Migration learning, in a nutshell, is a technique that allows existing model algorithms to be applied to a new area and function with minor adjustments. This concept is currently relatively rare in machine learning, but in fact its potential can be quite large. Prof. Yang Qiang mentioned a vision in the just-concluded CCF-GAIR speech. Using migration learning, even a small company that has no conditions to obtain a large amount of training data can apply the models trained by large companies according to their own needs. , thus popularizing the application of AI. At present, everyone is trying to improve the versatility of artificial intelligence applications, the rise of migration learning has become more apparent. However, Prof. Yang Qiang started studying migration studies as early as 2009. At that time, he was one of the few scholars in China who studied transfer studies. In 2010, Professor Yang Qiang participated in the publication of a paper that explained migration learning in detail on the IEEE Transactions on knowledge and data engineering: A Survey on Transfer Learning, in which the concept of transfer learning and the difference between traditional methods of machine learning And some common migration learning methods are explained. Let us select some of the more representative parts of the paper (extracting + editing) to show everyone how this machine learning new trend differs from the traditional approach. Summary Many machine learning and data mining algorithms are based on an assumption that the training data and the data to be processed in the future are all in the same feature space and have the same distribution rules. However, in many applications in the real world, this assumption is likely to be untenable. For example, we often face the need to complete a classification task in one area but only have enough training data in another area. The data of the two may have different feature spaces or obey different data distribution rules. In this case, conducting a successful knowledge transfer can greatly enhance the learning effect, thus avoiding the labor of a large number of data tags. In recent years, migration learning has been proposed as a new learning framework to solve this problem. This article focuses on the research process of classification review of existing migration learning projects to solve classification, regression, and clustering problems. In this article, we will also discuss migration learning and other related machine learning techniques such as domain adaptation, multitasking learning, sample selection bias, and covariate transformation. We will also explore some of the more promising ways to study migration studies in the future. Introduction: Data mining and machine learning techniques have achieved considerable success in the field of knowledge engineering including classification, regression, and clustering. However, when the data distribution pattern changes, most statistical models need to be rebuilt using new training data. In many applications in the real world, the cost of doing so is very large or even impossible. Therefore, reducing the necessity and workload of re-collecting training data becomes a very necessary matter. In other words, the knowledge transfer or transfer learning between different task areas can achieve satisfactory results. In many cases of knowledge engineering, migration learning is indeed very useful. For example, in web page classification tasks, the data characteristics and distribution of newly created web pages may be different from those used in previous training. Therefore, migration will be required at this time. Learn techniques to transfer models. If the training data is easily overdue, that is, the distribution of the data may be different at different time periods, then migration learning will also be needed to make the model not invalid. For example, in the indoor WiFi positioning problem, a technique for detecting the user's current location based on the collected user's WiFi usage data, but due to the WiFi signal strength is entirely possible with time, the user's equipment and some other factors. Changes in the changes, a static model obviously can not cope with this problem. The third example is: the problem of emotional classification. The method of collecting data in this problem is to collect and categorize the narration of a large number of equipment users, but because the distribution results of data generated by different products may be very different, if you want to use traditional methods with good enough classification results, you may need Different models are created for different devices, but the cost of doing so is obviously too great. Therefore, it is better to have a method that can establish a common model for each device. An introduction to the migration history and a comparison with traditional methods: Traditional data mining and machine learning algorithms can use the statistical model trained with previously marked or untagged data to make predictions about future data. Semi-supervised learning can solve the problem of too little data that can be used to build an available classifier by training with a small amount of marked data and a large amount of unlabeled data. Variants of supervised learning and semi-supervised learning to deal with imperfect training data have been studied in detail. Many good results have also been obtained, but many of them have the same distribution rules based on marked and unlabeled data. In contrast, migration learning allows the fields, tasks, and distribution patterns of data sets used for training and testing to differ. In the real world, we have observed many cases of transfer learning. For example, our process of understanding Apple may also help us to understand pears. Similarly, learning to play the keyboard may also help learn to play the piano. The study of migration learning was initiated because we discovered the fact that humans have the knowledge they learned before using this to solve new problems faster or better. The fundamental impetus for studying transfer learning in machine learning came from a seminar at the topic of “learning to learn†to develop a discussion of “ultimate†machine learning methods that can preserve and reuse existing knowledge. The study of transfer learning began in 1995 and has begun to attract more and more attention while changing names one by one: learning to learn, life-long learning, and knowledge transfer (knowledge transfer), inductive transfer, multi-task learning, knowledge consolidation, context-sensitive learning, knowledge-based inductive bias , meta learning and value-added/cumulative learning. Among them, a technology closely related to migration learning is a multitasking learning framework. The goal of this framework is to learn multiple tasks at the same time, even if they are different from each other. A typical multitasking learning approach is to discover some (potential) common rules among these separate tasks. In 2005, the announcement (BAA) issued by the Department of Defense Information Processing Technology Office (IPTO) of the US Department of Defense Advanced Research Projects Agency (DARPA) gave a new mission to migration learning: one that can recognize and will learn from previous tasks. The system of knowledge applied to new tasks. Under this definition, the goal of migration learning is to extract knowledge from one or more tasks and apply it to another goal task. Compared with multi-tasking learning, which focuses more on learning all sources at the same time, migration learning is more focused on the target task. The relationship between source and target in migration learning is no longer equal. Figure 1 shows the difference between the traditional technology and the migration process. We can see that when high-quality training data is not enough, traditional machine learning technology learns more from random tasks to learn from tasks, while migration learning is based on learning from previous tasks. Today, migration learning methods have been applied in several high-end areas. Especially in data mining, machine learning and its applications. Several migration learning classification tables, charts in conclusion: In this article, we review several trends in the current migration learning arena: There are three different types of migration learning: inductive migration learning, transmigration migration learning, and unsupervised migration learning. Most of the previous work focused on the first two categories. . However, in the future unsupervised transfer learning may receive more and more attention. Not only that, the specific implementation of the migration learning method can also be divided into four categories according to "what to migrate": immediate migration, representative feature migration, parameter migration and related knowledge transfer. The data in the first three environments satisfies the iid hypothesis, while the last one focuses on the relevant data in the migration study. Most of these methods assume that the selected source domain and target domain are related. In the future, there are still several important issues that need to be addressed in research. First: How to avoid negative transfers is still an open question. Many current migration learning algorithms assume that the source domain and the target domain have some degree of correlation. However, if this assumption is not true, there may be a negative transfer problem. It may cause the transferred algorithm to perform worse than before the transfer. Therefore, this is a problem that must be solved in the migration study. To avoid this situation, we must first assess the transferability of the source domain to the target domain. To define the transferability between the two domains, we can use the corresponding method to measure, and we also need a method to measure the similarity between source domains and tasks. And a related question is: If the whole of a domain cannot be used for migration learning, do we still have the opportunity to transfer some of the domain to help learning in new areas? In addition, most of the existing migration learning algorithms focus on generalizing migration methods between different source and target domains or distributions of tasks. To do this, they will assume that the feature space of the source domain and the target domain is the same. However, in many practical applications, we may need to transfer between source domains and target domains that have different feature spaces, and it may be necessary to transfer from multiple such source domains at the same time. Our management of this kind of migration learning is called "heterogeneous transfer learning". Ultimately, current migration learning technologies are used in small-scale applications with limited variables, such as sensor network-based positioning, text classification, and image classification issues. In the future, migration learning will be widely used to solve other challenging challenges. Applications, such as video classification, social network analysis, and logical reasoning. As mentioned above, most of the current AI research is conducted with the goal of making AI more adaptable and even capable of real learning. Migration learning is a technology that can achieve this goal to a large extent. Imagine if an AI has the same learning ability as the same person and can apply the experience learned in old things to new things. Even if the AI ​​is stupid at first, it should be able to become more and more through continuous learning. The smarter. Is this also a kind of intelligence? But no matter what, migration learning seems to be a very promising method. I believe we can shine in the near future.
XB4 series Pushbutton Switch are used in industrial for controlling circuits of AC50Hz or 60Hz,rated operation voltage 380V or below and DC operational Voltage 220V or below for controlling in magnetic starter, contactor,relay and other electrical circuits. The buttons indicators can also be used in the place with indicated light or signal. They meet with the standards: IEC 60947-5-1.
XB4 Series Metal Pushbutton Switch is the new design in the field of Push Button Switch. Which with stronger metal plate and very easy to be installed.
XB4 Series Pushbtton Switch,Push Button Power Switch,Mini Push Button Switch,Push Button Lamp Switch Ningbo Bond Industrial Electric Co., Ltd. , https://www.bondelectro.com
A Survey on Transfer Learning
Is perfect migration learning a kind of AI?
