Semi-supervised Learnin g
Learning method that uses a small amount of labels and a large amount of unlabeled data together.
Semi-reviewed learning (semi-supervised learning) is a machine learning approach that provides a medium way between using a large amount of unlabeled data together with a small amount of labeled data. Collecting data in the real world is often cheap and easy - millions of images or text can be downloaded from the internet- but labeling this data correctly (such as review of each image of a specialist or classification of each text) is often a process that requires cost-effective, slow and sometimes expertise. Semi-reviewed learning is designed to overcome this practical restriction: the starting knowledge learned from a small number of labeled examples is enriched using the structure in the unlabeled large data set.
One of the common techniques is "self-training" (self-training): the model is first trained with a small set of label data, then this model is used to predict unlabeled data, and the most reliable predictions are re-trained by adding the model to the training set as "sözde label" (pseudo-label). Another approach is promoted to provide consistent forecasts for example, slightly modified versions of the model (consistency regularization). Specialist labeling such as medical imaging is expensive, but semi-supervised learning is particularly valuable approach in areas where raw data is abundant.
One of the common techniques is "self-training" (self-training): the model is first trained with a small set of label data, then this model is used to predict unlabeled data, and the most reliable predictions are re-trained by adding the model to the training set as "sözde label" (pseudo-label). Another approach is promoted to provide consistent forecasts for example, slightly modified versions of the model (consistency regularization). Specialist labeling such as medical imaging is expensive, but semi-supervised learning is particularly valuable approach in areas where raw data is abundant.