Semi-supervised learning ideas learning record

Semi-supervised learning ideas sorting

1. Semi-supervised learning ideas semi-supervised learning (SSL)

Clear knowledge points are required :

1. First, make sure that the training set contains two kinds of data: labeled and unlabeled.

2. Our ultimate goal is to get a classifier, that is, a network model.

3. The condition for the end of training can be to use unlabeled data as the input of the network to obtain the output prediction label, and the data within a certain confidence level can be divided into labeled data until the data in the training set has labels. The classifier can be considered as the final classifier.

steps :

1. An initial network model f is obtained by training the network with initially labeled data.
2. Use the initial network model f to predict the unlabeled data x to obtain the predicted label y, and the predicted label is considered correct within a certain degree of confidence.
3. The data (x, y) that is considered to be the correct predicted label is removed from the unlabeled data in the training set to the labeled data.
4. Use the newly obtained data set to retrain the network until the condition is met.

The finally obtained network model is the final network model.