Deep learning mainly emphasizes characteristics, and reinforcement learning mainly emphasizes feedback, while migration learning mainly emphasizes adaptation.
The machine learning algorithms that introduced artificial intelligence have the top 5 categories. For details, please refer to related articles. Today we focus on the sixth category - Transfer Learning. ^_^
The traditional machine learning is to grow melons and grow beans, and the migration learning can be inferred.
Artificial intelligence competition, from the research and development competition of algorithm models to the competition of data and data quality, these successful models and algorithms are mainly driven by supervised learning, while supervised learning is extremely hungry for data and requires massive data (big data) support. To meet the precise requirements of the application. However, the development of artificial intelligence tends not to require massive data to meet the precise requirements of applications. Therefore, "small data learning" is becoming a new hotspot. The small data learning technology represented by migration learning and reinforcement learning can better reflect artificial intelligence. The road to the future. Some experts said that after supervised learning, migration learning will lead the wave of commercialization of the next wave of machine learning technology.
The concept of migration learning (TL) has received wide attention from academia since it was presented at the NIPS5 symposium in 1995. The definition of migration learning is too broad, and various specific terms have appeared in related research, such as learning to learn, life-long learn, mulTI-task learning, meta-learning, inducTIve transfer, knowledge transfer, context sensiTIve learning. Among them, migration learning is most closely related to multi-task learning (mulTI-task learning). Multitasking learns multiple different tasks at the same time, discovering hidden common features to help learn individual tasks.
What is migration learning?Transfer Learning (TL) is the migration of learned model parameters to a new model to help the new model training. Considering that most of the data or tasks are related, through the migration learning, the learned model parameters can be shared to the new model in some way to accelerate and optimize the learning efficiency of the model.
The basic motivation for migration learning:The basic motivation for migration learning is to try to apply the knowledge gained from one problem to another different but related problem. For example, a programmer who is proficient in C++ programming can quickly learn and master the Java language. To some extent, migration learning in machine learning has a certain correlation with the “learning ability migration†in psychology. In human evolution, it is very important to migrate to learn this ability. For example, it is easy for humans to ride a motorcycle after learning to ride a bicycle. It is much easier for humans to learn to play tennis after learning to play badminton. Humans can apply past knowledge and experience to different new scenarios, and there is an ability to adapt.
Migration learning main category method:1) Instance weighting method: The weighting calculation is performed on the training samples from the source domain in a certain way to determine the importance of each sample in the training process.
2) Common feature learning methods: transfer useful knowledge between source and target domains through several common features.
The importance of migration learning:1) From the data point of view: data is king, calculation is the core, but not enough data or collecting data is time consuming, it is difficult to label the data, and training the model with data is very cumbersome. How to learn machine? Migration learning is suitable for small data volume scenarios;
2) From the perspective of the model: the cloud-end fusion model is commonly used, and needs to be specifically adapted to the equipment, environment, and users. Personalized model adaptation is complex and requires different user privacy practices. Migration learning is suitable for personalization.
3) From the application point of view: the cold start problem in machine learning applications, the recommended system has no initial user data and cannot be accurately recommended. Migration learning can solve the cold start problem.
Migration learning reduces the reliance on calibrated data and better completes machine learning tasks by migrating with existing data models.
Migration learning implementation method:1) Instance-based Transfer Learning: Find data similar to the target domain in the dataset (source domain), multiply this data by multiples, and match the data of the target domain. It is characterized by the need to weight different examples; data training is required. Generally, the samples are weighted, giving a larger weight to the more important samples.
2) Feature-based Transfer Learning: By observing the common features between the source domain image and the target domain image, and then using the observed common features to automatically migrate between features of different levels. To migrate in the feature space, it is generally necessary to project the features of the source domain and the target domain into the same feature space.
3) Model-based Transfer Learning: Using a tens of millions of images to train an image recognition system, when encountering a new image field, there is no need to find tens of millions of images to train. Now, the original image recognition system can be moved to a new field, so in the new field, only tens of thousands of images can be used to obtain the same effect. One of the benefits of model migration is that it can be distinguished, that is, it can be combined with deep learning, and it can distinguish the degree of migration at different levels. Those with higher similarity are more likely to be migrated.
4) Relational Transfer Learning: The source domain is used to learn the logical relationship network and then applied to the target domain. Such as social networks, migration between social networks.
Migration learning tools:NanoNets (nano-network) is a simple and convenient cloud-based migration learning tool that contains a set of well-prepared pre-trained models, each with millions of trained parameters. Users can upload data online or through network search. NanoNets will automatically select the best pre-training model based on the problem to be solved, and build a NanoNets based on the model and adapt it to the user's data. The relationship between NanoNets and the pre-training model is shown in the figure below.
1) Separation of structure and content : When faced with a machine learning problem and want to discover the commonality between different problems, the structure and content of the problem can be separated. Although such separation is not easy, once it is completed, the ability of the system to be inferior is very strong.
2) Multi-level feature learning : The problem is divided into different levels, and some levels are easier to help with the migration of machine learning. With this hierarchical migration learning, different levels have different migration capabilities, and there is a quantitative estimate of migration capabilities at different levels. When it is necessary to deal with new tasks, it is possible to fix certain areas or certain levels, and to train other areas with small data, so that the effect of migration learning can be achieved.
3) Multi-step, transitive learning : Moving from the old domain to the new domain, moving from a multi-data domain to a less-data domain, this is called single-step migration. In many scenarios, multi-step conduction migration is required in stages, and a deep network can be constructed. This network middle layer can take care of the problem domain as well as the original domain. If there are some intermediate areas, then it can link the original field and the target field step by step. Two objective functions can be defined. When two objective functions work together, one optimizes the final target and the other selects the sample. With this iteration, the data in the original domain is migrated from multiple steps to the target domain.
4) Learn how to migrate : Given any migration learning problem, the system can automatically find the most appropriate algorithm in all the past tried algorithms, which can be feature-based, multi-layer network-based, sample-based. Or based on some kind of mixing. Or sum up the experience to train a new algorithm, the teacher of this algorithm is all these machine learning algorithms, articles, experiences and data. Therefore, learning how to migrate is like learning how to learn. This is the highest level of learning, that is, the acquisition of learning methods.
5) Migration learning as meta-learning : The migration learning itself is a method of meta-learning, which is given to different ways of learning. Suppose there was a problem with machine learning or a model. Now, as long as you have a cover for migration learning, it can become a model for migration learning.
6) Data generation migration learning : For the production-oriented confrontation network, the referee outside the Turing test is a student, and the machine inside is also a student. The purpose of the two people is to grow together in confrontation, and the two parties constantly stimulate each other to form A kind of confrontation (common learning characteristics). Small data can be used to generate a lot of simulation data, and the simulation data can be used to determine whether it is true or false, to stimulate the growth of the generative model. More data can be generated from small data, and migration learning can be achieved in new fields.
Recently, migration learning technology has been deeply studied in the field of machine learning and data mining.
Conclusion:
With the wave of machine learning in recent years, migration learning has become the hottest research direction. The future development of machine learning is small data, personalization, and reliability. That is migration learning. Migration learning embodies the unique analogy of human beings, and it is a divergent thinking of “one-for-threeâ€. Migration learning has been widely used in various artificial intelligence machine learning application scenarios.
Wireless Charging Coils,10W Wireless Charging Coil,Wireless Charging Coil For Cell Phone,Car Wireless Charging Coils
Shenzhen Sichuangge Magneto-electric Co. , Ltd , https://www.scginductor.com