Traditional machine learning process, different tasks will use a different machine learning models, however, are learning to migrate data for a certain type of construction can also be applied to other types of systems and data.
Transfer learning from one or more sources in the art by training the model, useful knowledge obtained and used in the new target task (not labeled the same class have similar characteristics to the article or articles of different types unlabeled ) is essentially a transfer of knowledge reuse.
Migrating the goal of learning is to extract useful knowledge from one or more source tasks in the field and use it on the new goals and tasks, in essence, it is the transfer of knowledge reuse.
As defined migration learning, learning can migrate into three types, the distribution of differences in transfer learning, learning characteristic differences migration and migration differences label study. Edge of the source distribution difference migration learning domain and the target domain data distribution or the conditional probability distribution of the different source characteristic differences migration learning domain data and target different data feature space, the label difference transfer learning refers to data marks the spatial source and target domains different .
For example with banana and apple classification, data source domain is already tagged text data bananas and apples, the target domain is not new to the labeled bananas and apples of text data, data source and target domains from different times, different places, different data distribution, but the mark and space are the same feature space data using the source domain to the target domain learning problem is to transfer learning problems belonging to the distribution difference.
The source domain data is text data marked apples and bananas, and apples and bananas target domain is not marked with a picture data source and target domains is a text, one image, are characterized by differences in the migration range of learning .
The source domain data is text data marked bananas and apples, are binary classification, text data in the target domain is not labeled pears, oranges and oranges belong to three categories of problems, different spatial data mark source and target domains , marked differences belong to the scope of the migration study.
Under supervised learning mode is now mature, traditional form of supervised learning in intensive training marked a large sample volumes of data, but in the use of this model to the case is more complex, more real change in the environment will always be great error,
So is the case of transfer learning fewer samples than training a classifier, will put this model can be applied to a variety of other situations.
1. unmanned vehicles, unmanned aerial vehicles. Complex and changing environment unmanned vehicles and unmanned aerial vehicles are facing, not more determined, so the need to migrate to adapt to the role of learning algorithms in different environments.
2. robot.
3. The speech recognition. Migration of language, language can be a (Putonghua) to migrate from the application of various regional dialects and other languages (English, German, etc.).
