Zhen Zi, head of intelligent direction of Alibaba Economic Front Committee, head of D2C intelligence team of Tao Department.
Dao: Why do front-end intelligence
At first, the direction of front end intelligence was proposed to bring about changes in front end technology and expand the front end with the help of AI and machine learning capabilities, just like the clairvoyant in myth with telescope, and the “super power” with AI front end will also be generated on the basis of their own abilities.
Later, while maintaining this goal drive, it is difficult to land. By thinking about AI’s capabilities and characteristics and the current situation of the front end, the author decided to “solve the problems of front-line r&d personnel and improve their happiness”. As a starting point, reconstruct intelligent strategy and tactics system.
Finally, the pointcut is to design the generated code. Today, as the building system is becoming more mature and the modular development mode is deeply rooted in people’s hearts, a large number of demand changes and complex business environment cause that the front end cannot simply solve the problem with the idea of “reuse”. Therefore, the front end intelligence is used to identify and understand the design draft, and then the code is generated through rules. Design to Code (D2C) can well solve the problems of daily module development.
How to make the front end intelligent
The tao gave birth to a
D2C is applied in the C-terminal business of Taoshi e-commerce, supporting the r&d of double promotion venue module 0, combining with intelligent UI and intelligent venue, generating venue intelligently based on venue planning constraints, venue template and operation configuration, and providing different UI and interaction modes for users in different circles. In addition, intelligent card insertion, intelligent rights pop-up retention ability, multi-pronged, truly achieve front-end intelligent technology enabling business, on the basis of research and development and improvement of efficiency, to make user value and business value.
Second life
Design+ to Code (D+2C)
Based on the recognition and understanding of the original design draft, PD’s description of the data, functions and interactions on the design draft (original PD’s wireframes and interaction drafts, etc.) is integrated to further improve the input of the model and generate more logical codes to deal with more complex business scenarios.
Pipcook
To solve the problem of using machine learning cost at the front end, 0 cost helps the front end quickly master and use OTA algorithm capability of machine learning. At the same time, help the business team to train their own model to ensure that the model can be self-iterative and more accurate in their own business scenarios.
Two gave birth to three
Do publicity:
Because people are unfamiliar with AI, they have concerns and doubts about front-end intelligence, which is difficult to scale promotion. From tao department business, auction, health, ants, Alipay, sports, CBU… Practical experience and cases are collected and sorted out in the landed scenarios to form an easy-to-understand series of articles to help everyone understand AI and front-end intelligence. Further implementation of promotion, so that front-end intelligence can become a universal technology in the front-end field, so that front-line front-end r&d personnel get happiness.
Good application:
Imitation is an efficient way of learning, and machine learning essentially imitates from human experience. Practice in more fields and form more sample projects are the key matters of good application. To have a clear and clear business domain map, technical domain map, engineering domain map, through the way of coloring, bit by bit in the map to expand the application of the field. The way to expand from us to use, we do a model, we imitate and derive more excellent application cases. At the same time, strengthen the link between the application field and the core field, fully reflux the application idea, application method, application ability…… Feed back to the core field, and drive lean iteration of the core field with data based on objective requirements from actual application scenarios.
Good technology:
Continuous cultivation of intelligent technology system is also one of the key points in the future. Today’s intelligent technology system still has a lot of problems, running too fast, in order to solve the problem, many applications and the coupling between the core technology, many of the open ability and customized ability is insufficient, many of the theoretical foundation of the technology is not solid, many of the engineering system of the technology is not complete. In addition, in the general direction of “human-computer collaborative programming”, how to create standardized, data-oriented and automated links with PD, design, operation and users? Truly support the application system to reduce the cost of understanding and using technology, achieve: requirements and code, requirements and production, collaboration online, so as to further, full link, end-to-end improvement of business delivery capabilities.
Our vision-be
Front-end technology system upgrade:
Upgrade D2C technology system, introduce S2C capability, and form P2C end-to-end business delivery platform: In stateless simple UI code and the front-end code generation based on business logic, to upgrade to the complex UI code and complex business logic code generation, at the same time, the complex in the UI code to recognize and deal with state generated with change of state of the UI logic code, client side to invoke JSBridge, data, and the business logic of the service interface layer of glue code, The server invokes the business logic glue layer code of the data and service interfaces.
The intelligent UI technology system is upgraded, and the capability is lowered. Meanwhile, the intelligent UI customization of channel business is supported. The partial business is handed over to the P2C end-to-end business delivery system with the technology convergence. The three-pronged approach ensures that the intelligent UI technology system is universal, cohesive and controllable, so as to perform attribution analysis and iterate in a data-driven manner in a wide range of business scenarios.
The front-end intelligent algorithm engineering system is upgraded, and the introduction of cloud native capability enables the intelligent algorithm framework Pipcook to support cloud native from the data side, model side, training side and deployment side, which can be integrated into the cloud native system more conveniently, ensuring that users can directly online model algorithm to support business and open up the full link of business.
Upgrade of business R&D mode:
P2C business RESEARCH and development platform building: demand and code, demand and production, collaboration online brand new business research and development model. Get through the whole link of demand, design and RESEARCH and development, ensure the integrity, rapidness and consistency of information flow, ensure the consistency of structured information of final deliverables and requirements, and data-driven demand iteration.
Design sinking and asynchronization, from the previous design for demand evolution to design for business, from the design trend and social development, technological progress three perspectives, based on the design language, design system as the norm, to provide a design paradigm, the model according to the design paradigm for automatic design and production. N UI schemes are derived from a UI scheme, and N Elements are derived from an Element.
The server side sinks and abstractions and atomizes, providing standard, atomic domain capabilities that are understood by the S2C architecture and generated by the requirements of the call logic: glue layer code.
Front-end capability model upgrade
Upgrade of programming ideas
- deterministic
First, the process of intelligent thinking to solve problems is simple, and simple process represents certainty: there must be an answer. In fact, this is just like the perplexity of life. Everyone has his own answer. Even if we tell the model the so-called correct answer, the model may not give us the expected answer to the unknown question after training, but the model will certainly give us an answer.
Secondly, from the perspective of experience, as long as it is a good problem solved in the industry, it should be able to solve it in their own field. For example, using the image classification model, rigorous experiments have proved that the image classification model can accurately identify the presence of a dog or cat in a picture. If I label my pictures of cats and dogs and feed them to the model, the model can be trained to recognize other people’s pictures of cats and dogs, which is another level of certainty.
Finally, take a realistic example, for example, if you come to a new team and you can’t put names to faces of all the people, after a period of time together, you talk to each other every day, eat and drink together, eventually you can remember everyone and recognize the name of any person you see. The model, too, initially failed to remember different people’s faces, but with the names of each person, from every Angle, and in every light condition, the model, with proper training, was able to pronounce each person’s name as we do, and that was certainty.
- robustness
First of all, the method of intelligent thinking to solve the problem is very simple. After telling the model the correct answer, the model trains the parameters and weights of the model in the sample data, and summarizes the idea behind the answer by itself, which is very robust. To return to the problems encountered by Da Vinci with OpenCV: Threshold fallacy. It is difficult to summarize and extract the features of image, Text and Controller in all cases. However, when there are correct answers of large enough sample data, the model can extract sufficient robust answers.
Secondly, if you are interested in algorithms, you can Google genetic algorithm and ant colony algorithm, etc. You will find that algorithms with strong robustness are beyond our expectation. To put it simply, it is difficult for people to summarize and refine the patterns and ideas behind these algorithms. However, That doesn’t stop us from writing genetic algorithms and ant colony algorithms to train these robust patterns and algorithms in simulated or real-world environments, which is inherently robust. Recall that in the past, we used to think before we wrote code. Today, when we solve problems with intelligent thinking, we can write code without thinking clearly. Isn’t software development robust enough?
- evolutionary
At least 10 years ago, I discussed with my colleagues the “self-healing” problem of the system. Ten years have passed, and few companies are close to this goal. But today, with intelligent thinking, they can do it! I’ll start with the evolution of intelligent thinking, and finally share some of my explorations and reflections on the “self-healing” of systems.
The evolution of intelligent thinking is reflected in the way we solve problems. In the past, we determined to understand the problem ideas before development, and realized the problem solving ideas in the development process, which is Hardcode in plain English. When faced with problems that go beyond flexible design, we still have to write new code, constantly patching our “ideas”. Intelligent thinking to solve the problem, we do not provide their thinking, model is extracted from the correct answer from their thinking, once encountered new situations we can put these things as a new feeding model, the model can self evolved, and similar problems to solve, this process is to realize the self evolution, The only thing we need to do is form this evolutionary loop: evaluate model answers, generate positive and negative samples of new answers, and construct pathways for online training.
Upgrading machine learning capabilities
Machine learning related frameworks, libraries, packages that can use Pipcook or Python technology ecology…
Capable of data acquisition, data processing, model configuration, model training, model verification, model deployment, data backflow closed-loop, able to use Pipcook to configure Pipline for streaming computing, processing massive data.
Can debug model, know how to evaluate model, locate model problems, tune model, compress pruning and knowledge distillation to reduce model algorithm complexity and improve model performance.
Key indicators of intelligence
Model selection of model function indicators: business is used to serve users, and the understanding of users can help us examine the correctness of business understanding. Through user understanding, user tasks can be determined. By understanding the user’s tasks, the model’s tasks can be determined; By understanding the model’s task, you can determine the model’s functionality. Only when the function of the model is determined, can the model be selected correctly. By sorting out the functional indicators of the model, these indicators can be monitored to judge whether the model is correct.
Model accuracy in model understanding: Understand what the model should do? Then the metrics can be set through the business feedback data from the model. The accuracy of the model in dealing with known problems, similar problems and brand new problems can be judged according to the performance of the index and storage business, similar business and brand new business respectively. The accuracy of similar problems and brand new problems is the generalization ability of the model. The accuracy index of the model can be judged more comprehensively through the accuracy data of the model itself and the evaluation indexes of the feedback data in the business.
Model generalization ability of accuracy understanding: The model generalization ability is the accuracy to deal with similar problems and new problems. Similar questions are easier to understand, just like the literal “same type” and “similar” questions. The accuracy of new problems should be understood as: undiscovered, unintuitive, difficult to understand, but behind the common problems. It should not be understood as: a problem of arbitrary creation and having nothing to do with experience. Because machine learning today is weak ARTIFICIAL intelligence, lacking the ability to perceive, understand and create unknown areas. When using accuracy and generalization capability indicators, the boundaries of each indicator must be clearly defined. Accuracy index is used to evaluate model discovery (recall) and problem solving ability (model deductive ability), while generalization ability index is used to evaluate model solving ability of non-training set problem (model generalization ability).
The direction of front-end intelligence is still developing rapidly and we are still on the way. We will continue to explore the use of intelligent ability to solve the problems of front-line research and development personnel, improve the happiness of front-line research and development personnel, we are full of confidence.