On September 6th, Tencent senior executive vice president and one of Tencent's seven business groups, SNG (Social Network Group), the head of the social network business group, Tang Daosheng appeared in Shanghai, and held the Tencent U-Computer vision for the day. Summit platform.
“So far, Uto’s technology accumulation and application have achieved some results, but this is not enough. We will continue to invest for a long time without KPIs. & rdquo; Tang Daosheng stressed in his speech.
Compared with the well-known fist products such as QQ, Tiantian P, National K song and Tencent Cloud in the SNG business group, the existence of U-Map Lab is still not well known. This time, Tang Daosheng's development of “Excellent Development” does not include KPI", which shows Tencent's bet on the further expansion of the AI map represented by U-Map Lab.
Looking back on the past of Tencent's layout of artificial intelligence, “Yutu” can be called a milestone.
In 2012, Tencent U-Lab was established, focusing on computer vision, focusing on image development, pattern recognition, machine learning, data mining and other fields to carry out technical research and development and business. At that time, among the four domestic AI unicorn companies that are now called “Visual Recognition Four Little Dragons”, Shang Tang and Yun have not yet been established, and contempt and Yitu are only in the nascent stage of development, with visual recognition as The artificial intelligence technology represented has not yet reached a large-scale scene.
The accumulation of basic technology has become an important move in the past six years. In order to prove its scientific research strength in the AI field, Utographs began to appear frequently in the top events of the artificial intelligence industry (MegaFace, LFW, ICDAR, MIREX), conferences (ICCV, CVPR, AAAI) —— only in 2017, Uto has published 18 A-category papers and launched the first AI open source project, ncnn.
At the same time as the technical reserve, the AI capability will be used to form a product with a lower threshold, and find the right scene to land, which has become the focus of U-Tou's development. On the other hand, AlphaGo defeated Li Shishi and the Boston Power Robot's stunning debut, and also promoted Tencent's AI “ ran out of the laboratory, turning scientific research into a tool to improve the efficiency of the physical industry.
“In the past two years, we have often heard companies say & rsquo; our abilities have reached 96.88%, and have increased by one percentage point & rsquo;. But now, this kind of sound has begun to slowly become less, artificial intelligence has entered a period of deep development of the scene. & rdquo; Tang Daosheng said.
Tang Daosheng, Senior Executive Vice President of Tencent
In order to realize the scene of Tencent AI faster. On the same day, Tang Daosheng announced that he would upgrade Tencent Uto Labs to Tencent Computer Vision R&D Center to penetrate into the specific applications of medical, autopilot, industrial, retail, office, culture, and social welfare.
From the basic ability to "run high scores", to find the vertical field technology landing, artificial intelligence has entered the "run scene" era. However, for Tencent, which focuses on its own “two and a half” (social, content + half finance), it is necessary to face the complexity of the unfamiliar field in order to integrate its own AI basic technology into the complex scenes under the line. Accumulation, while at the same time need to grasp the distribution of interests with partners.
In BAT's business division of artificial intelligence, compared to Alibaba's Dharma Institute and Baidu's AIG business group, Tencent chose a “laboratory” system that is more deeply tied to the business. Specifically, it is to establish a laboratory with business characteristics in each business group of Tencent, and it is deeply linked with the frontline functional departments.
Also in the SNG business group, in addition to the upgraded to the Tencent Computer Vision R&D Center's Uto Lab, there is the Tencent Audio and Video Lab established in 2016 and the Quantum Computing Lab established in 2017. Among the three business groups, the quantum laboratory has the meaning of long-term layout, and the birth of audio and video and excellent pictures are closely related to SNG's existing QQ, QQ space, NOW live broadcast, Tencent cloud and other business needs.
Taking audio and video labs as an example, Liu Xiaoyu, the person in charge, said in an interview that QQ had audio and video communication functions in 1999, but the network problem has always been a technical difficulty, whether it is detecting network bandwidth or processing Packet, jitter, and multi-terminal device adaptation require a deep technical accumulation, which in turn forces Tencent to produce self-developed technology.
However, in the AI layout lineup that Tencent is currently publishing, in addition to the Tencent U-Lab, the AI Lab and WeChat AI Labs were established in 2015-2016.
Among them, AI Lab is affiliated to Tencent TEG (Technical Engineering Group), and is led by Dr. Zhang Wei, the top scientist in the field of artificial intelligence. His basic research interests include computer vision, speech recognition, natural language processing and machine learning. AI“ 绝 & & , 、 、 、 、 、 、 、 、 、 、 、 、 、 、 、 、 、 、 、 、 、 、 、 、 、 、 、 、 、 、 、 、
Tencent AI Lab leader Zhang Wei
The WeChat AI Lab is part of Tencent WXG (WeChat Business Group), which focuses on natural language processing, image and video, data mining and document comprehension, and input and text in speech, sweeping/covering, translating, and chatting. The robot, shake the music TV, voice lock and other scenes for application.
Because Tencent has a “horse race” tradition, how to determine the priority of the AI lab, and whether it can enable multiple laboratories to implement technical and business linkages, becomes a difficult point for Tencent to consider when exporting AI capabilities. one.
“It’s not that any lab has priority or no development, and every lab has its strengths. "The general manager of Tencent Youtu Lab, the distinguished scientist Jia Jiaya told the titanium media.
In Jia Jiaya's view, there may be no more than ten candidates who understand quantum computing and are willing to join the quantum laboratory, but there may be tens of thousands of experts in artificial intelligence. The difference in the size of this talent will inevitably affect the experiment. The magnitude of the input between the rooms.
For Uto Labs, Jia Jiaya bluntly said: "The development of the entire map is based on the landing, the latest research technology quickly landed in the industry. & rdquo; He used the image of the human body tracking algorithm of the cooperation between Tentu and Tencent's short video products, “micro-vision”, from the receipt of the demand to the completion of the project, the excellent figure took less than half a year to realize the human body shape, including the human face. , real-time tracking technology for each joint.
“Why is the figure able to rise? It is because they can fight hard. There is a brand new industry to negotiate, even if you don't understand anything before, you can start this thing from scratch. & rdquo; Jia Jiaya told the titanium media.
Let AI “Run” into the scene
2017 is a veritable strategic planning year for Tencent AI. In November of this year, Tencent announced the first full deployment of artificial intelligence at the Global Partner Conference in Chengdu. Ren Yuxuan, chief operating officer of Tencent Group, said: Tencent's goal is not All in AI, but AI in All, which allows Tencent's AI technology to be integrated with the industry after its opening, so that the technical value of AI can be realized.
In March of this year, Tencent “ two sessions of media communication meeting & rdquo; scene,Ma Huateng also answered questions about the layout of Tencent AI in Titanium Media.He said:
Now AI is applied to some special scenes, and the scene is drawn narrow enough, clear enough, and then through deep learning to understand its features, data, and developed into an algorithm. I think starting with the vertical field, AI's GM may be far away. I think there are many opportunities for vertical.
“觅影” is the first benchmark for Tencent AI's large-scale industry. This AI medical imaging product, which was born in August 2017, can assist doctors in screening for esophageal cancer, pulmonary nodules, diabetic retinopathy, colorectal cancer, breast cancer, cervical cancer, etc. Among them, the shadow of early esophageal cancer The screening accuracy rate is as high as 90%, and it has already landed in more than 100 top three hospitals nationwide.
& ldquo;Tencent &影” Colorectal tumor screening AI system finds and identifies polyps in real time.
In addition to the breakthrough in technology, the development route of Yingying represents Tencent's overall layout direction for AI.
“Open, cooperation” is the key action of Tencent. Especially for professional vertical scenes such as medical treatment, Tencent naturally cannot rely on one's own efforts to complete the penetration.
Therefore, the Tencent medical team responsible for the film products constantly expands tools in WeChat smart hospitals, sugar doctors, medical clouds, etc., and also focuses on the whole industry chain cooperation, such as the AI accelerator open AI technology, investment, combined with Tencent MIG business group. Tutors and other resources, landing artificial intelligence medical laboratories in many top three hospitals in China.
“We hope that the partners will treat us as a supermarket and can choose the capabilities they need. We cannot force you to choose things you don’t like. & rdquo; Tencent Vice President Chen Guangyu once talked about Tencent's medical layout.
On the other hand, behind the Tencent MIG (Mobile Internet Business Group), behind the film, the technology linkage of Ten Vision, AI Lab and other AI laboratories is integrated.
“Everyone has a specialization in each other's work, and they can complement each other very well. We rarely have a crash in the same field. For example, the breakthrough of AI in the medical field is a very good case of internal coordination. ” Tencent Vice President Liang Zhu told Titan Media.
Retail is another concentrated area of Tencent's recent AI landing. In May of this year, the face payment system of Tencent Youtu and WeChat payment cooperation was put into use in Shanghai Carrefour. This scene application incorporates the excellent image recognition and 1:1 nucleus technology to determine the nuances of the face, and the error rate of one billionth can be achieved under 1:1 conditions.
In May of this year, the face payment system of Tencent Youtu and WeChat payment cooperation was put into use in Shanghai Carrefour.
Through cooperation with Tencent Cloud, Tencent also launched “Excellent Mall”, a smart retail system, where consumers can use face registration members, through the face search and recognition technology of Tencent Youtu, the store can be at the moment the consumer arrives at the store. Identify the identity of the customer. If you identify old customers and VIP customers, you can also push items that may be of interest based on their past purchase records; in the end, consumers can easily complete the payment action by “brushing face”.
It can be seen that whether it is To B's medical and retail scenes, or To C's short video and payment tools, Tencent AI is accelerating the technology “running” out of the lab in the past year. Even the general manager of Tencent U-Lab and the outstanding scientist Jia Jiaya mentioned in the performance demonstration last week that this is the first time that the U-Tech team has been established for the first time in six years. Combined cases of industries such as culture.
However, Jia Jiaya also frankly said that Uto's long-term accumulation of technology in the laboratory is directly related to the efficiency of the scene of the whereabouts.
In Jia Jiaya's view, unlike other visual identity AI companies on the market, Uto's barriers have a powerful computing platform. The process of setting up this middle platform is very difficult, but after completion, the time for Unitu's researchers to train a new model can be shortened from ten days to one hour, plus a rich application layer such as Tencent Cloud and Microvision. Data and product requirements can further revive the capabilities of the middle platform through these feedbacks.
“From the training of the algorithm model to the completion of the combination of the establishment of the middle platform and the application of the upper layer. After completing this cycle, only three or four people who invest in a core algorithm are available, but forty companies in the small company are not finished. , & rdquo; Jia Jiaya said to the titanium media.