Shulou(Shulou.com)06/03 Report--

The fish and sheep come from au Fei Temple.

Qubit report | official account QbitAI

"Visual Power Show", CVPR 2020 is being held online.

This CVPR, which claims to be the most difficult in a decade, received a total of 6656 papers and selected 1470, with an enrollment rate of only 22%, the lowest in a decade.

However, the Chinese Legion is still outstanding. Tsinghua took the lead in academia, while Shang Tang led the industry with 62 papers.

The achievements are eye-catching, but some are not strange. After all, this is not the first time for Shang Tang to "ride away the dust" at the top meeting.

In the previous ICCV 2019, Shangtang and its joint laboratory attracted attention with the results of 57 papers selected.

And this time the CVPR, in the number of selected papers, Shangtang continues to surpass Microsoft, Facebook and other technology giants.

In addition, he won three world championships, including the CVPR 2020 ActivityNet space-time action positioning track, the action classification track, and the CVPR 2020 NTIRE race.

So far, in the six years since its establishment, Shangtang has won 60 + world firsts in various important competitions.

So, the question is, why should Shangtang?

62 papers + 3 world championships to achieve technological breakthroughs in various fields

Let's take a look at the technological progress behind Shangtang's CVPR 2020 report card.

In the triathlon, ActivityNet is an important strength test competition in the field of video action recognition.

The competition is sponsored by top universities and research institutions such as Google, Facebook and UC Berkeley over the years. Related technologies are of great value in many practical applications, such as video analysis, living detection and so on.

In the AVA-Kinetics spatio-temporal action positioning competition, relying on the self-developed technical object-scene-object reasoning network (ACAR-Net) and its own deep learning supercomputing platform, Shangtang Research Institute, X-Lab and the Chinese University of Hong Kong-Shangtang joint laboratory team won the first place with an absolute advantage.

39.62mAP 's score is significantly ahead of this year's second place 6.71mAP.

In the action classification competition, Shangtang Research Institute, X-Lab and the joint laboratory team rely on their own deep learning supercomputing platform to train a number of video classification models with ultra-deep networks in a short time. In the stage of multi-model fusion, the spatio-temporal interleaving network (TIN) developed by Shangtang also plays an important role.

As a result, it tied for first place with Google Cloud AI in the competition.

As one of the most comprehensive low-level vision competitions in the world, NTIRE (New Trends in Image Restoration and Enhancement) can directly reflect the research progress and breakthroughs in the hot field of low-level vision.

In the real image (Real World RGB Images) spectral reconstruction (HS Reconsturction) track of CVPR 2020 NTIRE, the Shangtang research team successfully defended the champion by virtue of a new level 4 network structure that enlarged the receptive field while doing feature extraction in different levels of subnets.

In terms of papers, the 62 papers selected by Shangtang this year are also distributed in many fields, including adversarial generation model, 3D point cloud understanding and analysis, training acceleration and model quantification, video understanding and analysis, network structure search and so on.

For example, in the paper "INT8 training Technology for accelerating the training process of convolution Neural Network" selected by Shangtang in CVPR 2020, aiming at the problem of how to improve the training speed of deep learning, an INT8 training technology for accelerating the training process of convolution neural network is proposed. The use of 8-bit numerical training model can greatly improve the training speed, reduce the calculation loss, and the training accuracy is almost lossless.

So, what's the secret of soup?

The answer should be found in the positioning of Shangtang itself.

The "algorithm Factory" in the AI era

It is reported that at present, Shangtang has a total of about 4000 employees, of which more than 2500 are algorithm and product research and development personnel, more than half of the total number of employees.

The route of Shang Tang is to rely on such a technical force to complete the transformation from scientific research to products in the mode of "algorithm factory".

What is an "algorithm factory"?

Xu Bing, co-founder of Shangtang Technology, said in a public speech that how to mass-produce models for different objects and scenes has become a key issue driving AI growth and the evolution of the next generation of technology.

The mature deep learning and training platform can promote the generation of large-scale new models. This led to the formation of the concept of "algorithm factory".

How to embody the "algorithm factory"?

The most direct change, of course, is to reduce costs and increase efficiency.

Xu Bing mentioned that it would take 10 researchers and six months to complete a pedestrian recognition model with 1/100000000 accuracy in 15 years.

Now, with the same model, a researcher can achieve the same effect in three days and use only half the original GPU resources.

In this way, the number of trained models can far exceed the number of researchers. When a person can bring an average of 4-5, or even dozens of industrial-level models, the artificial intelligence algorithm can be regarded as a centralized service, through the framework platform through the end customers, the scope of application can be expanded more quickly.

Xu Bing said that with the algorithm factory and more front-end perception, what you can see in business is the further acceleration of the digitization process. The penetration rate of AI in each scene will increase rapidly, and the number of functions will increase rapidly. All aspects of industrial production, urban governance, work and study will be affected by subversion.

In short, the purpose of building an "algorithm factory" is simple--

Realize the large-scale landing of AI and further stimulate commercial value.

When the technology is on the ground

The energy brought by the "algorithm factory" was also reflected during the epidemic.

During the epidemic, Shangtang worked with the people's Hospital of Qingdao West Coast New area to develop a CT image screening system for COVID-19 within a week, and deployed an anti-epidemic front line to help doctors improve the accuracy and efficiency of diagnosis.

This speed is based on the SenseCare intelligent diagnosis and treatment platform built by Shangtang.

In fact, in addition to imaging departments, AI Medical can also actively play an auxiliary role in clinical departments such as cardiac surgery.

For example, a lot of data need to be studied and evaluated before cardiac stent surgery. In this process, AI can identify the core indexes such as curvature, length and diameter of blood vessels before operation, and then simulate the operation of stent placement. This is very helpful to reduce the risk of operation and improve the efficiency of operation.

Based on this background, Shangtang's SenseCare intelligent diagnosis and treatment platform provides services such as AI recognition, auxiliary diagnosis and treatment, and operation planning, covering many departments, such as Gastroenterology, Orthopaedics, Respiratory, Neurology, radiotherapy, Radiology, Stomatology, Cardiology and so on.

Moreover, in 2019, SenseCare has been certified by two SFDA and began the process of commercialization.

Another striking landing case is the smart city.

Recently, Shangtang took the lead in experimenting with AI+-net management in Qianlu Street, Changning District, Shanghai, and developed functions such as "intelligent patrol screen".

Based on Shangtang SenseFoundry Ark city-level open vision platform, it constructs a multi-scene, one-stop AI urban governance solution, and realizes the full closed-loop management of AI research and disposal, including automatic discovery, filing, intelligent dispatch, disposal, automatic verification and closing.

This kind of AI closed-loop management can effectively solve the urban pain point problems such as exposed garbage identification and shared bicycle stacking, and greatly improve the efficiency of urban management.

In addition, another trend of AI technology landing is the combination of algorithms and hardware.

For example, people are familiar with taking pictures on mobile phones. Limited by the limit of hardware, the emergence of billion-level pixel mobile phones is actually a combination of software, algorithms and multi-lens photography.

Shangtang, as a "AI factory" output algorithm, has helped mobile phone manufacturers achieve well-known functions such as 60x zoom and dim light shooting.

Today, AI has already penetrated into all aspects of life, especially since the beginning of this year, under the COVID-19 epidemic, digitization has become a new driving force for innovation and economic growth, and AI is an important underlying technology to promote this process.

Under this background, artificial intelligence has stepped into the era of landing as king.

The popularization of 5G technology and the opening of the curtain of new infrastructure are the east wind that the cutting-edge technology of AI enterprises turns into actual productive forces.

For tech startups such as Shangtang, this may be the best opportunity in history.

Just wait and see.

-end-

Https://blog.csdn.net/weixin_42137700/article/details/106900204

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.