multiplication of the quantization weight and the unsigned quantization feature map to obtain an intermediate result of convolution according to the encoded weight and a second preset formula. The method can improve quantization precision and calculation efficiency. The application also discloses a neural network quantization device, an accelerator and a computer readable storage medium, which have the above technical effects.
Need to check novelty before this filing date? Find Prior Art