当前位置：网站首页>Summary and future prospect of transfer learning | community essay solicitation

Summary and future prospect of transfer learning | community essay solicitation

2022-06-22 09:24:00 【InfoQ】

The article brief introduction

Study this article , You can gain the following knowledge :

What is transfer learning

Why use transfer learning

The advantages of transfer learning

Classification of transfer learning methods

The future of transfer learning

What is transfer learning ？

Generally speaking, it is most appropriate to use an idiom to describe ——

The lines

. Transfer learning utilizes and synthesizes knowledge extracted from similar tasks , And the valuable experience accumulated from the past , To promote learning new problems . The core is to find the common ground between existing knowledge and new knowledge / similarity .

Take a chestnut , After Xiaobao learned to ride a bike , Electric cars 、 Motorcycles can be used quickly , But driving a car requires re systematic learning . Here's a chestnut , such as , You learn programming , First I learned

C

Language , With

C

The basis of language , You can soon learn by analogy

Python

、

Java

Wait for computer language , but

C

Language will not help you learn Japanese .

Transfer learning , Researchers usually divide data into source data and target data . Source data refers to other data that is not directly related to the task to be solved , Usually with large data sets . Target data is data directly related to the task , The amount of data is generally small . The bicycle in the above can be regarded as the source data , Electric cars 、 The motorcycle 、 All cars are target data , But the similarity between cars and bicycles is low .

What migration learning should do is to make full use of the source data to help the model improve its performance on the target data .

Take a chestnut , Xiaobao is learning

NILM

Electric meter

V-I

Track recognition direction , Relevant public data sets can reach tens of thousands of data at most , And not targeted

NILM

A pre training model is proposed , But there are many image recognition models , for example

AlexNet

、

VGG-16

、

googlet

and

ResNet-50

etc. . These models are carefully trained based on millions of images , If we migrate these models to NILM in , Will greatly improve NILM Accuracy of trajectory recognition .

In transfer learning, if the knowledge correlation between the source domain and the target domain is low , The effect of transfer learning will also be poor , This is known as “ Negative transfer ”. For example, text data model is migrated to image data model , The migration performance will be poor . But for text migration to images , Not without a solution , We can connect two seemingly unrelated domains through one or more intermediate domains , This is known as **“ Transitive transfer learning ”**, Transitive transfer learning is also one of the hotspots that researchers pay attention to .

For example, in order to realize the migration between text and image , The literature 《Transitive Transfer Learning in Proceedings of the 21th ACM SIGKDD International Conference onKnowledge Discovery and Data Mining - KDD ’15》 Use annotated images as intermediate fields .

Why transfer learning ？

stay AI( Artificial intelligence ) and ML( machine learning ) field , The motivation for the use of transfer learning in AI is stronger than ever , Because it can solve the two limitations of large amount of training data and training cost .

About why transfer learning , Professor Wang Jindong and other professors summarized it into four aspects :

The contradiction between big data and less labeling ： In the era of big data, massive data is generated all the time , But the data lacks perfect data annotation , The training and updating of machine learning model depend on the annotation of data , At present, only a few data are marked .

The contradiction between big data and weak Computing : Massive data requires huge storage and computing power , Strong computing power is very expensive , In addition, the training of massive data takes a lot of time , Therefore, it leads to the contradiction between big data and weak Computing .

Contradiction between universal model and personalized demand : The purpose of machine learning is to build a model as general as possible to meet different users 、 Different equipment 、 Different needs of different environments , This requires the model to have high generalization ability , But in reality, the universal model can not meet the needs of personalization 、 Differentiated needs , This leads to the contradiction between the model and personalized requirements .

Application specific requirements : In reality, there are often some specific applications , For example, the cold start problem of the recommended system , This requires us to use the existing model or knowledge to solve the problem as much as possible .

in summary , When we use artificial intelligence to solve problems , The biggest obstacle is that model training requires a lot of data and parameters , On the one hand, we usually can't get the data of the scale needed to build the model ( Marked ); On the other hand , The training of the model needs a lot of time . And transfer learning , Valuable knowledge and previous experience from similar tasks can be used to significantly improve traditional AI Learning performance of technology .

Advantages of transfer learning

In conclusion, compared with the previous machine learning and deep learning , It has the following advantages :

Improve the quality and quantity of training data : Transfer learning selects and transfers knowledge from similar fields with a large amount of high-quality data

Speed up the learning process : Benefit from valuable knowledge and experience shared from other similar areas / Or what you've learned in the past , It can significantly improve the learning rate

Reduce computation : Most of the data in transfer learning is before the trained model is transferred to the target domain , Are trained in other source domains , Thus, the computational requirements of the training process of the target domain are greatly reduced .

Reduce communication overhead : You don't need to send a lot of raw data , Just transfer knowledge

Protect data privacy : Users do not need to learn from raw data in other fields , Just start with your trained model ( Expressed by weight ) Just study in , Therefore, data privacy can be protected .

Classification of transfer learning methods

Sample based migration

Sample based migration is to select samples with high similarity with the target domain data from the source domain data set according to a similarity matching principle , And migrate these samples to the target domain to help the learning of the target domain model , So as to solve the learning problem of insufficient labeled samples or unlabeled samples in the target domain .

Generally, the weight of samples is trained by the similarity between the source domain and the target domain , The source domain data samples with large similarity think that they have strong correlation with the target domain data, which is beneficial to the target domain data learning, and the weight is improved , Otherwise, the weight will be reduced .

The traditional method is the sample weighting method , Use discrimination to distinguish source data and target data 、 Kernel average matching method 、 Function estimation method to estimate the weight , but

The density ratio between the source domain and the target domain should be calculated for the weight (

MMD

And

KL

Equidistance measurement ), The calculation is difficult

Migration based on model parameters

Model based transfer learning is to share some common knowledge between source tasks and target tasks at the model level , Including model parameters 、 Model prior knowledge and model architecture , It can be divided into two types: knowledge transfer based on shared model components and knowledge transfer based on regularization . The former uses the model components or super parameters of the source domain to determine the target domain model ; The latter prevents over fitting of the model by limiting the flexibility of the model .

Generally speaking, it means , First, use a large amount of data in the source domain to pre train the model , Then migrate the obtained weight parameters , Finally, the whole connection layer is retrained with a small amount of target data .

Feature based migration

The core of feature transfer method is to find the typical features between the source domain and the target domain to further weaken the differences between the two domains, so as to realize the cross domain transfer and integration of knowledge .

Feature transfer methods can be further divided into feature extraction transfer learning methods and feature mapping transfer learning methods according to whether they are selected from the original features . The advantage is that the similarity between models can be used , The disadvantage is that the model parameters are not easy to converge .

Feature extraction and migration method

Definition : Reuse the pre trained local network in the source domain , Turn it into part of the target domain depth network .

Will usually

CNN

Model as feature extractor , Then use a small amount of data to fine tune the network , Trained based on different fine-tuning strategies

CNN

The performance of the model is also different , Therefore, fine tuning strategy is the focus of this kind of methods . At present, the fine-tuning strategy is mainly multi-scale feature migration , Learn the characteristics of target data by fine tuning different network layers .

Feature mapping migration method

Definition : Map the instance from the source and target domains to the new data space , In the new data space , Instances from both domains have similar data distributions , It is suitable for joint depth neural network

Feature mapping migration method performs feature mapping by adjusting the marginal distribution or conditional distribution of source domain data , To expand the size of the training set , Enhance the effect of transfer learning .

Compared with the feature extraction migration method , Feature mapping transfer learning is more cumbersome . First , To find a common feature representation between the source domain and the target domain , Then map the data from the original feature space to the new feature space .

Feature based transfer learning has a wide range of applications , It can be used regardless of whether the source domain and target domain data have labels . However , When data has labels , The measure of domain invariance is not easy to calculate ; When the data has no label , Learning common features across domains is also difficult .

The future of transfer learning

The combination of transfer learning and generation of confrontation network

Generative adversary network （GAN, Generative Adversarial Networks ） It's a deep learning model , It is one of the most promising methods of unsupervised learning on complex distributions in recent years . The model passes through the framework （ At least ） Two modules ： Generate models （Generative Model） And discriminant models （Discriminative Model） Learning from each other's game produces quite good output .

Generative antagonistic network GAN It is a widely used data enhancement method in recent years , This method can generate false samples similar to real samples , Thus, the training samples can be expanded , Achieve the effect of improving the performance of the model .

GAN

Combined with transfer learning to form

DANN

Domain versus network ( The picture below is

DANN

The structure of domain confrontation network ),

DANN

Domain countermeasure network directly optimizes the loss on the source domain ; Using confrontation method to optimize the communication between source domain and target domain

HΔH

distance ; The loss upper bound of the target domain is minimized . But for the

DANN

For this kind of network , The training will be more difficult , And it is difficult to expand from single source domain to multi-source domain , This is also the problem we need to solve later .

Transfer learning combined with attention mechanism

Attention mechanism （Attention Mechanism） From the study of human vision . In cognitive science , Due to the bottleneck of information processing , Humans selectively focus on part of all information , While ignoring other visible information . There are two main aspects of attention mechanism ： Decide which part of the input you need to focus on ; Allocate limited information processing resources to important parts .———— Baidu Encyclopedia

The literature 《Transferable Attention for Domain Adaptation》 Put forward

TADA

( The picture below is

TADA

chart ) Methods the attention mechanism was used to select the transferable images and the key areas in the images , Improve model performance . Two transfer processes combined with attention mechanism are proposed ：

Transferable Local Attention

and

Transferable Global Attention

Attention mechanism can improve the accuracy of the model to a certain extent , But it also takes up too much computing resources . Therefore, various lightweight attention mechanisms have been proposed in recent years , But the lightweight attention mechanism will lose some model accuracy . Therefore, how to ensure the accuracy under the premise of , Combine the lightweight attention mechanism with the transfer learning method more effectively , It is a problem to be studied .

Transfer learning combined with federal learning

Federated learning is a kind of machine learning , Many clients train the model together under the coordination of the central server , At the same time, keep the training data decentralized and decentralized . The long-term goal of federal learning ： Analyze and learn the data of multiple data owners without exposing the data .（ Purpose ： Solve data silos ）

2020

Global artificial intelligence and robotics Summit （ abbreviation “

CCF-GAIR 2020

”）, Professor Yang Qiang introduced the key technologies and application cases of federal learning , It further introduces the latest research on the combination of federal learning and transfer learning, as well as the next key research directions .

Professor Yang Qiang said , We built

AI

looked after all the time , Protecting people's privacy is the moment

AI

A particularly important point in development , This is also from the government to the individual 、 Requirements of enterprises and society ; in addition ,

AI

Also protect the security of the model , Prevent malicious or non malicious attacks . Data privacy has become

AI

A difficulty that development has to overcome .

And federal transfer learning (

FTL

) By applying homomorphic encryption （

Homomorphic Encryption

） And polynomial approximation instead of differential privacy （

Polynomial Approximation instead of Differential Privacy

） Methods , Provides a safer way for specific industries 、 A more reliable way . meanwhile , Based on the characteristics of transfer learning ,

FTL

Participants can have their own feature space , There is no need to force all participants to own or use data with the same characteristics , This makes

FTL

Suitable for more application scenarios .

Research on the measurement of transfer learning domain

Because the performance of transfer learning depends largely on the similarity between domains . Therefore, the research on measurement methods is one of the important fields of transfer learning in the future , Accuracy of measurement , Computational convenience will affect the development of transfer learning .

Multi source domain migration knowledge

It is limited to migrate knowledge from only one source domain , If we can realize the comprehensive learning of knowledge in multiple fields , That is to combine multiple learning and transfer learning , This can increase the opportunity to find and learn more beneficial knowledge for the target domain , So as to improve the learning efficiency and effect of transfer learning , Make the transfer learning more secure and stable , Effectively avoid negative migration .

summary

Transfer learning is a new research field , By using the original knowledge to help the training in the target field , It has excellent performance . Transfer learning can also be combined with many methods , For example, federal learning 、 Attention mechanism, etc , Have excellent performance . Similarly, there are many problems and challenges in transfer learning : Negative migration problem , Failed to improve the model capability , Instead, it reduces the accuracy ; Inter domain metrics are too complex , There is no good measurement method, etc .

But as Wu Enda, a famous professor at Stanford University, once said in

2016

year

NIPS

The meeting said ：" Transfer learning will be the next driving force of machine learning business success after supervised learning ." With the deepening of research , Transfer learning will become another twinkling star in the field of artificial intelligence .