Course outline
In recent years, with the accumulation of big data, the innovation of theoretical algorithms, and the improvement of computing power, artificial intelligence has once again attracted extensive attention from the academic and industrial circles, and has made breakthroughs in many application fields. However, custom models often require deep learning models built by AI algorithm scientists, and continuous training and testing. In practice, the business side often has to continuously improve the model as scenarios and data change. At this point, the ability to quickly customize and deliver AI at low cost is critical. So what are jd’s explorations in AI? What is the whole process of AI development? How can model development be accelerated? This article will answer these questions.
Click on the”readingView the course video!
Quickly build AI capabilities based on model training platform
— Jd Zhilian Cloud Zhu Ertao —
AI capability development process
At present, the field of artificial intelligence has not formed a perfect frame of reference. However, the White Paper on STANDARDIZATION of ARTIFICIAL Intelligence (2018 edition), based on the development status and application characteristics of ARTIFICIAL intelligence, puts forward an ARTIFICIAL intelligence reference framework from the perspective of INFORMATION flow of ARTIFICIAL intelligence, which describes the generation of the entire AI capability.
Infrastructure Provider (Computing)
Infrastructure providers provide computing power support for AI systems, enable services to communicate with the outside world, and support through the underlying platform. The computing power on AI is usually provided by smart chips; Communication with the outside world is through new sensors, such as cameras, microphones and various Internet-of-things devices that collect terminal data for business use; Basic platforms include distributed computing framework providers and network providers to provide platform protection and support, including cloud storage and computing, and interconnection networks. For example, JD Zhilian cloud is the technical platform for everyone, including computing power support, cloud main online storage, micro services, etc., and can also create an overall service platform based on the entire product system.
Information provider (data)
Information provider is the source of intelligent information in the field of artificial intelligence. Through the knowledge information perception process, the data provider provides intelligent perception information, including original data resources and data sets. The perception of original data resources involves the recognition of graphics, images, voice and text, as well as the Internet of Things data of traditional devices, including the service data of existing systems and the perception data of force, displacement, liquid level, temperature and humidity.
Information processor (algorithm)
Most people’s understanding of ARTIFICIAL intelligence may be more inclined to algorithm, such as data cleaning, extraction of feature values, etc., which is equivalent to a process of data modeling, based on which some model training is carried out. And then provide intelligent reasoning and intelligent decision support for business scenarios. Intelligent reasoning and intelligent decision-making will eventually output valuable intelligent information, which comes to the “intelligent execution and output” link, which outputs the results of the entire intelligent information flow process, including functions such as movement, display, sound, interaction and synthesis, which are also the results of the entire AI capability.
01 Three major ai chips
Chips for cloud inference are also what we call the second type of chips. At present, mainstream artificial intelligence applications need to provide services through the cloud, transfer the collected data to the cloud server, use the CPU/GPU/TPU of the server to process the inference task, and then return the data to the terminal.
The third type of chip runs on the device side, including smart phones, smart security cameras, robots, autonomous driving, VR, etc. It can make some inferences and decisions quickly, and is not affected by the network. How to understand? Let’s say we have a camera, and if we want to send all the data that this camera collects to the server for processing, we can imagine how many points of network bandwidth it would take. However, if we put some data processing tasks on the terminal to complete, and only upload the processing results, so the requirements on the network will be much lower, and the user experience can be significantly improved.
02 How to measure the quality of data?
Data is the starting point of our success. In ai, there are two main types of data: structured data and unstructured data. Structured data is highly organized and neatly formatted data. It is the data type that can be put into tables and spreadsheets, can be queried with SQL, and is suitable for machine learning, data analysis and mining.
Unstructured data cannot be represented as a table, and there is no unified format for it. This data is typically stored in a non-relational database and queried using NoSQL, making it suitable for deep learning. After the development of the Internet, the amount of unstructured data has been increasing, and now 80% of enterprise data is unstructured data.
For enterprises, there are usually three ways to obtain data:
- Buy from industry data providers
- Self-accumulation or collection, including industry accumulation, crawler, etc
- Obtain industry data from partners
The more accurate the data is and the more matching it is with its own business, the closer the results of business inference using the data training model will be to the real scene, thus bringing greater value to the enterprise itself. How to evaluate the quality of the data? There are four latitudes:
- A) Relevancy B) Relevancy
- Recency
- Range: The Range affects the effect
- Reliability: Marked by professionals
Simply speaking, the higher the degree of relevance between data and their own business, the better, this is the primary standard. Also care about the timeliness of the data, and the needs of users. For credibility, there are professionals corresponding to different business scenarios, who can develop industry standards and annotate data to a large extent to ensure the quality of business data.
Machine learning process disassembly
In the whole process of forging AI capabilities, algorithm research is one of the steps, but also an important step. Machine learning is a new category of traditional artificial intelligence, which focuses on how to simulate and implement human learning habits. In simple terms, machine learning uses algorithms to train models and use those models to identify and predict new problems.
Training data collection
Raw data is the first step in the machine learning process. It is collected from various sources. Data sets are usually divided into two parts: one for training, the training set; The other part is for validation, which is the test set.
Data preprocessing
As the collected data is often rough or noisy, it is necessary to preprocess these data, including desensitization of business-sensitive data, and cleaning, format conversion, or feature extraction of unqualified data to obtain effective training data.
Training prediction model
When it comes to training predictive models, algorithm engineers can step in. Before formally starting the model training, we need to select the appropriate algorithm combining the characteristics of data and business. Machine learning can be divided into: classification, regression, clustering, anomaly detection and so on. In the early stage, algorithm engineers need to do some Demo tests among possible algorithms through the test set and training set, and then select specific algorithms according to the test results, so as to avoid the loss caused by large-scale training model changes.
Model to evaluate
After selecting a model, we need to verify whether the following model is qualified. How do you evaluate it? This is to test the model using the test set described earlier. Because the test set is completely new to the model, you can objectively measure how the model performs in the real world. In fact, this process is a cyclic iterative process, in which we will transform the algorithm or adjust the parameters, and finally get an ideal model. We usually make this model into a service for people to use.
Take JINGdong Zhilian Cloud as an example. Jingdong Zhilian Cloud provides trained models into online service APIS like the one shown in the figure below. It’s split into several directions, each with a generic AI capability for everyone to apply. For example, in the retail field, some of JD’s leading algorithm models will be encapsulated into a general API and provided to users. Each API provides some interface for calling methods that users can apply for online. Of course, we also provide offline version or SDK to facilitate in-depth integration with the user’s business.
AI capability development process
Jd.com’s Zhaopin Cloud artificial intelligence department has launched a product called NeuFoundry. As the best practice of JD to enrich scenarios, NeuFoundry provides a one-stop AI development platform from data annotation – model development – model training – model evaluation – model release to improve the efficiency of AI development and reduce development costs. The platform not only supports image and text deep learning, but also machine learning, such as sales prediction. In addition, it will train massive data through preset models to make the intelligent supply chain more intelligent.
The main services NeuFoundry provides include:
Data management
It provides massive storage space in the form of database and object storage, stores and manages user data, and provides data annotation, data analysis and other functions.
AI capability customization
Customized AI capabilities can be trained according to the capabilities and scenarios required by the actual business and combined with the actual business data of the enterprise;
AI Service Support
The intelligent Middle Platform provides the basic environment support for the operation of various AI capabilities, and ensures the stability and reliability of AI services.
Among NeuFoundry’s data management services, the platform provides data set management and annotation services with built-in public and custom industry data sets; In addition, tools are provided to enable users to annotate online or import existing annotated data, including common texts, images, videos, and audio files.
In the model training tasks provided by NeuFoundry, the platform supports common engine frameworks such as TensorFlow, Caffe, And PyTorch, as well as multiple project execution modes, and provides a variety of interaction modes for different types of developers. For example, we provide task submission and Jupyter NoteBook for algorithm engineers who are proficient in algorithms. For business developers, it provides convenient and rapid customization of AI models through graphical drag and drop or automated training, without the need to master the ability of algorithm development.
Why NeuFoundry can help companies accelerate AI capabilities? Because in NeuFoundry, we preload some industry-leading algorithm models for common business scenarios. As long as users put up their own business data, through automated training, these algorithms can be used to quickly iterate their own customized models. Typically, users can create their own API capabilities in just two steps on NeuFoundry:
(1) Select the scene and data and submit the task;
(2) View the model evaluation results and publish them as services.
Create API
In fact, AI capability building is only a part of “intelligent platform”. Jd Zhilian cloud will provide enterprises with solutions in different fields from business, technology, data and other perspectives based on the financial field and specific application scenarios.
In the financial AI scenario architecture diagram, based on the underlying operational resources and data resources, the development layers of different application scenarios are basically the same. Looking up, there is a set of preset services including traffic analysis, content moderation, user behavior analysis, intelligent decision making, user profiling, and risk control, which are optimized and customized for the financial scenario.
Through learning this course, we have learned how to develop AI capabilities, and jd Zhaopin cloud’s exploration and capabilities in this area. NeuFoundry — a one-stop AI capability foundry is based on JD’s retail and retail infrastructure, rich scenes and mass data polishing the best practice of AI development platform. Accumulated mature one-stop model development tools and high-quality data sets, and precipitation for mature models and AI services. NeuFoundry can provide a one-stop development platform for small and medium enterprises, helping enterprises build their own intelligent platform quickly and at low cost to complete the intelligent transformation.
Click on the”readingView the course video!
Q&A
Q NeuFoundry different framework, model training after deployment, what deployment support?
A NeuFoundry offers A JupyterNotebook environment where users can train and deploy in A Juypter environment. NeuFoundry can help users publish their models as an online service if they are trained on automated tasks provided by NeuFoundry.
Q For this epidemic, the mask correction model, face recognition with large area of occlusion, epidemic prevention and control consulting customer service and so on have been derived. I want to know how to control the accuracy of these data acquisition and model?
A Let me explain this to you. In AI model training, the more relevant the training data is to the business, the better the accuracy or effect of the training will be. If your business is in some public places, or the lighting or camera is not good, of course, there will be blocking of faces and so on. Then we should do targeted data collection for this scenario and find typical data. And then we do some optimization and training for this data. This model, then, works better for your scenario. There is no one model that fits all scenarios, there are some areas of applicability.
Q sample is relatively small, how to train, the effect can be good?
A This is A typical problem encountered by many enterprises when they want to do AI construction, that is, their business data may not be easy to extract or refine. I also mentioned this just now. If you don’t have enough accumulation in this industry, you can do some procurement in this industry. As for the requirement of data magnitude, take image classification as an example, when the magnitude is about several hundred to one thousand, its accuracy can meet the business needs to a certain extent, and the improvement effect of increasing data is not so effective. If your business scene and JINGdong zhilian cloud this piece of ability is more suitable, then we can also help you to do some data preparation, or industry scene support.
