Rapid DRBM adaptation method based on meta-learning

A meta-learning and fast technology, applied in the field of machine learning, can solve the problems of DRBM network underfitting, the initial value of network parameters cannot reach the global optimum, etc.

Inactive Publication Date: 2021-05-28
NAT UNIV OF DEFENSE TECH
View PDF0 Cites 2 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

[0006] The purpose of the present invention is to solve the problem that the DRBM network is underfitting and the initial value of the network parameters cannot make the network reach the global optimum after training under the condition of small samples

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
  • Rapid DRBM adaptation method based on meta-learning
  • Rapid DRBM adaptation method based on meta-learning
  • Rapid DRBM adaptation method based on meta-learning

Examples

Experimental program
Comparison scheme
Effect test

example 1

[0083] In Example 1, the HRRP simulation data output by the three-dimensional aircraft model electromagnetic simulation software designed by Saibo Company is used. The three types of aircraft simulated are F-35, F-117 and P-51. The specific parameters of the aircraft are as follows: Figure 4 shown.

[0084] The radar simulation band is x-band. The frequency range is 9.5GHz-10.5GHz with a step size of 5MHz. The polarization mode is vertical polarization. The pitch angle of the target is 0°~10° with a step size of 0.1°, and the azimuth angle is 0°~90° with a step size of 0.1°. Therefore, our data set has a total of 201 frequency points, 101 elevation angles and 901 azimuth angles, that is, a total of 101×901=91001 samples, and each sample has 201 dimensions.

[0085] Figure 5It shows the changes of HRRP data during preprocessing, where the x-axis represents the data dimension and the y-axis represents the signal amplitude. (a) is the HRRP data before preprocessing, (b) is...

example 2

[0103] In Example 2, the handwritten digits (MNIST) dataset was used for experiments. The MNIST data set consists of 70,000 pictures and corresponding labels, of which 60,000 are used to train the neural network and 10,000 are used to test the neural network. Each picture is a handwritten digital picture of 0-9 with 28*28 pixels. The picture is white on a black background. Black is represented by 0, and white is represented by a floating point number between 0 and 1. The closer to 1, the whiter the color. The MNIST data set provides the label corresponding to each picture, which is given in one-hot form, that is, the label vector is a one-dimensional array with a length of 10. A sample of the MNIST dataset such as Figure 8 shown.

[0104] We form 784 pixels into a one-dimensional array with a length of 784. This one-dimensional array is the input data we want to input into the neural network. For the RBM network, the input data of the network can only be binary, so the MN...

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 belongs to the field of machine learning, and particularly relates to a DRBM method based on meta-learning, which divides an algorithm into two stages of meta-learning and model learning by improving a training-testing algorithm of a network. In the meta-learning stage, a training task is utilized to update network parameters, and the updated network parameters are used as initial values of network parameters in the model learning stage, so that the initial values of the network parameters can enable a loss function of network training to descend more quickly and achieve global optimum more easily; and network parameters are updated and tested by using the test task in the model learning stage. According to the algorithm, a meta-learning method is introduced to improve the training process of the DRBM, so that the gradient descent direction of a meta-learning stage of network parameters is descent towards a 'most adaptive' point, and the network can quickly adapt to a new task.

Description

technical field [0001] The invention belongs to the field of machine learning, and in particular relates to a meta-learning (Meta Learning)-based fast-adaptive discriminative restricted Boltzmann machine (Discriminative restricted Boltzmann machine, DRBM) method. Background technique [0002] Restricted Boltzmann machine (RBM) network is one of the most popular basic models in machine learning, and it is also the most commonly used basic component in deep neural networks. RBM can use its hidden units to extract features and learn the probability distribution of data, and use the learned probability distribution to generate new samples. It has been extensively studied by scholars in the fields of target recognition and probability models. DRBM is an extended form of RBM. Its core idea is to construct a discriminant function in a certain number of sample sets, and use the feature vector and label together as the input of RBM for training, so that RBM has the function of classi...

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): G06N3/08G06N3/04G06F17/18
CPCG06F17/18G06N3/04G06N3/08
Inventor 张新禹刘子衿任祖煜霍凯刘振张双辉刘永祥姜卫东黎湘卢哲俊
Owner NAT UNIV OF DEFENSE TECH
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Try Eureka
PatSnap group products

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

© 2024 PatSnap. All rights reserved.Legal|Privacy policy|Modern Slavery Act Transparency Statement|Sitemap