NEURAL NETWORK POST-PROCESSING METHOD AND APPARATUS, CHIP, ELECTRONIC DEVICE, AND STORAGE MEDIUM

WIPO Patent Application WO/2023/004800

A1

The present application relates to the field of artificial intelligence, and provided in an embodiment thereof is a neural network post-processing method. In response to receiving a convolution result of a neural network layer, an address table entry that matches the neural network layer is looked up in a pre-constructed lookup table on the basis of channel information of said neural network layer; an activation and quantization parameter corresponding to the convolution result is selected from within a matched address table entry; and then an activation and quantization value corresponding to the convolution result is calculated on the basis of the convolution result and the selected activation and quantization parameter. Further provided in the present application are a neural network post-processing apparatus, a chip, an electronic device, and a computer-readable storage medium. Activation and quantization are merged together, the expressive ability and adaptability of a neural network model are able to be enhanced, and computing power costs are reduced.

JP2024044810	Evaluation program, evaluation method and evaluation device
WO/2016/127045	DISTRIBUTED TRAINING OF REINFORCEMENT LEARNING SYSTEMS
WO/2021/086688	SMART-DEVICE-BASED RADAR SYSTEM PERFORMING GESTURE RECOGNITION USING A SPACE TIME NEURAL NETWORK

LIU GENSHU (CN)
XIAO YANNAN (CN)
ZHANG YIHAO (CN)
ZUO WENMING (CN)
ZHOU YUCONG (CN)
ZHONG ZHAO (CN)

PCT/CN2021/109848

February 02, 2023

July 30, 2021

Click for automatic bibliography generation  

HUAWEI TECH CO LTD (CN)

G06N3/04; G06N3/063

CN109389212A	2019-02-26
CN109146057A	2019-01-04
CN110610237A	2019-12-24
US20210216859A1	2021-07-15

SHENZHEN SCIENBIZIP INTELLECTUAL PROPERTY AGENCY CO., LTD. (CN)

View/Download PDF      PDF Help

Previous Patent: ELECTRONIC DEVICE AND NEURAL NETWORK QUANTIZATION METHOD

Next Patent: TASK PROCESSING METHOD AND APPARATUS