Semi-supervising method without consistency constraint and mobile terminal

Abstract

The invention provides a consistency constraint-free semi-supervising method and a mobile terminal, the method comprises the following steps: presetting a data set and a classification model, the dataset comprises a labeled data set and a label-free data set, and the classification model comprises a feature extractor; based on a gradient iterative optimization algorithm, performing minimization problem iteration on the data set in the classification model to obtain minimization distribution information; performing maximization problem iteration on the data set in the classification model to obtain maximization distribution information; and obtaining a calculation result according to the minimized distribution information, the maximized distribution information and a semi-supervised algorithm formula without consistency constraint. Training of the feature extractor is restrained through the distance between the empirical distribution of the labeled data in the feature space and the empirical distribution of the unlabeled data in the feature space, so that the feature extractor enables the labeled data to be close to the feature distribution of the unlabeled data in the feature space, and the classification effect of the model on real distribution data is improved. And the generalization ability of the model is improved.

Description

technical field [0001] The present application relates to the technical field of deep learning, in particular to a semi-supervised method and a mobile terminal that do not require consistency constraints. Background technique [0002] Semi-supervised learning algorithm refers to a class of algorithms that can use both labeled data and unlabeled data for learning. It is usually used in scenarios where there is a large amount of unlabeled data but only a small amount of labeled data. By finding ways to utilize unlabeled data, semi-supervised learning can achieve better and sometimes substantially better results than supervised learning algorithms. [0003] As far as we know, semi-supervised learning algorithms can be traced back to the 1960s. The idea at that time was self-training, that is, first train an initial model with labeled data, then use this initial model to label part of the unlabeled data, and then use these labeled data and labeled data to train the second model...

AI Technical Summary

Problems solved by technology

[0007] This application provides a semi-supervised method without consistency constraints and a mobile terminal to solve the problem that the hypothesis-based model in the prior art has poor effect on the real distribution of data and the generalization ability of the model is low

Images