Dynamic label deep learning algorithm

Abstract

The invention discloses a dynamic label deep learning algorithm. The algorithm comprises the steps of S1, inputting a data set D; S2, randomly initializing a classifier f (x, 0); S3, randomly dividingthe data set D into a training set Dtrain and a verification set Dvalid; S4, for each sample of the training set Dtrain, randomly selecting a new label according to the label migration probability matrix T to form a new training set Dnew; S5, training a classifier f(x; 0) on the new training set Dnew; S6, testing the precision of the classifier f(x; 0) on the verification set Dvalid; and S7, repeating the steps S4 to S6 until the classifier precision reaches a preset value. According to the method, model over-fitting can be effectively avoided, the model with better generalization performanceis learned, the method has good anti-noise capability and anti-over-fitting capability, and the optimal classifier learning problem of the noise training set is effectively solved.

Description

technical field [0001] The invention belongs to the technical field of deep learning, and in particular relates to a dynamic label deep learning algorithm. Background technique [0002] In recent years, machine learning and deep learning classification models based on large-scale noise-free training sets have achieved tremendous progress. However, obtaining a large-scale noise-free training set is very difficult in real life. Manually annotating large-scale noise-free training sets is very expensive and time-consuming. When the labels of the training set are noisy, it is called label noise, and the training set is called a noisy training set. In machine learning and deep learning, the label noise of the training set seriously affects the performance of the model. Solving the classification problem of noisy training sets has very important practical significance, and is one of the hot spots in machine learning and deep learning research, and has aroused great interest in t...

AI Technical Summary

Problems solved by technology

[0007] The purpose of the present invention is to address the above-mentioned deficiencies in the prior art, to provide a dynamic label deep learning algorithm to solve the over-fitting problem of the existing classification model in the training process and to learn the optimal noise-free subset in the noise training set. The Problem with Optimal Classifiers

Images