Shulou(Shulou.com)06/02 Report--

Artificial intelligence, an old topic, reviews the history of yesterday's rise and fall, and after more than 60 years of development, it has finally entered the golden age, and its application has gone deep into various fields of enterprises. Up to now, artificial intelligence has entered the third stage, which is inseparable from the combination of artificial intelligence and application scenarios. We all say that we must embrace AI, but how to embrace it?

On October 17th, the 10th China system Architects Conference, with the theme of "Ten-year Architecture, the Road to growth", gathered domestic technical groups such as CTO, R & D director, senior system architect, development engineer and IT manager. Second, the Daily main Line 4 artificial Intelligence (part two) special show, senior technical experts from iqiyi, a little information, Zhihu and Xiaomi shared the application and exploration of AI in their respective enterprise scenes. How can you miss this door-to-door welfare?

Liu Guohui, head of iqiyi Advertising algorithm: application and Exploration of AI in iqiyi Commercial Advertising

Advertising is an important means to realize Internet traffic, and it is also one of the most successful application scenarios of AI technology in industry. In this lecture, Mr. Liu Guohui introduced the business characteristics and core challenges of iqiyi video advertising, and shared how to use AI technology to build iqiyi intelligent advertising algorithm engine. The content includes the construction method of industrial algorithm engine, how to improve the operation efficiency of advertisers through inventory estimation, intelligent inquiry volume and suggested bidding in the advertising sales phase. in the advertising execution stage, how to improve the advertising effect through global optimal allocation and personalized recommendation, as well as iqiyi's intelligent exploration of balancing business goals and user experience.

In order to better support business demands, iqiyi built a set of intelligent advertising algorithm system framework, the system framework is mainly divided into two parts, in the advertising sales link, will provide advertisers with a wealth of intelligent delivery tools; in the advertising delivery link, iqiyi advertising algorithm team built a real-time delivery engine.

At the algorithm level, the framework is divided into three parts. In the middle is the research of some basic algorithms, the first part is the inventory allocation system of iqiyi brand, the core of the system is guaranteed quantity, and the latter part is a set of effective personalized recommendation system, which mainly does some pre-advertising recall.

Teacher Liu said that there are many links involved in the whole machine learning application, and there are many problems to be solved before finally starting the adjustment of the algorithm, which will affect the final effect of the algorithm. It is mainly reflected in three aspects: at the business level, the business is complex and the business objectives can not be directly optimized; at the data level, the efficient use and analysis of massive data becomes a problem; at the engineering level, how to ensure offline engineering quality and online consistency becomes a problem.

Tian Mingjun, Senior Technical Director of one Point Information: adaptive recall in Personalized recommendation system

Recall is a very important part of personalized recommendation system, which directly affects the quality of the final output recommendation results. Modern recommendation systems usually include a variety of recall methods, content-based recall, recall based on user behavior, recall based on user attributes and usage scenarios, recall driven by deep learning model, and so on. Limited by the computing power of the upstream sorting phase, the number from the recall to the sorting phase is usually limited. For different product lines and different users, the importance of value generated by different recalls is also different, so it is necessary to dynamically determine the combination of recalls and the matching mode of quota. In this lecture, Mr. Tian Mingjun shared a little information about some of the work on systems and algorithms for this challenge, as well as the impact of related work on the quality of the final recommendation.

Teacher Tian Mingjun divides the adaptive recall into intention analysis and unified intention expansion, recall selection and dynamic selection recall service, query builder automatic assembly recall request, configuration center flexible control of recall scope, user status, use of scene-assisted regulation and control of recall quota allocation.

Its later work mainly includes two aspects: on the one hand, it can improve the recall ability of model2news, reduce the recall path, optimize the iterative recall method based on machine learning, improve the unique docs coverage, reflect the characteristics that affect the recall decision into the recall model, and generalize it through learning. On the other hand, it can improve the intelligence of recall adaptive decision-making, transform more policy control into model-driven, and improve scalability and optimization efficiency.

Huang Bo, Technical Leader of Zhihu AI team: the application of AI in Zhihu production, consumption, connection and governance

As a well-known knowledge sharing platform in China, Zhihu has 160 million registered users and more than 100 million responses. at present, AI has fully participated in all aspects of Zhihu, which has greatly improved its efficiency. AI plays an indispensable role in the four fields of production, consumption, connection and governance. Huang Bo, a teacher in this special session, will elaborate on these areas and how AI is applied.

Zhihuyu uses AI to build intelligent communities in knowledge graph, content analysis, user analysis and business applications.

Teacher Huang Bo said that at the application level of the knowledge graph, it is mainly divided into three parts, namely, voice search, interest graph in the recommendation system, and the construction of the knowledge graph. Zhihu is mainly constructed around the entity, content, type, domain and relationship elements of the knowledge graph, and various relationships can be constructed among the various elements.

Content analysis refers to the labeling of all kinds of content. Why do you want to do such a multi-granularity semantic tag? Teacher Huang said that in the first-and second-level field, we hope that it has a coarse granularity and a complete and orthogonal classification system as far as possible, so as to ensure that any question / article can be classified into a certain category. In the topic area, we want the model to have high accuracy, and there can be multiple topics on the same question / article. In entities / keywords, we require high accuracy of the model and give priority to ensuring that hot entities / keywords are recalled.

At the level of user analysis, it includes user basic profile, user interest profile and user social representation and mining. In its service architecture, real-time computing includes real-time interest calculation, real-time behavior calculation, and final login behavior location; offline computing includes user basic profile prediction, user representation and user clustering, and the service architecture also includes HBase multi-cluster synchronization and online services.

Knowledge graph, content analysis and user analysis will all be based on the final application goal, the most important goal is the information flow recommendation on the front page, the whole content analysis and user analysis will be a very important feature at the bottom of the whole recommendation system, used for recall and sorting, recall in addition to this recall based on tags, there will also be recall methods based on collaborative algorithms and neural networks.

Li Yin-MACE, Xiaomi Software engineer: the Design practice of Xiaomi Mobile Deep Learning Framework

With the in-depth development of mobile Internet and the popularity of IoT smart devices, users' demand for intelligence, low latency and privacy protection is becoming higher and higher, and offline deep learning applications on mobile devices are becoming more and more common. MACE is a deep learning model prediction framework optimized for mobile devices. MACE can be easily deployed to heterogeneous devices, make the model run on CPU, GPU and Hexagon DSP, and support the mainstream deep learning networks such as visual DNN (classification, object detection, semantic segmentation, style transformation, etc.) and NLP (machine translation, etc.), and apply scene portrait patterns, scene recognition, offline translation and other projects in Xiaomi. In this lecture, Mr. Li Yin will focus on the framework design of MACE, operator optimization on heterogeneous devices, and the use and deployment of solutions.

Teacher Li Yin introduced that MACE focuses on in-depth learning, how to predict on the mobile side. It takes two steps to deploy the training model to the business scenario. First, we need the resources of the GPU server to train a model after we get the data, tags, and samples. After the model training, we will get a result, which includes two parts, one is the model and its own structure, and the other is the parameters of the model after training. Take these two parts to form the final cloud training model, and then we need to deploy this model to the mobile side. Next, we need the execution engine of the mobile side to execute all the operators contained in these model structures on the mobile side, and this is what MACE needs to do.

"Ten years sharpen a sword, sharpen the fragrance of plum blossoms", the 10th China system Architects Conference prepared a three-day traditional technology conference speech, two days of in-depth theme training, more exciting topics welcome to visit the conference special page (http://zt.it168.com/topic/sacc2018/)).

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.