Table of Contents
1. A new era created by artificial intelligence
2. The mission opens the flying paddle to the exclusive spring
3. Technological breakthroughs establish the flying paddle brand Yiqijuechen
4. Industry application accumulation The flying paddle brand is thriving
5. Ecological communication makes the flying paddle brand unique Yan
6. The status quo and future thinking of deep learning platforms
The status quo and future thinking of the deep learning platform
As China's first feature-rich, open source and open deep learning Chinese platform, Baidu Flying Paddle fully highlights the characteristics of self-developed technology and innovative breakthroughs. It not only completely realizes the comprehensive independence of technology in the underlying framework, It also provides a foundation and practice for industry, academic, and scientific research innovation with continuous open source core capabilities. Its benchmarking significance and demonstration reference have global value. Flying Paddle, as a leader, leads the top brands in the domestic technology industry to participate in the common progress of industrial-level deep learning platforms such as MegEngine, Huawei MindSpore, and Jittor, as well as Xiaomi MACE, Ali MNN, A number of reasoning engine tools such as Tencent NCNN and OPEN AI LAB (Open Intelligence) Tengine have been open sourced, promoting the innovation and development of open source algorithm frameworks. At the same time, the construction of AI open platforms in basic technologies such as voice and vision, open computing, and vertical fields continues to advance and develop rapidly, gradually expanding the domestic open source and open platform ecosystem.
According to incomplete statistics, there are currently more than 40 domestic open source and open platforms, and their service capabilities have been further enhanced, which has the ability to accelerate the innovation and development of the domestic artificial intelligence industry. In terms of basic technology service platforms, such as Ali, Didi, Tencent, Netease, JD.com, etc. have established and operated comprehensive AI capability open platforms. Open platforms for visual image recognition AI capabilities such as Cloudwalk, Megvii, Meitu, EZVIZ, ArcSoft, etc., open platforms for speech recognition AI capabilities such as iFlytek, Yitu, Xiaomi Xiaoai, Sogou, Xueersi, and Youdao. In terms of open computing service platforms, such as Alibaba Cloud, Huawei Cloud, Baidu Cloud, JD Cloud, Kingsoft Cloud, and Tencent Cloud have established and operated AI cloud computing service open capability platforms. In terms of open source and open platforms in the vertical field, there are Baidu Apollo automatic driving open platform, Ali city brain open platform, Tencent smart medical open platform, etc. Among them, the Baidu Apollo platform has brought together 210 ecological partners around the world. More than 80,000 developers in 135 countries around the world use Apollo open source code, and the number of open source code exceeds 700,000. The Apollo autonomous driving platform has become the most powerful and open in the world. , the most active autonomous driving platform.
The AI application open platform accelerates the construction of the industrial ecology. For example, it has organized the establishment of the first open source foundation in China - the Open Atom Open Source Foundation. Huawei OpenHarmony, Tencent Tiny OS, Alibaba Ali OS and other 10 projects have been donated to the foundation for incubation. "Owned by one family" is "open source and shared", attracting the upstream and downstream of the industrial chain to jointly build an ecology. The new generation of artificial intelligence industry technology innovation strategic alliance organizes all parties in the industry, academia and research to cooperate to build the "Openl Qizhi Open Platform", which brings together open source software, open source hardware and open data, aiming to promote open source, open and collaborative innovation in the field of AI, and open up the AI technology chain , innovation chain, ecological chain and other links in the industrial chain resources, promote the wide application of AI in various fields of social economy, and accelerate the construction of industrial ecology.
The exploration of multi-modal large models is active. The pre-training model is analogous to the infrastructure of the algorithm layer. By building a general-purpose model device that can solve multiple problems in the fields of voice, vision, and natural language processing at the same time, when AI algorithms are used by thousands of industries, It can directly call the general model interface and use a small amount of data for fine-tuning to solve application tasks in specific fields. It is the basic device for AI algorithms to realize large-scale application replication. Since 2020, domestic and foreign technology giants such as Google, Microsoft, Nvidia, Zhiyuan Artificial Intelligence Research Institute, Ali, Huawei, Baidu, and Inspur have launched large-scale research and exploration. At present, the implementation of ultra-large-scale pre-training models in the fields of content idea generation, language and style conversion, and dialogue is progressing rapidly. With the continuous improvement of the performance of the future model and the continuous improvement of the platform, the large model may become the basic platform of the next generation of AI and empower all walks of life. At present, domestic ultra-large-scale pre-training models are progressing rapidly. First, the large-scale self-developed models of leading enterprises have reached hundreds of billions of parameter scales. For example, the Pangu series of ultra-large-scale pre-training models released by Huawei in April 2021, including the world's largest visual (CV) pre-training model with 3 billion parameters, and the 100 billion parameters and 40TB training model jointly developed with Cycle Intelligence and Pengcheng Lab The world's largest Chinese language (NLP) pre-training model of data, opening a new model of industrialized AI development. The ultra-large-scale intelligent model "Enlightenment 2.0" released by Beijing Zhiyuan Artificial Intelligence Research Institute has a model parameter of 1.75 trillion, surpassing the Google Switch Transformer model and becoming the world's largest pre-training model. Inspur released the world's largest AI massive model "Yuan 1.0" in September 2021. The Ali M6 large model has become the world's largest AI pre-training model with 100,000-level parameters. Second, efficiency continues to improve, and commercialization scenarios are increasingly enriched. In the past 10 years, the computing resources used for AI training models have exploded, and the computational complexity of AI training has increased by 10 times every year. Compared with OpenAI's GPT-3 at the same parameter scale, the Ali M6 large model consumes only 1% of the energy consumption of GPT-3, and its efficiency is nearly 11 times higher. As the first multi-modal large-scale model commercialized in China, M6 has been applied in more than 40 scenarios, with hundreds of millions of daily calls.
Alibaba Cloud's AI open service, relying on the technological innovation of DAMO Academy, leads the development of visual AI capabilities. Ali AI (Ali Lingjie) relies on Alibaba Cloud infrastructure, big data and AI engineering capabilities, scene algorithm technology, etc., to provide enterprises and developers with a cloud-native AI capability system. Alibaba Cloud is the position where Alibaba provides AI open services. Currently, the AI open services provided by Alibaba Cloud include visual AI capabilities in more than ten subdivision directions and more than a hundred scenarios (Alibaba Cloud Visual Intelligence Open Platform), as well as speech recognition and speech synthesis. , voice analysis, machine translation, intelligent growth engine and other AI capabilities. Relying on the underlying technology innovation of Dharma Institute, Alibaba Cloud's commercialization and ecological capabilities, Alibaba Cloud's visual intelligence open platform provides easy-to-use and inclusive visual API services for visual intelligence technology companies and developers (including developers). Provides more than 200 visual capabilities. Taking Alibaba Cloud ET City Brain as an example, it opens up platform AI capabilities to all ecological participants related to urban governance, serves medical care, urban management, environment, tourism, urban planning, security, people's livelihood and other fields, and provides comprehensive artificial intelligence for cities. center. ET City Brain's open AI capabilities come from the four major facilities of Alibaba Cloud's intelligent integrated computing platform, data resource platform, intelligent platform, and application support platform, and are widely connected to ecological partners such as governments, scientific research institutions, and entrepreneurs.
Tencent's AI open platform integrates internal superior AI resources and continuously builds a leading edge in all dimensions. With Tencent Cloud and AI Open Platform as the core, Tencent has opened up Tencent's full-stack AI product capabilities. Tencent Cloud is the commercial export of Tencent's AI capabilities. It mainly promotes the combination of AI technology and various industries, and continuously releases the value of AI applications. Tencent AI Open Platform gathers technology, professional talents and industry resources. Through continuous integration of internal advantageous AI resources, relying on the AI technology capabilities of Tencent AI Lab, Tencent Cloud, Youtu Lab and partners, it will continue to build an all-dimensional leading edge and build an industry Intelligent new engine. At present, Tencent Cloud has completed the overall layout of AI new infrastructure from multiple levels, based on the needs of internal business (games, social) audiences, and the landing projects incubated by internal research topics, coupled with some needs and pain points (medical care) from internal strategic research research. . Tencent's AI technology strength is mainly concentrated in the four major sectors of computer vision, speech recognition, natural language processing, and machine learning. In addition to social and games that are more linked to the main business, AI+medical and AI+pharmaceuticals are also the main directions of attack in the past two years .
HKUST Xunfei Open Platform aims to dig deep into industrial application value for industrial digitalization. On October 25, 2021, at the opening ceremony of iFLYTEK Global 1024 Developers Festival, iFLYTEK Open Platform 2.0 was officially released, from simply providing various AI capabilities to partners, and advanced to deeply mining industries for industrial digitalization application value. Xunfei Open Platform 2.0 and industry leaders jointly build a baseline base, open application scenarios to developers through industry low-code and zero-code, solicit application ideas, and use standard systems, testing platforms, certification systems, training platforms, low-code development platforms, and development The six major measures of the winners contest support the implementation of the open platform 2.0 strategy. In the future, the Xunfei open platform will open up a number of technologies, including global coverage and flexible customization of multiple languages. Through the integration of production and education, more artificial intelligence talents will be cultivated, and more high-quality resources will be released to empower outstanding entrepreneurs. At the same time, formulate industry standards, promote the healthy development of the industry, and join hands with ecological partners to create more industry benchmark solutions and create more valuable industry scenarios.
It is predicted that driven by policy dividends and demand in downstream application fields, the deep learning open platform market will grow steadily. On the basis of the current domestic market size reaching 10 billion yuan, the golden decade is expected to achieve a compound annual growth rate of 60%. Adhering to the spirit of open source, building an open ecology, and creating an open platform will greatly reduce the cost of enterprise application, increase the willingness of enterprises to use it, and continue to promote the prosperity and development of the AI market. The power and potential of the deep learning open platform to bring together industry experts, professionals, developers and scientists is unlimited. In the face of such a rapidly changing society, the urgency and value of working together and sharing information is obvious.
The new generation of artificial intelligence open innovation platform led by the Ministry of Science and Technology of the People's Republic of China, the top brands in the technology industry have become the main force in its construction. The Ministry of Science and Technology pointed out that the construction concept of the open innovation platform is: focus on key subdivisions of artificial intelligence, give full play to the leading and demonstration role of industry leaders and research institutions, effectively integrate technical resources, industrial chain resources and financial resources, and become the core of continuous output of artificial intelligence An important innovation carrier for R&D capabilities and service capabilities. Its essential connotation is to build an ecological chain based on the deep learning open platform, build a full-stack capability extending to the upstream and downstream of the industrial chain, and empower traditional enterprises to transform intelligently. Driven by the influence of technology companies to create an intelligent ecology for vertical industry applications, traditional enterprises rely on their own huge product system and market share to actively promote their own intelligent transformation strategies. It can be seen that "opening and sharing" has become an An important concept for industrial development. For domestic enterprises, building a deep learning platform and ecology through leading brands, and then driving small and medium-sized related enterprises and technical teams to develop together, can effectively strengthen the comprehensive support of AI technology and industry for technology, economy, social development and national security.
From a global perspective, how the deep learning open platform can give full play to the greatest advantages of infrastructure and provide value and support for ecological partners has become the key to determining its competitiveness. AI is being internalized into a digital base, gradually becoming a basic capability in the digital age. In 2001, when the MIT Technology Review first launched the "Top Ten Breakthrough Technologies", natural language processing was on the list. In the next two decades, technological breakthroughs in machine learning methods and learning models such as deep learning, reinforcement learning, adversarial neural networks, and GPT-3 (new technological breakthroughs in the field of natural language processing) have continuously made machine intelligence move toward "human intelligence." "Get closer, step by step to push the implementation of AI technology in multiple business scenarios to the climax of application. From smart watches, voice assistants, AR, VR and other hardware to payment, autonomous driving, medicine and other industry-level applications, they have become the basic capabilities and smart bases of smart products in the era of digital economy.
AI has reached its current level of development in the past ten years, and deep learning is undoubtedly the biggest technical contributor. The core technology of deep learning is inspired by the neurons and synapses of the brain, learning from large-scale data, making predictions, and making machines smarter. Intelligent speech, natural language processing, computer vision and reinforcement learning are the main technical lines of AI in the first ten years and have achieved remarkable results. In terms of natural language processing alone, the development of deep learning has led to major breakthroughs in natural language processing technology at the cognitive level, which has been widely used in part-of-speech tagging, grammatical analysis, sentiment analysis, information retrieval, automatic question answering, machine translation, and relationship classification, etc. key tasks in natural language processing.
Artificial intelligence includes three levels, namely computational intelligence, perceptual intelligence, and cognitive intelligence. Driven by the continuous upgrading of computing power and storage means, computing intelligence has been realized. Driven by the development of technologies such as the mobile Internet, big data, and cloud computing, the value of unstructured data has been valued and tapped, and perception intelligence such as voice, image, video, and contact is developing rapidly. Based on the development of computational intelligence and perceptual intelligence, AI is extending to cognitive intelligence capable of analysis, thinking, understanding, and judgment. At the critical point of technological evolution, the dividends of deep learning need to be further explored. In the process of driving AI to realize the progress from perception to cognition, deep learning will continue to promote the breakthrough and development of core technologies such as natural language processing and other cognitive layers. And in basic theoretical research, from particle physics, structural biology to cosmology, deep learning can learn features in large data sets, classify different objects, discover scientific laws, and innovate basic scientific theories. For example, the University of Sussex and University College London use deep learning technology to simulate the dynamics of the sun, planets and large satellites in the solar system by learning 30 years of trajectory data through the graph neural network, and found that the analytical expression of the mechanical laws of the model is equivalent to that of Newton The law of universal gravitation and estimate the mass of celestial bodies.
In the golden decade, with the spiral innovation of underlying technologies such as higher dimensions, more autonomy, multi-modality, and large models, and the deepening penetration of industry applications, quantum computing, unsupervised learning, shallow learning networks, and computing power progress will become The center of gravity, computer vision, natural language understanding and communication, cognition and reasoning, game and ethics, and other technologies, the development trend of mutual penetration and integration is more obvious. Self-supervised learning is the core breakthrough direction of unsupervised learning. At present, the development of self-supervised training large models is conducive to the systematic development of AI. At the same time, the future development of the computing power layer will also present three characteristics, one is the state of a variety of architectures blooming, the other is the rapid development of centralized computing power and edge terminal computing power, and the third is the growing trend of dedicated computing power. Technological breakthroughs will allow AI to gradually possess the cognitive flexibility and autonomous awareness of the human brain. In terms of algorithms, the recent new algorithms focus on improving the quality of data and the scale of model parameters, digging deep into the development potential of existing technology paths, and making up for the depth through large-scale pre-training models, self-generated data, relying on knowledge graphs and common sense relationships, and using multi-source data. Learning has limitations in terms of generalization, small data, interpretability, and independent learning ability, and continuously improves the level and depth of problem solving, leading the industry's innovative development and structural upgrading. For example, Google proposed ModelSoup, a method based on pre-trained model fine-tuning, which refreshed the accuracy of ImageNet to a new height of 90.94%.
The construction of an open platform for deep learning must be done for a long time in order to be the cornerstone of future prosperity. In the next ten years, China is expected to become the world's largest artificial intelligence market, accounting for a quarter of the global market share. Therefore, the open platform and ecology of deep learning are the only way to overtake the car in a corner and occupy the leading position in global technology. The deep learning open platform provides a common infrastructure for the win-win cooperation of all levels and links of the industry, which is the top priority of redefining China's creation.
Looking forward to the golden decade, the deep learning platform will evolve and iterate around the three dimensions of technical strength, functional experience, and ecological model. From the perspective of technical strength, the deep learning platform will better integrate cutting-edge technologies such as self-supervised learning, reinforcement learning, transfer learning, and small data on the basis of deep mining and optimization of mature technologies, and help continue to expand by creating simple and efficient innovative tools Apply vision and capability boundaries to promote innovation and progress in all dimensions of production and life. From the perspective of functional experience, Chinese enterprises and industries have their own characteristics, and their requirements for AI functions are also unique. In the future, the deep learning platform will fully pay attention to the individual needs of Chinese enterprises and deconstruct the business form, and further optimize the platform functions based on the characteristics of subdivided scenarios. Details, to achieve flexible adaptation to different industries, enterprises and business links, and continuously improve the technical capabilities for auxiliary decision-making, forecasting, reasoning and other links, to help enterprises establish leading advantages in the intelligent era. From the perspective of the ecological model, the iterative opening of the ecology will be an important symbol of the deep learning platform system. With the rise of consumer sovereignty, the functional architecture of the deep learning platform system will become more and more complex. This complexity comes from the personalization of customers, the diversity of scenarios, and the complexity of the technology itself. The main body of the platform From the platform owner to the joint development of "platform owner + industry customers", the content of the platform has changed from limited, closed, and customized functional modules to massive, open, and versatile technical tools, so as to build a mutual relationship between the platform and users. A promotional, two-way iterative ecosystem.
The intelligent transformation of the whole industry driven by the deep learning platform has kicked off, helping enterprises and even the country to gain development opportunities in the era of digital economy and intelligent economy. In the face of the sudden impact of the epidemic, the deep learning platform has been widely used to effectively prevent the spread of the epidemic, accelerate drug research and development, promote enterprise transformation, and help resume work and production. The social value has begun to show. From the perspective of a longer cycle and a broader perspective, the deep learning platform will continue to increase the proportion of high value-added products in traditional industries, further optimize the industrial structure, and enhance the national industrial economic resilience and anti-risk capabilities. It is an urgent task of sex, times and reality. It requires the government, scientific research institutions, artificial intelligence companies, and traditional industry companies to work together to create a positive and healthy industrial ecology.
In fact, the competition of deep learning platforms shows that it is a technical competition, but it is actually a competition for the entrance, standards, rules, and voice of the artificial intelligence industry. The domestic deep learning open source platform represented by Flying Paddle shoulders heavy responsibilities and is brave and diligent. Using Paddle as a boat, it teaches people how to fish, serves tens of millions of users, and empowers thousands of industries. As mentioned in "Top of the Wave", there are three things China's computing power industry should do: one is to build the scale of infrastructure, the other is to improve the efficiency of computing power, and the third is to provide sufficient development capabilities in infrastructure. "To accomplish these things, computing power can truly benefit the enterprise and individuals. And self-cultivation, saving people, flying paddles all the way, are the supreme wise men and practitioners with this universal feeling.