Three experts appeared at the machine learning venue of the WOT AI Summit, revealing how AI empowers industry applications

[51CTO.com original article] In 2018, artificial intelligence, as one of the most dazzling technologies for mankind to create the future, is profoundly affecting the global industrial structure, business model, urban form, and human life and work methods. How to enable artificial intelligence technology to empower the industry and let many industries share scientific and technological productivity is an important topic under the wave of digital transformation. As an important annual event to promote pragmatic innovation in artificial intelligence, the WOT2018 Global Artificial Intelligence Technology Summit was grandly held at the Beijing Yuecai JW Marriott Hotel from November 30th to December 1st. 60+ domestic and foreign artificial intelligence elites and more than a thousand industry professionals gathered on site to share artificial intelligence platform tools, algorithm models, voice vision and other technical content, and discuss how artificial intelligence can give new vitality to the industry.

On the afternoon of November 30, the Machine Learning Sub-Forum was held at Venue A. Three senior experts were invited to attend and give wonderful speeches. After the meeting, 51CTO compiled the experts' speeches into text, hoping that the essence of their speeches will be helpful to everyone.

Yang Xuefeng, Senior Researcher, Shenzhen Zhuiyi Technology Co., Ltd.

Exploration of reading comprehension technology and its application in enterprise services

Yang Xuefeng's sharing mainly focused on two parts. First, he introduced the current status of machine reading comprehension, cutting-edge technologies in the market, and the background, significance, and methodology of machine reading comprehension. Secondly, he shared the difficulties in the implementation of AI technology and the productization practice of reading comprehension technology.

In Yang Xuefeng's view, machine reading comprehension is to find the answer that the user needs for a given question. The answer may be text, or it may be a picture, number, symbol, or fragment. Different answer forms have different difficulties and are usually processed through different data sets. He introduced the four mainstream English data sets at the scene, namely the CNN & Daily Mail data set invented by Google, the Microsoft MCTest data set, the SQuAD1&2 data set created by Stanford, and the Microsoft MS MARCO 10,000-level data set. In addition, the Chinese data sets that are relatively good include Baidu's DuReader and iFlytek's CMRC data set.

Yang Xuefeng also gave several application scenarios of machine reading comprehension on the spot:

The first is the customer service and new employee training scenario. Due to the high mobility of customer service, companies need to minimize the training process, so that customer service personnel can take up their posts as soon as possible and become familiar with the operation manual and common business documents. Through machine reading and comprehension products, employees' questions can be answered immediately and they can be helped to become familiar with the business.

Second, it provides professional consulting services in the financial field. Due to limited manpower, fund managers cannot answer everyone's questions, but through machine reading comprehension, users can query detailed documents and data independently, improving user experience;

The third is children's interest-based early education, which can provide children with a general knowledge question-and-answer system for educational applications, such as children's early education machines, infant tutoring, etc., so that children can enrich their knowledge while having fun.

Yang Xuefeng also used the solution they provided for China Southern Airlines to illustrate his experience. He said that China Southern Airlines launches various activities every week, and customer service staff need to answer a large number of questions at this time. Zhuiyi Technology will select the most easily consulted documents from thousands of documents, build a model through automatic extraction functions, recommend many question and answer fields, simplify the enterprise annotation process, and make detailed annotations to make it as simple as possible for users to use and meet high-concurrency access scenarios. "There are still many challenges in this field in the future, especially since most financial customers are privately deployed, and data is a closed-loop system. They hope to be able to superimpose their own data to train new models, and let the service provider only provide product logic."

Yu YuanyuanCTO of Hangzhou Weipei Network Technology Co., Ltd.

Application of Deep Learning in E-sports Industry

Yu Yuanyuan said at the beginning that the data analysis methods of the traditional sports industry are not applicable to the field of e-sports. The traditional sports industry collects data by both manual and technical means. For example, if a player scores a goal, it needs to be manually registered according to the on-site situation, and more on-site professionals are required to interpret it. However, these data analysis methods do not work in the field of e-sports. For example, e-sports players can often play dozens of games a day. The game iterations are fast, and the game rules and logic are changeable. It is difficult for traditional data analysis to keep up with the pace of change. In addition, the characters in the game are complex, and multiple indicators such as equipment, team contribution value, and damage of each character need to be analyzed and displayed. What is more special is that the subjective consciousness of participants in the e-sports industry is very strong, and the cognitive differences are large, so it is difficult to apply the analysis model of traditional sports experts. "In the past e-sports data analysis, many teams used traditional mathematical models, which were often accused by players and users of being not objective or comprehensive enough. If deep learning is used, these problems can be effectively avoided."

Compared with traditional mathematical models, deep learning has three major advantages: plasticity, universality, and efficiency. Yu Yuanyuan pointed out that the model of deep learning can be many, many building blocks, which can turn many problems into engineering problems. After training, the model can be built quickly, and the model can be easily moved to another project. Of course, Yu Yuanyuan also admitted that deep learning in the e-sports industry also has shortcomings. It requires a lot of training data, and it cannot be learned and understood directly. "When deep learning is combined with e-sports games, processing large amounts of data does not require interference from human factors, it is more objective, and large amounts of data can be processed quickly and completely."

At the end of the speech, Yu Yuanyuan introduced a win rate analysis project AlphaMao in detail to show how they use deep learning technology and models to solve e-sports problems. In half a year, the model has been updating its model every day by learning new game samples, and the training data comes from 60 million e-sports games. Because the rules of the game change very quickly, they have to learn from past data and update the latest data, and adjust and increase the weight ratio of the updated data at any time. The final model has an overall accuracy of 60% and a relative accuracy of up to 97.6%.

"There are three main directions we will focus on in the future." Yu Yuanyuan revealed that they will add time data to the model to strive for real-time win rate analysis, and will also analyze the win rate to provide players with more intuitive and effective suggestions. In addition, they hope that the deep learning model can be easily transplanted to other games.

Jia Rongfei Taobao Senior Algorithm Expert

Redefining people, goods and places - intelligent terminals and situational computing

Jia Rongfei mentioned in his speech that when consumers buy things, they have different needs in different scenarios. For example, in the office, due to time constraints, people want to place orders faster, but if they are at home, they prefer to lie comfortably in bed and browse the shopping pages slowly. In view of this, Taobao hopes to redefine people, goods and places, increase its understanding of information on users' mobile phones, change the interaction mode between e-commerce and consumers, and thus promote the intelligent development of e-commerce.

As more and more goods are sold through e-commerce, it becomes more important to accurately understand consumer needs. Jia Rongfei believes that the era of smartphones makes all this possible. It can locate the real-time status of users, and then Taobao uses scenario computing technology to analyze and understand the real-time status of users, and more focused on understanding what customers need at this moment. "This is a better understanding of users than the original recommendation search." He admitted that there are also many difficulties here. First, the amount of user information is very large. How to extract effective information from complex information? Second, as the amount of data processed increases, how can the computing power of e-commerce support a larger amount of information and more complex models? This is a very challenging topic.

According to Jia Rongfei, Taobao's demand for scenario computing is mainly to process user orders in real time and to have greater computing power to serve individual users. Under such demands, Taobao chose to deploy computing power mainly on the client without hesitation, and to build global models on the cloud to achieve information sharing between users. The overall scenario computing architecture design is "terminal-based, cloud-assisted." The deep learning model is used to identify and understand user information, which is very helpful for judging the user's environment.

Since these models are complex models, and complex models will inevitably encounter computing power issues on the end, Taobao began to study how to support computability through improvements to the end framework. Jia Rongfei introduced that first, there will be raw data collection on the end, and the real-time status of users will be obtained regularly, and sent to the basic model to produce intermediate data to serve the application products. At the same time, in the cloud, Taobao will complete the model training, and then regularly synchronize it to the user's mobile phone, using experiments and various parameters to ensure that users have a good user experience.

"Next, we hope to make more attempts in three directions," Jia Rongfei revealed. The first is to build a user demand discovery engine that can analyze user preferences in different scenarios. The second is to achieve a change from one thousand faces to one thousand models for one thousand people, giving full play to the computing power of the user's mobile phone and the in-depth understanding of the user, and training a more personalized thing for a single user on each mobile phone. The last one is the joint learning of cloud + terminal. Taobao hopes to combine the information of cloud and terminal for training to achieve better results.

The above content is compiled by 51CTO reporters based on the speech content of the "Machine Learning" sub-forum of the WOT2018 Global Artificial Intelligence Technology Summit. For more information about WOT, please pay attention to .com.

[51CTO original article, please indicate the original author and source as 51CTO.com when reprinting on partner sites]