[paper reading] internally semi supervised methods

The first two categories are equivalent to making an adapter , The method originally belonging to supervised learning and unsupervised learning is used in semi supervised learning . This section mainly introduces the semi supervised learning method in essence . It adds the unlabeled data loss to the objective function .

1 Maximum-margin methods

The classic in this category is to SVM From supervised to semi supervised , Consider unlabeled samples to construct the optimal hyperplane .

2 Perturbation-based methods

be based on Disturbance Methods , The reason for this method is very simple , Unlabeled samples have no labels , So whether it's regression or classification , There is no right or wrong , So how to produce a kind of Loss Added the concept of Objective function What about China? ？

Answer: Train two networks , Here we need some tricks To make the two networks different , In short, it can not be the same , Otherwise, the predictions of the two models are the same , Loss is meaningless . Disturbance That's one of them trick, The more common is in each layer of the model noise .

2.1 Ladder networks

Portal

2.2 Pseudo-ensembles

Compared with the previous model, the data is Disturbance , This model is right Model Make a disturbance .

2.3 $\mathrm{\Pi }$-model

3 Manifolds

4 Generativemodels