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

Neural networks using intra-loop data augmentation during network training

A neural network, data technology, applied in the field of improving the performance of neural networks

Active Publication Date: 2020-04-14
INT BUSINESS MASCH CORP
View PDF8 Cites 1 Cited by
  • Summary
  • Abstract
  • Description
  • Claims
  • Application Information

AI Technical Summary

Problems solved by technology

Determine class accuracy difference and total loss for each class (based on training)

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
  • Neural networks using intra-loop data augmentation during network training
  • Neural networks using intra-loop data augmentation during network training
  • Neural networks using intra-loop data augmentation during network training

Examples

Experimental program
Comparison scheme
Effect test

example 1

[0060] Example 1. A machine intelligence model to distinguish between images showing normal (no pneumonia) and abnormal (pneumonia) features.

[0061] In this example, a machine intelligence model is trained to distinguish between normal and abnormal features to identify patients with pneumonia. In cases involving imaging and diagnosis of diseases, more images of normal patients can be obtained from the data set rather than images of diseased patients. Therefore, the techniques presented in this article are used to enhance disease data sets in order to improve the accuracy and precision of neural networks.

[0062] In this example, 1000 normal images and 100 diseased images showing pneumonia are obtained. The enhancement amplitude value for each class (L c ) Is set so that the initial value of the normal image is 1 and the initial value of the pneumonia image is 10. Generally, the initial value is set so that the number of samples for each class (in this case, the normal class an...

example 2

[0065] Example 2. Determine the harmonic mean of the accuracy for each class.

[0066] In this example, 1000 normal images and 100 pneumonias are provided. Similar to Example 1, the initial value of the enhancement magnification is set to 1 for the normal case and to 10 for the pneumonia case. It is determined that the total loss is below the loss threshold, and for pneumonia, the accuracy for each class is reduced by 0.8. Correspondingly, the enhancement magnification for pneumonia has been updated to 11 ( <-10+1), and perform updates and training until the enhancement magnification for pneumonia reaches 20.

[0067] Additionally, the harmonic mean of the accuracy for each class is determined for the enhancement magnification of 10 to 20, and the remaining rounds of training are performed at the magnification with the largest harmonic mean. The harmonic mean can be expressed as:

[0068] Harmonic mean

[0069]

[0070] (acc c : Class c precision)

[0071] When the variable (for exa...

example 3

[0072] Example 3. Enhanced test

[0073] A data set with 800 original training samples for pneumonia and 7640 training samples for cell invasion was obtained. A test data set with 200 test samples for pneumonia and 200 test samples for cell invasion was also obtained.

[0074] During experiment 1 of data enhancement, the enhancement magnification was set to a value of 10 for pneumonia (8000), and the enhancement magnification was set to a value of 1 for cell invasion (7640). After 20 rounds, the loss is determined: 0.64-> Accuracy: 0.1 for pneumonia, 0.95 for invasion (with harmonic mean: 0.18).

[0075] During trial 2 with increased data, for pneumonia (16000), the enhancement magnification was set to a value of 20, and for invasion (7640), the enhancement magnification was set to a value of 1. After 20 rounds, the loss is determined: 0.58-> Precision: 0.82 for pneumonia and 0.35 for invasion (harmonized average: 0.49).

[0076] After 100 rounds, the accuracy of pneumonia was 0.76 ...

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

Embodiments of the disclosure relate to neural networks using intra-loop data augmentation during network training. Methods and systems are provided for training a neural network with augmented data.A dataset comprising a plurality of classes is obtained for training a neural network. Prior to initiation of training, the dataset may be augmented by performing affine transformations of the data inthe dataset, wherein the amount of augmentation is determined by a data augmentation variable. The neural network is trained with the augmented dataset. A training loss and a difference of class accuracy for each class is determined. The data augmentation variable is updated based on the total loss and class accuracy for each class. The dataset is augmented by performing affine transformations ofthe data in the dataset according to the updated data augmentation variable, and the neural network is trained with the augmented dataset.

Description

Technical field [0001] The embodiments of the present invention relate to machine learning systems, and in particular to perform dynamic data enhancement during training to improve the performance of neural networks. Background technique [0002] Deep neural networks usually require large amounts of training data to meet specified performance standards. However, collecting such training data is often time-consuming and can be expensive. Insufficient training data or using a biased training data set may cause the trained neural network to also have bias, and the classification of the training data is prone to overfitting. Therefore, when applied to other data sets, the trained neural network may not meet the desired performance standards. Summary of the invention [0003] According to an embodiment of the present invention, a method, a system, and a computer-readable medium for training a neural network using enhanced data are provided. Obtain a data set including multiple class...

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): G06N3/04G06N3/08G06N20/00G06V10/772
CPCG06N3/08G06N20/00G06N3/044G06N3/045G06V10/82G06V10/772G06F18/214
Inventor 后藤拓矢阪本正治中野宏毅
Owner INT BUSINESS MASCH CORP