Reference: Concise Handbook of Transfer Learning
table of Contents
1 Transfer learning concept
To learn transfer learning, we must first clarify three questions, what is transfer learning, why and how?
1.1 What is it-introduction to transfer learning
Migration learning, as the name suggests, is to transfer. Putting it into our subject of artificial intelligence and machine learning, transfer learning is a learning idea and mode.
As an important branch of machine learning, transfer learning focuses onApply the learned knowledge to new problems。
Transfer learningthe core issueYes,Find the similarities between the new problem and the original problem, Can successfully realize the transfer of knowledge.
example:
For example, China, the United States, and Japan are in the northern hemisphere. Based on the similarity in the northern hemisphere, we can judge that the climate in the United States and Japan is similar to that of China.
The climate in Australia and South Africa may be the opposite of the above, because Australia and South Africa are in the southern hemisphere.
For another example, if you can play table tennis, then you are more likely to play tennis. You can play chess, then you may also play chess.
1.2 Why-why use transfer learning, why can it be used
Why use transfer learning boils down to three points:
1.2.1 The contradiction between big data and less labeling.
As we all know, the training and updating of machine learning models all rely on data labeling. However, labeling is often difficult to implement and costly. On the other hand, in certain fields, because there is not enough calibration data for learning, these fields have not been able to develop well.
1.2.2 The contradiction between big data and weak computing.
Big data requires large equipment and equipment with strong computing capabilities for storage and calculation. However, the large computing power of big data is a game that only "the rich" can afford. Such as Google, Facebook, Microsoft, these giant companies have strong computing power to use these data to train models. And how can ordinary people use these data and models?
1.2.3 The contradiction between universal models and individual needs
The goal of machine learning is to build a model that is as universal as possible, so that this model can meet different users, different devices, different environments, and different needs. But specific to each individual and each need, there is its uniqueness and specificity, and a universal model cannot meet it at all. So, can this universal model be transformed and adapted to better serve people's individual needs?
1.2.4 Requirements for specific applications.
Machine learning has been widely used in real life. In these applications, there are also some specific applications, and they are facing some practical problems. For example, the cold start problem of the recommended system (Lack of data)
1.3 What to do-how to transfer (transfer learning method)
In response to the above four problems, transfer learning has proposed corresponding strategies.
1.3.1 Big Data and Less Labeling: Migrating Data Labeling
It is impossible to accurately train highly available models based on a small amount of labeled data alone. In order to solve this problem, our intuitive thought is: just add more annotation data? But without relying on manual labor, how to add label data? Using the idea of transfer learning, we canFind some labeled data that is similar to the target data, So as to use these data to build models and increase the annotation of our target data.
1.3.2 Big Data and Weak Computing: Model Migration
It is impossible for everyone to be able to use big data to quickly train models. Using the idea of transfer learning, we can combine thoseModels trained by big companies on big data,Migrate to our mission. For our missionFine-tuning, So we can also have a model trained on big data. Furthermore, we can adaptively update these models for our tasks to achieve better results.
1.3.3 Universal model and individual needs: adaptive learning
In order to solve the challenge of individual needs, we use the idea of transfer learning,Perform adaptive learning (form a model according to the user's situation). Taking into account the similarities and differences between different users, we make flexible adjustments to the universal model in order to complete our tasks.
1.3.4 Requirements for specific applications: knowledge transfer in similar fields
In order to meet the needs of specific applications, we can use the methods introduced above to perform transfer learning from data and model methods.
Table 1 summarizes the necessity of transfer learning.
2 Classification of transfer learning
2.1 According to the target domain label
This classification method is the most intuitive. Analogous to machine learning, transfer learning can be divided into the following three categories according to whether there is a label in the target field:
- Supervised Transfer Learning
- Semi-Supervised Transfer Learning
- Unsupervised Transfer Learning
2.2 Classification by learning method
-
Sample-based transfer learning method (Instance based Transfer Learning)
That is to say directly toDifferent samples give different weightsFor example, for similar samples, I give it high weight.
-
Feature based transfer learning
It means that if the features of the source domain and the target domain are not originally in the same space, or they are not similar in the original space, then we will find a way to combine themTransform a space, Aren’t these characteristics similar?
-
Model based transfer learning
The structure of the neural network can be directly transferred
-
Relation based Transfer Learning
This is mainly about mining and using relationships for analogy transfer. For example, the teacher is in class and the students are in class.analogyFor the company's meeting scene. This is a kind of relationship migration.
At present, the hottest is the migration based on features and models, and then the migration method based on examples is used in combination with them.
2.3 Classification by characteristics
According to characteristic attributes, transfer learning can be divided into two categories:
- Homogeneous Transfer Learning
- Heterogeneous Transfer Learning
For example, the migration of different pictures can be considered as isomorphic; while the migration from pictures to text is heterogeneous.
2.4 According to offline and online forms
According to offline learning and online learning, transfer learning can also be divided into:
- Offline Transfer Learning
- Online Transfer Learning
At present, most of the transfer learning methods have adoptedOffline mode. That is, both the source domain and the target domain are given,Migrate onceThat's it. The disadvantage of this approach is obvious: the algorithm cannot learn from the newly added data, and the model cannot be updated. The opposite is the online method. That is, with the dynamic addition of data, the migration learning algorithm can alsoConstantly update。