Artificial intelligence has been detonated by Google in the past two years, so startups in this field have sprung up, some special chips, some algorithms, some robots ... What is artificial intelligence? Is it for machines to replace humans for repeated production line labor? This is too simple to understand. Artificial intelligence definitely does not just allow machines to complete imitations, thereby replacing repetitive work. Real artificial intelligence requires machines to be trained and learned to approach or surpass human intelligent behavior. When artificial intelligence trains models, data is the most important factor. Only by training with big data can the trained model be closer to the actual needs. For example, the model training for autonomous driving is now basically based on autonomous vehicles driving on the road to collect data to train the model. Only when the cars on the road collect more possible scenarios can we ensure that the trained model is closer to the actual road conditions. There are generally two problems encountered in artificial intelligence training: First, how to train the model when the amount of data is insufficient? Secondly, how to quickly annotate too much data? MathWorks Product Marketing Manager Zhao Zhihong Transfer learning: Can the artificial neural network be trained by generating data from the model if there is no data? This is a confusion for many startups. In the author's concept, the normal AI development process is to first collect data and then train the model. At the recent Matlab conference, Mr. Zhao Zhihong, MathWorks product marketing manager, gave a different answer. He said, "Data, output, and models are a step in the entire AI development process. If developers want to do artificial intelligence, there is no Too much data, you can use the method of transfer learning. "What is the concept of transfer learning? Do not understand from practical examples. For example: to predict the failure of wind turbines through artificial intelligence, users cannot wait to collect a large amount of failure data before predicting, which does not meet the goal of equipment maintenance, but first use Simulink model to build a model of wind turbines, Through correction, this model is very close to the wind turbine, and then the fault data is generated through the model, and then the neural network for machine learning or deep learning is trained. Mr. Zhao Zhihong said, "The advantage of model-based design is that it builds a model before actually making the product. This model is very close to the actual product, whether it is from the results of digital calculations and behavioral calculations and the real results. Very close. Many engineers have already built this model when designing products. It is easy to generate failure data by adding failure conditions to this model, which is much easier than failures that occur on actual equipment. Therefore, failures can be carried out Forecast and maintenance. " Marking big data through deep learning With the large-scale adoption of sensors, the amount of data has entered the era of flooding, which has brought difficulties to artificial intelligence feature labeling. At this time, deep learning can be used for labeling. Mr. Zhao Zhihong explained, "The characteristic of deep learning is that there is no need to manually find the feature values. The system can automatically extract the feature values ​​from the data. There are many tools and functions for automatic labeling in MATLAB. LiDAR 3D point cloud technology can be used for each point. Annotation, cluster this point together into a target model, and then identify the actual object represented by the target. Some users have developed using MATLAB tools, and the famous auto parts company AUTOLIV is using this method Automatic data annotation. " Taking intelligent detection of puffed food as an example, researchers can extract features when users bite the food, and use the bite sound and bite force to measure the crispness of the food. With these two characteristics, a machine learning classifier needs to be developed. And MATLAB provides a classification learner. With this tool, developers do not need to try the algorithms of various classifiers one by one, and use the APP provided by MATLAB to try all the algorithms at once. After starting to run the APP, the user selects the data and the classifier to be trained, and then trains. During the training process, users can see the overall results of each classifier, select the one with the highest accuracy, and then conduct more investigations and research. If users have not studied AI, they can use the APP provided by MATLAB to learn and try all machine learning algorithms. Is it difficult to do deep learning in MATLAB? In the words of Mr. Zhao Zhihong, only need to write 5 lines of MATLAB code, users can build a network that can recognize food. The author has not been in contact with MATLAB for many years, so it is impossible to verify. Engineer friends can try it if necessary. How to collaborate with open source code? Open source on the market is very hot now. Engineers can quickly obtain the required code in the open source community. There are also many open source resources in the field of artificial intelligence. MATLAB as a closed resource may be slower to update than open source code, and plus With the charging model, engineers also worry about the integration of the two. When the author asked this question, Mr. Zhao Zhihong explained, "There are a lot of development teams behind the open source code. They develop voluntarily. We also have our own development team. We do not define open source code and MATLAB as a competitive relationship. In fact, We have their own characteristics and strengths. Sometimes we can share with open source code. Some open source code can be run in MATLAB or directly called, we are contributing to the progress of science and technology. Although we are paid software, but not only provide Tools, and many services. " "When users encounter problems when using MATLAB, we have a strong technical support team behind to help users solve these problems. In addition, our tool not only produces the correct results for a certain function, we also consider to make this function very It is easy to use. For example, for AI applications, we have made the interface very suitable for the professional field, and also give the user a feeling of using the familiar MATLAB or Simulink interface. This is our advantage. If the user adds new in MATLAB or Simulink The function, the learning threshold will not be very high, and it is easy to get started. In addition to providing tools, we also provide users with a lot of convenience, which is something that can not be obtained from open source tools. "Mr. Zhao Zhihong added. 20 Awg Tinned Copper Wire,Tinned Copper Conductor,Stranded Tinned Copper Wire,Tinned Copper Wire Price Sowell Electric CO., LTD. , https://www.sowellsolar.com