Official account |-com

Source | Dige.com

Author｜Lou Yuge

Robot butler, flying cars... These are scenes that have appeared in science fiction movies such as "Blade Runner" and "Mechanical Girl". People had imagined artificial intelligence centuries ago.

Today, people’s imagination of artificial intelligence in the movie may not have been fully realized, but in fact, AI has begun to enter our daily lives, moistening things silently.

Although it is not easy for the general public to perceive it, it is undeniable that AI is playing an important role in many scenarios such as finance, manufacturing, and fire protection. Even in daily life, AI has become closely related to us for AR navigation, smart cities, etc.

However, the development of AI in China has entered the industrial stage from laboratories, so the more urgent question at present is how to implement it and achieve large-scale commercialization.

Against this background, at the recent World Artificial Intelligence Conference (WAIC), SenseTime proposed a solution - a new type of artificial intelligence infrastructure, called SenseTime AI large device, integrating computing power, platform, and algorithm capabilities, focusing on resolving the "long-tail application" problem in the implementation of AI applications.

What role will this device have in the implementation of AI applications?

Artificial intelligence heads toward the era of big models

Although it is no longer difficult to integrate AI into various industries, different training of AI models is required when dealing with different scenarios. AI development has become more fragmented. This also leads to the burden on AI developers.

Therefore, the importance of a general model that can integrate more capabilities and stronger computing power is obvious.

Since 2018, large-scale pre-trained language models such as Google BERT and GPT-3 have gradually become the mainstream in the NLP industry and are developing rapidly.

In June 2020, GPT-3, a natural language deep learning super-large model with 175 billion parameters, was released. It is the most breakthrough research result in the field of artificial intelligence since then. It can perform original analogies, generate recipes, and even complete basic code writing.

Just half a year later, in January 2021, this record was broken. Google researchers have developed a new language model that contains more than 1.6 trillion parameters.

In April this year, Dr. Tian Qi, chief scientist of Huawei Cloud's artificial intelligence field and IEEE, said in his speech that artificial intelligence will show two trends in the future. One is the trend of neural networks from small models to large models. In the past decade, the demand for computing power of AI algorithms has increased by 400,000 times. The second is the deep integration of artificial intelligence and scientific computing. In many fields, including industry, meteorology, energy, biology, medicine, etc., the profound impact of artificial intelligence on them has been seen.

He believes that big models like GPT-3 are a way to solve the fragmentation of AI applications. Large models can absorb massive amounts of knowledge, improve the generalization ability of the model from the inside, and reduce dependence on domain data annotation.

"80% of the current applications are low-frequency and long-tail needs. If these needs are not solved by unified artificial intelligence methods, they will face a lot of manpower and collect huge amounts of data, and cannot solve the small data and small sample problems that we can cause." Xu Li, co-founder and CEO of SenseTime, said this at the WAIC that ended a few days ago.

The real world is undergoing a digital process, and the boundaries between the real world and the virtual world are becoming blurred. It is particularly important to promote the connection between the real world and the digital world.

The establishment of a digital world usually requires several steps in the dataization of scenarios, the structure of elements, and the interactive process. However, at present, 80% of structured applications are low-frequency, long-tail scenarios. Without general artificial intelligence, we are facing a large amount of manpower invested in a single project, and it is not possible to solve these problems that are essentially small data and small samples.

Sang Tang AR hanging window

However, if you still focus on a single problem process in technology, it is difficult to have good generalization capabilities in many general scenarios, resulting in unstable performance. Faced with this problem, “the general hyperscale model coupled with the segmentation optimization of a single scenario in small samples becomes the core of solving production costs and accuracy.”

At the same time, with the development of general artificial intelligence technology, AI algorithms no longer rely entirely on labeled big data, but the exploration of huge unknown possibilities must rely on large computing power.

In this case, industrial applications can gradually verify the boundaries of "machine conjecture", help us more cautiously promote the implementation of new technologies and gradually try to find the boundaries of reasonable application of new technologies.

Big models are accelerating our entry into the AI ​​era.

Cities need AI "heart"

Around 2012, artificial intelligence was in full swing in China, with many star companies such as SenseTime, Yitu, Yuncong, Megvii, and Cambrian, which also attracted the participation of many venture capital institutions.

Capital has driven the rapid development of the industry, but the commercialization path of AI companies is not clear in the early stages. Megvii co-founder Tang Wenbin once said bluntly that he was very anxious two years before starting his business and couldn't figure out what he was going to do.

Huawei's rotating chairman Hu Houkun said that the current bottleneck in popularizing AI technology is no longer in technology or demand, but in development efficiency. AI application development is too slow, which seriously hinders the combination of technology and demand.

At present, smart cities have a wide range of AI application scenarios, including urban governance, transportation, parks, scenic spots, etc.

But it is undeniable that there are still challenges in the long-tail application of artificial intelligence.

Most of the current AI development models are still relatively traditional handicraft workshop models, which require high reliance on experts and data. They not only spend a lot of time collecting and processing data, but also consume a lot of manpower to optimize parameters, making it difficult to meet productization requirements in a short period of time.

Therefore, from the perspective of the entire industry, how to promote closer relationship between AI and the industry, let AI get rid of its intensive dependence on artificial intensiveness, and improve the output efficiency ratio of AI is an urgent problem that the industry needs to solve.

There have been many attempts regarding AI commercialization, and in recent years, big models have begun to become an emerging development direction for improving AI business capabilities. The main reason is that after big models become a public AI infrastructure, tens of thousands of industry models will be based on this and can quickly meet the needs of AI applications.

SenseTime, an artificial intelligence platform company, has also launched SenseTime's AI device. The so-called SenseTime AI device refers to integrating a strong computing power foundation and leading algorithm capabilities, being able to disassemble and collide data, deeply tap potential value, and realize innovation and application of AI costs, efficiency improvement and scale.

SenseTime AI device

According to reports, SenseTime's large AI device includes three layers: computing power layer (AI chip and processing card AIDC AI sensor); platform layer (model production training platform data platform) and algorithm layer (algorithm toolbox open source framework).

However, with the rapid development of AI algorithms, more and more model training requires huge amounts of computing power to be implemented quickly and effectively. Some industry insiders even said that computing power is the decisive factor in the future breakthroughs in artificial intelligence applications. Against this background, SenseTime invested 5.6 billion yuan to build an artificial intelligence computing center AIDC.

At the end of this year, SenseTime AI Intelligent Computing Center (AIDC) will be put into trial operation. It is said that after all the projects are completed, the peak calculation speed can reach 3740 (1 equals 10 trillion floating-point operations per second), which can meet the use of four super cities with a population of 20 million. AIDC will become an important part of SenseTime's large AI device, promoting the implementation of larger-scale parameter models.

The major trend in the AI ​​industry has emerged. To accelerate the implementation of AI, computing power must become an accessible and affordable urban public resource like water and electricity. Without sufficient AI computing power, just like without water and electricity, it will greatly restrict the process of urban digitalization.

AI is moving from technological competition to building industry infrastructure.

AI nourishes everything, ethical governance helps AI development

It is not difficult to find that the domestic artificial intelligence industry is ushering in a period of change. On the one hand, it needs to solve how to achieve commercialization on a large scale; on the other hand, the development of new technologies has also brought about a series of problems.

At present, the commercialization scenarios of AI are mainly concentrated in smart cities, mobile phones, automobiles and other businesses, as well as the implementation of industries such as medical, education, satellite remote sensing and even games.

But as a technological revolution, artificial intelligence has entered a wider consumer end from the B-end to the inevitable trend. Moreover, according to the goals of the "New Generation Artificial Intelligence Development Plan" issued by the State Council, by 2020, the artificial intelligence industry will become a new important economic growth point; by 2025, the basic theory of artificial intelligence will achieve major breakthroughs, some technologies and applications will reach the world's leading level, artificial intelligence will become the main driving force for my country's industrial upgrading and economic transformation, and the construction of an intelligent society will make positive progress.

This means that artificial intelligence must go out of the "boudoir", be closer to the industry and the consumer side.

At this year's WAIC World Artificial Intelligence Conference, many AI companies also showed off their blooming scenes in various fields. For example, in a very common elevator scenario in life, AI companies can achieve full-process intelligent management from pre-prevention, in-process intervention to post-tracement traceability through joint efforts, making active prevention of safety accidents possible. Earlier, Jiangsu Road Street, Changning District, Shanghai, built a multi-scene, one-stop AI urban governance solution based on Shangtang's Ark City Open Platform, effectively solving urban pain points such as exposed garbage identification, random stacking of shared bicycles, random drying, and road water accumulation.

In urban cultural and tourism scenarios such as museums, exhibition halls, scenic spots, shopping malls, airports, and railway stations, SenseTime uses the original Mars mixed reality platform to deploy 3D modeling and AR application.

It is worth mentioning that at this World Artificial Intelligence Conference, Tang Jueying, an independent brand of SenseTime intelligent automobile solutions, also debuted.

According to reports, SenseTime is a full-stack system that relies on original AI technology and sustainable output, covering from intelligent driving, intelligent cockpit to vehicle-road collaboration. Before Jueying's official debut, SenseTime had cooperated with more than 30 domestic and foreign OEM factories, and the mass production project covered a total of more than 20 million vehicles.

Based on SenseTime's AI large installation, whether it is the "Pearl of Industry" automobile or the stations and exhibition halls in daily life, SenseTime truly penetrates AI into all walks of life.

Empowering all industries seems to be the mission that AI has born with. The steam era, the electricity era and the information era in history are all named after breakthroughs in science and technology, and these technologies have three common characteristics: versatility, practicality, and durability, and can be deeply integrated with all walks of life.

But on the other hand, with the development of artificial intelligence, the boundaries of technological ethics and privacy have received more and more attention and attention from all walks of life. For example, related issues such as privacy management and employment gap have caused widespread discussion in the industry.

In this regard, Xu Li, co-founder of SenseTime Technology, recently shared the AI ​​ethical governance concept that advocates "development". "There is a very important goal of artificial intelligence ethical governance, which is to use artificial intelligence to promote the overall universal development of society." In addition, SenseTime previously proposed the "AI governance triangle" of "sustainable development, people-oriented, and technology controllable."

In order to implement specific actions, in March this year, SenseTime and Shanghai Jiaotong University held the unveiling ceremony of the "Center for Computational Law and AI Ethics Research" to work together to create a rich case basis for artificial intelligence governance and provide a basis for case studies for AI ethics and governance; in addition, SenseTime Technology has also established an ethics committee, security committee and product committee internally to formulate and implement strict ethical standards for the use of artificial intelligence technology to ensure maximum respect and protection of personal privacy, so that AI technology can be used correctly...