Multi-example learning method using deep learning technology

Abstract

The invention provides a multi-example learning method using a deep learning technology. A series of data filling and segmentation means are used to convert multiple-example samples into characteristic matrixes of the same size, and a convolutional neural network is used to carry out learning with monitoring and classification. According to the method of the invention, hidden abstract concepts in a multi-example data set can be discovered, errors of the data set can be tolerated effectively, and the generalization capability is high.

Description

Technical field [0001] The present invention relates to a machine learning method, in particular to a multi-instance learning method using deep learning technology. Background technique [0002] Multi-instance learning is a branch of machine learning, and has received extensive attention from machine learning researchers since its emergence. Multi-instance learning was originally proposed in the problem of drug molecule activity analysis, and then it has been widely used in image classification, speech recognition, text understanding and other fields. In the classification problem of multi-instance learning, the sample used to train the learner contains multiple examples, and only some of the samples will play a decisive role in the classification label of the example, while other samples in the example have no effect on the classification . For example, an image is composed of many small local areas, and only certain local areas can determine the classification of the image (s...

AI Technical Summary

Problems solved by technology

[0004] (1). Most of the existing methods are based on modifying the existing single-instance machine learning algorithm to make it suitable for multi-instance learning scenarios. However, this modification is subject to many restrictions, which will increase the complexity of the algorithm. Efficiency decline, information loss of training data sets and other issues, so that the effect of the algorithm is not very ideal;

[0005] (2). Existing methods are mostly based on supervised learning, so they are very dependent on the characteristics of the data in the data set and the quality of the labels. The robustness of the algorithm is relatively poor, and the small errors in the data set will be magnified, which has a greater impact on the accuracy of the model. big impact;

[0006] (3). Existing methods are mostly based on statistics and model the probability distribution of input features, which is not conducive to the discovery and modeling of hidden abstract concepts in multi-instance data sets, so that the accuracy and generality of the final classification model performance is severely limited

Images