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

A method and structure for training deep neural network on unbalanced data set

A deep neural network, unbalanced data technology, applied in the field of deep learning, can solve problems such as inability to train

Pending Publication Date: 2019-07-26
NORTHWEST UNIV(CN)
View PDF0 Cites 3 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

[0003] Aiming at the deficiencies and deficiencies in the prior art, the present invention provides a method and structure for training a deep neural network on an unbalanced data set, which overcomes the fact that the existing methods cannot handle extremely unbalanced data with few minority samples. Effective training on set and other deficiencies

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 method and structure for training deep neural network on unbalanced data set
  • A method and structure for training deep neural network on unbalanced data set

Examples

Experimental program
Comparison scheme
Effect test

Embodiment 1

[0053] like figure 1 and figure 2 As shown, the present invention provides a method for training a deep neural network on an unbalanced data set, the method first divides the training set and the verification set by stratified sampling or the leave-one-out method imposed on the minority class samples; then, The adaptive normalization layer is used as the first layer, the sample diversification layer is used as the second layer, and the target deep neural network is concatenated after the sample diversification layer to obtain the overall deep neural network; finally, batch equalization is used to optimize the overall deep neural network. The network is trained. The method comprises the steps of:

[0054] Step 1, divide the training set and validation set:

[0055] Divide the original unbalanced data set into a training set and a validation set according to the ratio and unbalanced ratio required for the number of samples in the training set and the validation set;

[0056...

Embodiment 2

[0086] like figure 1 As shown, the present invention also provides a structure suitable for training a deep neural network on an unbalanced data set, the structure comprising: an adaptive normalization layer as the first layer, a sample diversification layer as the second layer And the target deep neural network connected in series after the sample diversification layer; the target deep neural network is any neural network that needs to be trained on an extremely unbalanced data set.

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 a structure for training a deep neural network on an unbalanced data set. The method comprises the following steps: firstly, dividing a training set and a verification set through hierarchical sampling or a leave-one method applied to a few types of samples; then, taking the adaptive normalization layer as a first layer, taking the sample diversification layeras a second layer, and serially connecting the target deep neural network to the sample diversification layer to obtain an overall deep neural network; and finally, training the overall deep neural network on the training set by adopting batch equalization. According to the method, the problem that effective learning cannot be carried out on an extreme unbalanced data set with few types of samplesin the prior art is effectively solved; therefore, the method for training the deep neural network on the unbalanced data set does which not need to preprocess the unbalanced data, does not introducesuper parameters, can execute self-adaptive normalization, and is compatible with the existing deep learning technology is provided.

Description

technical field [0001] The invention belongs to the technical field of deep learning, and in particular relates to a method and structure for training a deep neural network on an unbalanced data set. Background technique [0002] How to efficiently learn on unbalanced datasets is a common and widely studied problem in the field of machine learning. The methods used to deal with unbalanced data in traditional machine learning can be roughly divided into two categories: methods based on data preprocessing and methods based on algorithm modification. The method based on data preprocessing is to improve the unbalanced degree of the data set through data preprocessing, such as upsampling method and downsampling method. The method based on algorithm modification is to try to modify the algorithm itself to make it better support training on unbalanced data sets, such as through cost-sensitive learning and the method of integrating models learned on multiple sub-training sets. In ...

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): G06K9/62G06N3/04
CPCG06N3/045G06F18/214
Inventor 冯筠胡景钊刘阳
Owner NORTHWEST UNIV(CN)