Looking for breakthrough ideas for innovation challenges? Try Patsnap Eureka!

Capsule network training method, classification method and system, equipment and storage medium

A technology of training method and classification method, which is applied in the direction of biological neural network model, text database clustering/classification, special data processing application, etc. It can solve multi-label text without providing iterative objective function, fixed number of iterations, capsule network training effect In order to solve problems such as poor classification effect, the number of iterations can be flexibly determined and the classification effect is better.

Inactive Publication Date: 2019-07-23
SHENZHEN INST OF ADVANCED TECH CHINESE ACAD OF SCI
View PDF0 Cites 14 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

[0005] The purpose of the present invention is to provide a training method, classification method, system, equipment and storage medium of a capsule network, aiming to solve the problems existing in the prior art due to the fact that the iterative objective function is not provided in the dynamic routing mechanism and the number of iterations is fixed. The capsule network training effect and the poor effect of multi-label text classification

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
  • Capsule network training method, classification method and system, equipment and storage medium
  • Capsule network training method, classification method and system, equipment and storage medium
  • Capsule network training method, classification method and system, equipment and storage medium

Examples

Experimental program
Comparison scheme
Effect test

Embodiment 1

[0039] In this embodiment, the capsule network includes such as figure 1 Structure shown:

[0040] The convolution layer 101 is used to extract feature scalar from the training sample text and construct the feature scalar group.

[0041] Specifically, the training sample text may be the text corresponding to the Wikipedia webpage, the text corresponding to the customer comments, the text corresponding to the biological gene sequence, or the text corresponding to the disease diagnosis certificate. Then, using the trained capsule network, it can finally be used for the label classification of the corresponding scene, such as: the classification of heterogeneous labels on Wikipedia web pages, the ratings of different aspects of customer reviews, gene function prediction in biology, disease prediction or diagnosis, etc. . Especially for large-scale multi-label text data sets, the amount of data is huge and the number of tags is tens of thousands. For example, the European Union Laws a...

Embodiment 2

[0058] Based on the first embodiment, this embodiment further provides the following content:

[0059] The training method of this embodiment also includes:

[0060] First, compress each initial vector to obtain a compact vector;

[0061] Secondly, the compact vector is subjected to affine transformation to obtain the intermediate vector.

[0062] In other words, as figure 2 As shown, in the capsule network, between the primary capsule layer 102 and the polymer capsule layer 103, a corresponding compact capsule layer 201 is also provided. The compact capsule layer 201 is used to perform the initial vector transfer from the primary capsule layer 102. The compact vector is obtained by compression processing, and after the compact vector reaches the polymerized capsule layer 103, the intermediate vector is obtained by the affine transformation of the polymerized capsule layer 103.

[0063] The compression processing may specifically include the following content: performing denoising pro...

Embodiment 3

[0066] On the basis of this embodiment and the first or second embodiment, the following content is further provided:

[0067] The training method of this embodiment also includes:

[0068] Under the condition of minimum classification loss, the transformation parameters corresponding to the affine transformation processing and the convolution parameters corresponding to the capsule network processing are updated iteratively.

[0069] Specifically, the routing parameters used in routing processing are updated iteratively through the above iterative routing mechanism. In the capsule network, other parameters, such as the transformation matrix of affine transformation, convolution parameters, etc., can all be based on preset classification loss targets. Under the condition of minimizing the function, the corresponding iterative update is performed to complete the training of the entire capsule network.

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 method is applicable to the technical field of computers. The invention provides a training method, a classification method, a system, equipment and a storage medium of a capsule network. In the method, a capsule network is constructed by a convolution layer, a primary capsule layer, a polymerization capsule layer and a category capsule layer; in the polymerization capsule layer, affine transformation is utilized to obtain an intermediate vector, routing processing is utilized to obtain a secondary vector, and under the condition that the average distance between the secondary vector and each intermediate vector is the shortest, a routing parameter corresponding to the routing processing is dynamically updated based on an iterative routing mechanism of adaptive kernel density estimation. Therefore, an iterative routing mechanism of adaptive kernel density estimation can be utilized to provide an explicit iterative objective function, and the number of iterations is flexibly determined, so that the capsule network training effect and the multi-label text classification effect are better.

Description

Technical field [0001] The present invention belongs to the field of computer technology, and in particular relates to a training method, classification method, system, equipment and storage medium of a capsule network. Background technique [0002] The task of multi-label text classification is to classify each text to be processed into a subset of different labels. At present, traditional machine learning algorithms such as tree-based models and tag embedding methods can be used in the prior art, and convolutional neural networks or capsule networks can also be used to implement multi-label text classification. [0003] The existing use of the capsule network to realize the classification of multi-label texts is mainly that the capsule network first extracts the features of the text to obtain the feature scalar and constructs the feature scalar group, and then converts the feature scalar group into an initial vector, and then the initial vector is obtained by affine transformatio...

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
IPC IPC(8): G06F16/35G06N3/04
CPCG06F16/35G06N3/045
Inventor 杨敏曲强陈磊
Owner SHENZHEN INST OF ADVANCED TECH CHINESE ACAD OF SCI
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Patsnap Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Patsnap Eureka Blog
Learn More
PatSnap group products

Browse by: Latest US Patents, China's latest patents, Technical Efficacy Thesaurus, Application Domain, Technology Topic, Popular Technical Reports.

© 2024 PatSnap. All rights reserved.Legal|Privacy policy|Modern Slavery Act Transparency Statement|Sitemap|About US| Contact US: help@patsnap.com