from: Alibaba Cloud MVP 2020-01-09 5651
speaker Profile: Gao Jie, senior algorithm expert, Machine Intelligence Laboratory, Damo Academy
make an appointment to watch a Speech Video
the following content is organized according to the speech video and PPT. This article will share the AI product matrix that Alibaba Cloud Intelligent DAMO Academy can provide from a technical perspective.
In recent years, the word "artificial intelligence" is very popular. Just two or three years ago, people's understanding of AI was just "AI is an advanced technology, which can surpass the world champion in playing go." But nowadays, AI has become a technology wave, even a social trend, which also reflects the high expectations of the whole society for AI technology. We expect AI to enable robots to explore nature, AI to build a digital nervous system comparable to nature, and AI to overcome major diseases. These expectations will certainly occur in the long run. For Alibaba, the exploration of AI is more down-to-earth. Therefore, Alibaba is also concerned with scenarios where AI interacts with various production activities in the current human society and produces value, and also made a variety of attempts.
Today, Alibaba has implemented large-scale commercial application of AI. A very practical case is the past "Double 11". What we don't know is that our shopping spree during the Double 11 shopping festival has been covered by all AI technologies.
Nowadays, most users often use mobile Taobao for shopping, and there is a function in the upper right corner of the homepage of Mobile Taobao, which is called paili Tao. Behind pailitao, more than 0.4 billion items and 3 billion pictures are indexed. When you enter a specific store, you will see many billboards, especially the home page Banner. Nowadays, the Banner on the homepage of many Taobao stores is automatically generated by the AI designer-Luban system. At present, Taobao has millions of shops on the whole website, and each shop wants to make its own shop page design more attractive. The design of the shop page itself requires very professional technicians and personnel, if you want to design home pages and pictures for these stores by manpower, it needs about 10 million designers to participate, while there are only hundreds of thousands of designers in China, and most of them are distributed in large companies. Therefore, for small and medium-sized enterprises, the AI-based Luban designer system can be used. Users can generate very beautiful banners and pictures for their stores by simply calling the API, and can automatically publish and update them every day. Furthermore, when you purchase products, you may not know the information about some products. Therefore, you need to open a consultation window to communicate with customer service. Alibaba's AI technology is also behind this. Based on the knowledge base behind alimebot, it can answer about 80% of users' questions. Finally, when a user purchases a product for delivery, an AI system also exists in the logistics process. An example here is Cainiao voice assistant. During the "Double 11", couriers will be very busy because they need to deliver a lot of goods every day, and they also need to confirm whether the consignees are at home one by one when delivering goods. After Cainiao voice assistant is available, this step can be completely completed with the help of the machine. Only on the day of double 11, Cainiao voice assistant helped the courier make a total of 160,000 hours of phone calls, saving huge labor costs and generating great value for the courier company.
DAMO Academy: innovating the world with science and technology
before introducing the AI products of Alibaba Cloud DAMO Academy, let's introduce the AI lab of Alibaba DAMO Academy. Alibaba DAMO Academy was formally established on October 11, 2017, focusing on application-driven basic science and innovative technology research. Different from the research institutes of other large companies, Alibaba hopes to rely on various kinds of research within the Alibaba economy to drive technological progress and drive these technologies into real application scenarios, finally, systematic innovation can be realized.
Global layout and full-stack technology innovation
the research field of DAMO Academy is simply "4 + X", where "4" includes machine intelligence, data computing, robots and financial technology, "X" refers to X laboratory. Machine intelligence includes visual intelligence, speech intelligence, language technology, decision intelligence, urban brain, etc. Data computing includes underlying computing technology, Intelligent Computing, database and storage, etc; robots include intelligent logistics and unmanned vehicles. Financial technology includes block chain, biometrics, and financial intelligence. While some more cutting-edge technologies, such as quantum computing, are put in X laboratory. Today, Damo Academy has become a global research institution. More than 300 researchers of DAMO Academy are distributed in 7 cities in 4 countries around the world.
Introduction to Machine Intelligence Technology Laboratory (MIND)
the main contents introduced in this article are basically related to the Machine Intelligence Technology Laboratory of DAMO Academy. MIND: Machine IntelligeNce of Damo is composed of a group of outstanding scientists and engineers in Hangzhou, Beijing, Seattle, Silicon Valley, Singapore, etc, alibaba is the core team responsible for AI AI technology research and development. Machine Intelligence Technology Lab is based on Alibaba's valuable massive amounts of data, deep learning, machine learning, and machine learning technologies, core AI technologies such as natural language understanding fully empower important Alibaba Group businesses such as e-commerce, finance, logistics, social networking, and entertainment, as well as the output of ecosystem partners, implement Alibaba's "AI" (Alibaba Inside) technology strategy.
Core technical capabilities of Machine Intelligence Technology Laboratory
the main research directions of Machine Intelligence Technology Laboratory of DAMO Academy are divided into four directions, namely decision intelligence, Image Vision, voice interaction and natural language processing.
As a department focusing on algorithm and basic technology research, the Machine Intelligence Technology Laboratory of DAMO Academy has many original technologies and papers, and has won numerous awards in many competitions.
AI applications cover more than 2000 scenarios in Alibaba's economy
the AI technology of the DAMO Academy machine intelligence technology laboratory has covered various scenarios of the Alibaba economy. Specifically, it covers about 2,000 scenarios, including retail, home furnishing, customer service, driving, logistics and office scenarios.
After these AI technologies are applied on a large scale by Alibaba itself, we can only achieve the first loop shown in the following figure, that is, technologies are available and applied. Now, Alibaba hopes to provide these technologies to the public as a public product and to every developer who wants to use AI technologies, finally, it falls into the specific industry scenario to create real commercial value.
now, Alibaba Cloud has opened all the intelligent technologies of the DAMO Academy machine intelligence technology laboratory, such as intelligent speech, NLP, Knowledge Graph, face recognition, and machine translation. You only need to log on to ai.aliyun.com to access each scenario of each subdivision technology. There are about 130 products applicable to more than 300 scenarios on the official website. These are basic capabilities. In Alibaba Cloud, these capabilities are called atomic capabilities. In order to generate the final commercial value for everyone, atomic capability alone is far from enough. Therefore, it is necessary to bridge the gap between atomic capability and commercial value.
AI. Machine Vision the atomic capabilities mentioned above have been widely used in practical fields. Take Machine vision as an example, such as face recognition, image tagging, the scenario classification capability has been applied to various scenarios with a large number of calls. However, there is a big gap in how to make these basic abilities generate value in the final industry practice. To fill the gap between capabilities and business value, there are two main ways. One is that Alibaba, as a provider of AI basic capabilities, actively takes a step forward, find scenarios that can be implemented for AI capabilities, and then combine AI capabilities with scenarios. This has two advantages. The first is that the value of AI algorithms and technologies ultimately depends on the final industrial practice of commercial implementation. The second is that the progress of AI technology itself is driven by industrial practice. The future development direction of algorithms depends on requirements. In addition, most AI technologies are data-intensive, while data sources are still in the industry, only when AI technology is widely applied in the industry can real data be obtained to iterate the AI model, making it more accurate and the model itself better.
Here is a more specific example-OCR technology. The Chinese meaning of OCR is optical entity recognition. In short, it is to read the text from an image. OCR is a very simple technology and also a very old technology, which has existed in the industry for more than ten or even twenty years. As a basic atomic capability, OCR technology can accurately identify every word in the whole picture, however, whether the recognized words can generate value and how to recognize words more accurately and better are the biggest problems currently faced by OCR technology.
Alibaba has deeply integrated OCR technology with the industry, such as identifying various receipts and invoices in the medical industry. This has two advantages. The first is to make full use of the prior information of these industries, so that the AI algorithm can be optimized for such prior information and the accuracy is higher. The second is to transcribe words into structured information, which is the basis for the subsequent use of data mining to generate commercial value. Therefore, both higher-precision recognition rate and structured information solutions focus on reading optical literacy OCR solutions. In addition, Alibaba provides vertical solutions for a series of scenarios such as finance, medical care, justice, and customs.
Intelligent speech interaction the example here is intelligent speech interaction. In China, Alibaba's voice synthesis and other technologies have the best standards or effects. The General voice broadcast providers will provide a standard male voice and a standard female voice, while Alibaba's speech synthesis idea has taken another step towards the industrial scenario. Alibaba has optimized and customized different scenarios, such as children's story reading and customer service. However, it is not easy to do this. Only with good technical support can the overall cost and time be reduced. In addition to speech synthesis, it also has speech recognition and speech interaction capabilities, including mobile phone AMAP speech assistant, Tmall Genie, Haier, Konka and other speech assistants. Currently, speech recognition technology has been widely used. Only on the public cloud, there are thousands of speech recognition users, one of the typical specific scenarios is court hearing speech recognition. Now, more than 10,000 courts across China have installed Alibaba's intelligent court hearing system.
Some problems may also be encountered when such intelligent speech interaction technology is applied to a wider range of scenarios. The core problem in the application of intelligent speech technology to various scenarios is accuracy. Although our computing power is very powerful and the data volume is very large, it may be several orders of magnitude better than that of ten years ago, and people also constantly claim how high their speech recognition rate is. However, in fact, "The reality is very bony". When Alibaba applied its speech recognition technology to thousands of partners, and when speech recognition was deployed to tens of thousands of courts across the country, you will find that the current speech recognition technology is far less robust. To give two specific examples, for a call center of an electric power operator, if the speech recognition technology is not optimized for the electric power industry, the accuracy rate may rapidly decrease from 95% to 81%. As for legal trials, in southern cities such as Fujian, the recognition rate will be greatly reduced due to accent and other reasons. For the preceding scenarios, there is only one solution, which is to optimize these scenarios.
Intelligent speech interaction-self-learning platform so, should Alibaba solve all the above problems? This leads to the second point that this article wants to express, that is, Alibaba hopes to provide some basic capabilities to work with partners to achieve the best application results in the entire industry. Therefore, Alibaba provides a complete set of self-learning platform for intelligent speech interaction. The idea is that Alibaba focuses on providing atomic capabilities, focusing on what Alibaba is good at, and creating better algorithms to improve the model capabilities. If the recognition rate is reduced in some scenarios, Alibaba provides a complete set of process tools for partners to perform targeted optimization. Some specific cases are also listed here. In the call center, partners can use the self-learning platform to quickly optimize acoustics or language, thus increasing the recognition rate from 85% to 91%. However, some intelligent voice service providers in the market believe that they can cover the world. However, in tens of thousands of scenarios, one capability is difficult to do well in each scenario. Alibaba has chosen to give the key to these capabilities to its partners so that they can perform unlimited optimization based on their own scenarios and the cost is permitted, this enables partners to achieve higher accuracy than Alibaba Cloud's end-to-end solutions when using the basic atomic energy and self-learning platform to finally create solutions. In addition, combined with partners' insight into the industry, end users can enjoy the best experience and help partners create more value, which is also reflected in Alibaba's customer service partner Alliance.
NLP basic services: make machines understand you better in addition to basic word segmentation and other technologies, Alibaba also provides search and recommendation, public opinion, advertising, Justice, customer service, and other services. As we all know, natural language processing is more sensitive to industries and fields. Therefore, Alibaba's NLP basic service has also launched a self-learning platform that is the same as intelligent speech interaction, capable of targeted entity recognition, key tasks such as classification enable partners to obtain industry data without background knowledge, thus achieving the best effect of industry application.
AI has never left the computing Foundation. Many AI algorithm-related scenarios were described earlier. In fact, the development of AI technology still depends on the progress of data and computing. Behind the AI technology of the DAMO Academy machine intelligence technology laboratory is Alibaba Cloud's solid computing foundation. As a global leader in cloud computing, Alibaba Cloud provides the best cloud computing technology, A single cluster can support more than 100,000 machines and provide good support for CPU and GPU computing.
Alibaba Cloud machine learning platform PAI today, in addition to the various algorithm products mentioned above, the PAI platform, the cornerstone of Alibaba Cloud Machine Learning, is also open to everyone as an official product of Alibaba Cloud. In addition to AI technology capabilities, developers can also use the Alibaba Cloud PAI platform if they need the final application deployment. On the one hand, developers can use the PAI platform for final prediction services, and on the other hand, they can also use it to implement self-learning capabilities.
Three core boosters of AI at the beginning of this article, we introduced the capabilities of the DAMO Academy AI platform. In addition, we also introduced the AI computing capabilities of Alibaba Cloud, in the end, when these basic atomic capabilities are put into the industry, the real value still depends on AI Inside peers. Alibaba hopes to narrow its scope of work to what it is good at, such as better algorithms, better AI capabilities, and better computing capabilities, how to generate better value and how to do better, all partners have more advantages. Therefore, Alibaba hopes to combine the capabilities of itself and its partners to generate practical business value for end users. Alibaba hopes to enable every enterprise to have AI capabilities and make all kinds of expectations for the future come true little by little through the joint efforts of Alibaba and its partners.