Title:
NEURAL NETWORK PARAMETER QUANTIFICATION METHOD AND APPARATUS
Document Type and Number:
WIPO Patent Application WO/2023/231794
Kind Code:
A1
Abstract:
Provided in the present application are a neural network parameter quantification method and apparatus in the field of artificial intelligence, which method and apparatus are used for quantifying a neural network, and reducing precision losses during low-bit quantization, so as to obtain a lightweight model with more accurate outputs. The method comprises: firstly, acquiring parameters of neurons in a model to be quantified, so as to obtain a parameter set; then, clustering the parameters in the parameter set, so as to obtain a plurality of kinds of classification data; and quantifying each kind of classification data among the plurality of kinds of classification data, so as to obtain at least one quantification parameter, wherein the at least one quantification parameter is used for obtaining a compression model, and the precision of the at least one quantification parameter is lower than the precision of the parameters in the model to be quantified.
More Like This:
Inventors:
NIE YING (CN)
HAN KAI (CN)
LIU CHUANJIAN (CN)
MA JUNHUI (CN)
WANG YUNHE (CN)
HAN KAI (CN)
LIU CHUANJIAN (CN)
MA JUNHUI (CN)
WANG YUNHE (CN)
Application Number:
PCT/CN2023/095019
Publication Date:
December 07, 2023
Filing Date:
May 18, 2023
Export Citation:
Assignee:
HUAWEI TECH CO LTD (CN)
International Classes:
G06N3/04
Foreign References:
| CN115081588A | 2022-09-20 | |||
| CN110309904A | 2019-10-08 | |||
| CN113396427A | 2021-09-14 | |||
| CN110874627A | 2020-03-10 | |||
| CN113222098A | 2021-08-06 | |||
| CN109859281A | 2019-06-07 | |||
| US20200250539A1 | 2020-08-06 |
Attorney, Agent or Firm:
SHENPAT INTELLECTUAL PROPERTY AGENCY (CN)
Download PDF: