First, let's first understand what co-training is:
In computer vision, we all know that when training a classifier, we need two types of samples: positive and negative. Supervised training can become off-line training, that is, preparing samples already labeled in advance, training the classifier, and classifying the trained classifier. The other is online training. At the beginning, we initialized the classifier with some labeled samples. In this way, the classifier has a certain effect, and then classifies unlabeled samples, and then uses relevant methods to identify positive and negative samples, then, the classifier is continuously improved in the process to continuously improve the classifier classification effect.
Co-training was first proposed by Blum he Mitchell in 1998 as Combining labeled and unlabeled data with co-training.
Early Application: 1: Web-page classification (Page text \ Hyperlink text ))
2: bilmetric recognition systems (appearance and voice)
In object detection, co-training is mainly used for vehicle detection and mobile object recognition in monitoring.
Co-training is a popular semi-guidance machine learning method. Its basic idea is to construct two different classifiers and use small-scale labeled corpus, the method for marking large-scale unlabeled corpus. the biggest advantage of the Co-training method is that it can automatically learn knowledge from unlabeled corpus without manual interference. The Co-training method is a compromise between guidance and no-guidance machine learning. Its principle is to use as many unlabeled data as possible without sacrificing performance, it starts from a small-scale labeled Corpus and uses large-scale unlabeled materials for learning. The Co-training algorithm is applied to datasets with natural severability of attributes, that is, certain attributes of a dataset can depict certain characteristics of the data from a certain angle, and these attributes are not unique, there are many different attributes that can depict the same feature from different perspectives, so that the data property set is naturally severable, ignoring the severability of the dataset feature. the Co-training algorithm uses two different learners to independently learn in the dataset/Segmentation feature set, and makes the final learning conclusion based on the learning results of the two learners, in this way, the error rate can be reduced.
The description of Co-training is as follows:
- Two views are a classification problem;
- Create a model based on different angles and train each model under the annotation set;
- Mark unlabeled sentences, and then find sentences with high confidence in each model;
- Pick out these highly confident sentences in different ways;
- Add these sentences to the training set and iterate this process until no labeled data is exhausted;
What is co-training?