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

A redistribution method for multi-view classification

A multi-view and view technology, applied in the field of multi-view classification, can solve problems such as unsatisfactory performance and inability to fully explore the interaction of different views

Active Publication Date: 2022-07-08
TIANJIN UNIV
View PDF5 Cites 0 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

[0006] Although existing methods have achieved great success in multi-view learning, there are still some major limitations
First, most existing multi-view methods assume that different views of the data are highly correlated and maintain consistency by maximizing the correlation between views, but there is usually consistent and complementary information between multiple views. , overemphasizing only one side may make the performance unsatisfactory
Second, most existing models usually construct the model in each view and then constrain it to consistency or complementarity among multiple views, which cannot fully explore the interaction between different views

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
  • A redistribution method for multi-view classification
  • A redistribution method for multi-view classification
  • A redistribution method for multi-view classification

Examples

Experimental program
Comparison scheme
Effect test

Embodiment 1

[0033] 1. Given a set of multi-view samples

[0034] Suppose a set of multi-view samples is represented as in is the feature matrix of the vth view, is the label set of all samples, V represents the number of views, y N Indicates the label of the Nth sample, N represents the number of samples, d v represents the feature space dimension of the vth view, means size d v ×N space. Define an integration network F(x;Θ f ), a merged spatial network G(x;Θ g ), multiple attention networks represents the mth attention network), and is randomly initialized.

[0035] Among them, x is the sample; Θ f refers to the parameters in the integrated network, Θ g refers to the parameters in the combined spatial network; is the parameter in the attention network corresponding to the mth learner.

[0036] 2. Integrate the network F(x;Θ f ) integrate features from multiple views to obtain a unified representation

[0037] 3. The unified representation that will be obtained In...

Embodiment 2

[0046] The scheme in Embodiment 1 is further introduced below in conjunction with specific calculation formulas and examples, and is described in detail below:

[0047] 1. Data

[0048] Get a multi-view dataset, assuming a set of multi-view samples represented as in is the feature matrix of the vth view, is the label set of all samples, V represents the number of views, N represents the number of samples, d v Represents the feature space dimension of the vth view, y N is the label of the Nth sample.

[0049] Second, the construction of the loss function

[0050] According to the technical solution, a loss function suitable for multi-view data classification is constructed, including the following steps:

[0051] 1) Define the integration network F(x;Θ f ), where x is the input sample, Θ f Refers to the parameters in the integrated network. Integrate features from multiple views for a unified representation

[0052] 2) Define multiple attention networks, where ...

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 redistribution method for multi-view classification, which includes defining an integration network, a combined spatial network, and multiple attention networks, and randomly initializing the parameters of the networks; the integration network integrates features from multiple views to obtain Unified representation; input the unified representation into multiple attention networks, and output multiple weight vectors as multiple sets of attention masks; then multiply the unified representation and the attention mask element by element to select useful features, forming the mth viewz m ;Define preset rules for z m Update to obtain the mth pseudo-view in the public space; splicing the implicit representations of multiple pseudo-views learned by different learners to obtain a complete representation to construct a loss function, update the integration network, merge spatial network, pay attention to the loss function according to the loss function The parameters of the force network; if the loss function has converged, the training ends. The present invention explores the consistency and diversity among multiple views in a unified framework, so that the classification performance is significantly improved.

Description

technical field [0001] The present invention relates to the field of multi-view classification, in particular to a redistribution method for multi-view classification. Background technique [0002] In practical applications, data can be described from different perspectives (multi-view), for example, a video usually includes voice signals, text information and image information; a piece of news usually includes text information and image information; even data in a single modality , and different types of descriptions can also be extracted to characterize different attributes. It is of great significance to study how to better integrate the information of multiple views in real life. In medical data, the diagnosis of diseases usually needs to combine the results of multiple medical detection methods; in the field of autonomous driving, the perception of the external environment by the car requires Fusion of results from multiple sensors. Therefore, multi-view learning has ...

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 Patents(China)
IPC IPC(8): G06V10/774G06V10/40G06V10/82G06N3/04G06N3/08G06K9/62
CPCG06N3/084G06V10/40G06N3/045G06F18/214
Inventor 张长青张宇桐付海娟
Owner TIANJIN UNIV