Unlock instant, AI-driven research and patent intelligence for your innovation.

Method and device for training neural network for image recognition

A neural network and image recognition technology, applied in neural learning methods, biological neural network models, probabilistic networks, etc., can solve problems such as large DNN models

Pending Publication Date: 2022-07-19
SAMSUNG ELECTRONICS CO LTD
View PDF1 Cites 0 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

[0004] Furthermore, state-of-the-art DNN models may be too large and may not be sufficient to be implemented in limited usage environments such as mobile devices

Method used

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
View more

Image

Smart Image Click on the blue labels to locate them in the text.
Viewing Examples
Smart Image
  • Method and device for training neural network for image recognition
  • Method and device for training neural network for image recognition
  • Method and device for training neural network for image recognition

Examples

Experimental program
Comparison scheme
Effect test

example 310

[0074] The ten instances 310 may be instances where different mixed-precision quantization is applied to the neural network (eg, ResNet50), and for ease of description, it may be assumed that the ten instances 310 satisfy the convergence criteria. exist image 3 In the instances 310 shown in , the vertical axis may indicate different instances, and the horizontal axis may indicate the layers included in each instance. The numbers in the boxes that indicate each layer indicate precision. For example, a number in a box may indicate the precision of the layer of the corresponding instance. For example, "4" may indicate INT4 precision, "8" may indicate INT8 precision, and "16" may indicate INT16 precision. image 3 Some layers of instances 310 shown in can converge to a predetermined accuracy (eg, INT4), and ten instances 310 selected to be best suited for at least one of the multiple objectives can have convergence characteristics (eg, convergent layer accuracy ). The neural ...

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

PUM

No PUM Login to View More

Abstract

The invention discloses a method and equipment for training a neural network for image recognition. The method comprises the following steps: receiving training image data; pre-training a neural network by using the training image data; applying different mixing precision quantization to the pre-trained neural network to obtain a population comprising a plurality of instances; determining whether a portion of the population satisfies convergence criteria; in response to determining that the portion satisfies convergence criteria, generating a new instance using the portion; updating the population by adding a new instance to the population; and selecting, from the updated population, an instance to which the optimized hybrid precision quantization of the neural network is applied as a trained neural network.

Description

[0001] This application claims the benefit of Korean Patent Application No. 10-2021-0004075, filed in the Korean Intellectual Property Office on Jan. 12, 2021, the entire disclosure of which is incorporated herein by reference for all purposes. technical field [0002] The following description relates to methods and apparatus for training neural networks for image recognition. Background technique [0003] Multi-objective optimization (MOO) can be an important and practical task in hardware (and / or hardware-implemented software) design because multi-objective optimization can enable the generation of models, architectures that can satisfy multiple objectives simultaneously and device. In order to optimize deep neural network (DNN) models for image recognition in terms of conflicting goals (eg, accuracy, memory capacity, latency, power consumption, etc.), attempts have been made to apply MOO methods. [0004] Furthermore, prior art DNN models may be too large and may not be...

Claims

the structure of the environmentally friendly knitted fabric provided by the present invention; figure 2 Flow chart of the yarn wrapping machine for environmentally friendly knitted fabrics and storage devices; image 3 Is the parameter map of the yarn covering machine
Login to View More

Application Information

Patent Timeline
no application Login to View More
Patent Type & Authority Applications(China)
IPC IPC(8): G06N3/04G06N3/08G06N3/00
CPCG06N3/04G06N3/08G06N3/006G06N3/126G06N7/01G06N3/045G06N3/048G06N3/086G06N3/047
Inventor 伊霍尔·瓦希尔特索夫张祐锡
Owner SAMSUNG ELECTRONICS CO LTD