In the second part, we will introduce the characteristics and advantages of PAI. First PAI has abundant machine learning algorithms built into it. PAI's algorithms have been accumulated by Alibaba Group's large scale business. They not only support basic algorithms, such as clustering and regression, but also support complex algorithms such as text analysis and feature processing. It supports not only traditional machine learning algorithms, but also a variety of deep learning algorithms. It also includes data pre-processing, feature engineering, statistical analysis, and other functions for data processing. On the basis of these algorithms, we can develop applications in vertical fields such as text, analysis, searching recommendation, image processing, etc. A key feature of PAI is visual modeling.

PAI supports the construction of machine learning experiments by dragging and dropping components to achieve zero code development of artificial intelligence services. The figure on the left is a complete machine learning task in PAI-Studio, which uses the training data to train the machine learning classification model, and uses the texting data to verify the training effect. From the data source to the machine learning classification model, to the prediction and evaluation, each step is encapsulated as visual components. Users only need to connect them together and set some parameters, and the model will be run with one click. 

The advantage of visual modeling is that there's no need to code manually. Users only need to understand the functions of each module so they can build machine learning tasks conveniently and quickly to complete a variety of practical applications. Even beginners can use it easily. PAI provides a very convenient one-click model deployment service. When the user generates the model with training, PAI-EAS can be used to deploy the model as a RESTful API interface with one click, and then invoke the service through HTTP request to realize connection between the model and the business.

PAI supports several prevailing deep learning frameworks such as TensorFlow, Caffe, PyTorch and mxnet. At the same time, PAI provides a powerful GPU computing cluster. TensorFlow and mxnet allow users to write their own Python code. And Caffe supports user defined network files. In order to use deep learning in the cloud, GPU resources need to be opened when you create a project. Users can select and switch the resource specifications of CPU and GPU according to the actual computing needs.

Finally, PAI supports interfacing with other products with Ali Clouds. The model's trained by PAI are stored directly in a big data computing service, MaxCompute, and can be used in conjunction with other Ali Cloud's products. From data processing to data application, it forms a complete closed loop. And each link has the accumulation of large scale practical application business, which can achieve good results in the application. 

Next, we will use PAI for artificial intelligence development and the traditional development way to compare. First of all, in terms of basic requirements, traditional development methods require developers to master the basic knowledge of Python and neural networks. For beginners, it takes a long time to master this knowledge. PAI development is suitable for developers of different levels. even if there's no code and machine learning foundation, but also through the virtual interface operation direct called PAI modules for development for developers with certain code and neural network foundation.

PAI also supports the development of self written code, custom algorithm package with specific functions, etc.. In summary PAI development does not impose too many restrictions on the developers basic knowledge requirements. In terms of develop method, the traditional development method requires developers to write code by themselves. For practical applications. The amount of engineering and code is often large, which is time consuming the barriers.

In addition to supporting self written code, PAI development also supports dragging and dropping development, which can construct models conveniently and quickly. And the visual modeling method is more clear and intuitive. Easy to understand. In terms of operation platform, traditional development method requires self configuration of development environment, and self installation of algorithm package, which often leads to a series of problems such as version incompatibility.

In addition, there's also high requirements of GPU and other hardwares, which requires considerable investment. It's not easy to deal with for general developers, especially beginners. PAI development is completely based on Ali cloud's platform, and you don't need to configure the development environment independently. It integrates a variety of mainstream deep learning frameworks, supports a variety of computing resources, as well as alternative payment methods, such as paying in advance and paying after usage. And the cost is controllable. In terms of supported algorithm library with traditional development method, you can only use the algorithm library inherit in the open source framework.

While PAI development has a large amount of algorithm library accumulated based on Alibaba's long-term practice. In terms of initial investment, the traditional development mode requires consideration of all hardwares and softwares investment, which usually costs a lot. However, through PAI development, you can make an investment according the actual demand of computing resources, which requires more investment due to higher allocation of computing resources and longer use time, and vice versa.