This article is for understanding the basics of transfer learning! ! !
1. What is transfer learning?
Transfer learning is also known as inductive transfer, domain adaptation, and its goal is to apply knowledge or patterns learned in a certain domain or task to a different but related domain or problem. For example, the skills of learning to walk can be used to learn to run, the experience of learning to recognize cars can be used to recognize trucks, etc.
2. The main idea of transfer learning?
Transfer labeled data or knowledge structures from related auxiliary domains, and complete or improve learning in the target domain or task.
3. What is the significance of transfer learning research?
In many engineering practices, it is very expensive or even impossible to collect labeled data for each application domain, so it is necessary to transfer existing knowledge structures from auxiliary domains or tasks to complete or improve the target domain tasks. Important research questions arising from practical needs.
4. What are the characteristics of transfer learning compared to traditional machine learning?
Transfer learning relaxes the assumption that the traditional machine learning training data and test data obey the independent and identical distribution, so that the fields or tasks involved in the learning can obey different marginal probability distributions or conditional probability distributions.
5. Comparison of transfer learning and semi-supervised learning?
Although traditional semi-supervised learning can solve the data sparsity, it requires a considerable degree of labeled data in the target domain; when the labeled data is very scarce and the acquisition cost is too high, it is still necessary to transfer knowledge from the auxiliary domain to improve the learning effect of the target domain.
6. A description of the transfer learning problem?
Transfer learning designs two important concepts of domain and task.
Domain D is defined as consisting of a d-dimensional feature space X and a marginal probability distribution P(x);
The task T is defined as consisting of a class space Y and a prediction model f(x) (conditional probability distribution)
7. See the differences in probability distributions between domains?
The data in the two domains were reduced to two dimensions for visualization using the PCA method.
8. What is the relationship between input space and feature space?
The space in which all eigenvectors exist is called the eigenspace. Each dimension of the feature space corresponds to a feature, and sometimes the input space and the feature space are assumed to be the same space; sometimes the input space and the feature space are assumed to be different spaces, and instances are mapped from the input space to the feature space.
9. Classification of transfer learning?
Classification according to feature space, category space, marginal probability distribution, conditional probability distribution
Divided into two categories: heterogeneous transfer learning (source and target domain feature spaces are different or different category spaces), homogeneous transfer learning (source and target domain feature spaces are the same and category spaces are the same)
10. Feature representation for unsupervised transfer learning?
Unsupervised transfer learning is a transfer learning task without labeled data in the target domain.
By learning a new feature representation Φ(x), the shared properties between domains are enhanced and the exclusive properties are weakened.
It is based on the assumption that some features in the feature space are domain-exclusive, while other features are domain-shared and generalizable; or there is a shared and generalizable implicit feature space between domains, which It can be extracted by the feature learning algorithm under the criterion of reducing the probability distribution difference between domains.
Feature representations can be divided into two subcategories: implicit representation learning methods and probability distribution fitting methods.
Implicit feature representation: Construct abstract feature representations by analyzing a large number of unlabeled examples in auxiliary and target domains, thereby implicitly narrowing the distribution differences between domains;
Probability distribution fitting method: Explicitly improve the sample distribution similarity between the auxiliary domain and the target domain by penalizing or removing features that are statistically variable between domains, or by learning a subspace embedding representation to minimize a specific distance function.
11. The main problem challenge of transfer learning problem?
Including over-fitting and under-fitting problems of classical machine learning, as well as under-fitting and negative transfer problems unique to transfer learning;
Negative transfer: Auxiliary domain tasks have a negative effect on the target domain task. The main idea of the current research on negative transfer from the perspective of algorithm design is to reduce the knowledge structure transferred between domains, such as only sharing the prior probability of the model between domains, while No model parameters or likelihood functions are shared.
Underfitting: The problem of fitting probability distributions across domains has not been adequately corrected.
Underfitting: The learned model fails to adequately characterize important structure of the probability distribution.
Overfitting: The learned model overfits irrelevant information about the sample distribution.
To sum up: overfitting and underfitting are aimed at the performance of the learning model in a certain field, and underfitting and negative transfer are aimed at the influence of the auxiliary domain knowledge structure or pattern on the performance of the learning model in the target domain.
12. What are the existing probability distribution similarity measurement functions?
Maximum Mean Difference, Bregman Divergence, etc.
13. What is the relationship between transfer learning and machine learning?
Transfer learning emphasizes the transfer of knowledge between different but similar domains, tasks and distributions. Essentially, transfer learning is a new machine learning method that applies information and knowledge from existing fields to different but related fields. Transfer learning does not require similar fields to obey the same probability distribution. Its goal is to transfer the existing knowledge and information in the source field to the new field through certain technical means, so as to solve the problem that the target field has less labeled sample data. Not even labeled learning problems.
14. Classify transfer learning based on whether there are labeled samples in the source and target domains?
归纳迁移学习：目标领域带有标签数据可以用于归纳目标领域中的预测模型，目标任务与源任务是不同但相关的。
直推式迁移学习：目标领域中数据没有标签，源领域有少量数据标签是可用的。源任务和目标任务是相同的。
无监督迁移学习：目标领域和源领域数据都没有标签，通过挖掘领域间内在结构特征进行迁移学习，源任务和目标任务是相关但不同的。
参考资料：
1. 龙明盛 博士论文《迁移学习问题与方法研究》
2. 张景祥 博士论文《迁移学习技术及其应用研究》