Author: Hao Junhui Source: IT Times
"Have you seen a blind man in the last month?" On December 12, at the WAVE SUMMIT+2021 Deep Learning Developer Summit hosted by the National Engineering Laboratory for Deep Learning Technology and Applications, Gao Hongzhi, a senior at Northeastern University, threw a question to the audience.
There are 17 million visually impaired people in China's 1.4 billion people, accounting for 1.2% of the total population, but they are rarely seen on the streets. "Discoloration" and obstructed blind lanes have become obstacles to their travel, while there are only more than 200 trained guide dogs in the country, and each guide dog takes 3-5 years to train.
"We trained a robotic guide dog in less than half a month." Using the flying paddle deep learning open source open platform and community, Gao Hongzhi and his team let a machine guide dog achieve vision-based blind lane navigation, common obstacle detection, traffic light recognition to guide crossing the road and other functions, and in the future, through AI training, in addition to guiding obstacle avoidance, machine guide dogs can also provide owners with warmer and safer voice companionship, many things that existing guide dogs can not do, will become a reality through it.
The young Gao Hongzhi is a flying paddle developer technical expert (PPDE), one of the 4.06 million flying propeller developers. As China's first self-developed industry-grade deep learning platform, as of now, Feipao has created 476,000 models, serving 157,000 enterprises and institutions, ranking first in the comprehensive market share of Deep Learning platforms in China.
"Artificial intelligence presents the characteristics of 'fusion innovation' and 'lowering the threshold'." Wang Haifeng, chief technology officer of Baidu and director of the National Engineering Laboratory for Deep Learning Technology and Application, believes that on the one hand, there are more and more integration and innovation of AI technology and industry; on the other hand, although AI technology is becoming more and more complex, the threshold for AI development and application is getting lower and lower.
This means that deep learning is pushing artificial intelligence into the stage of industrial mass production.
At the summit, Feipao released ten latest technological and ecological progresses, including the new panorama of Feipao, the new Wenxin model of the industrial model library, the industry's first industrial practice example library, and the Feipao "Great Navigation" 2.0 co-creation plan.
An AI project that a person can also complete
Li Sangyu, a 25-year-old railway worker in the Xiangyang railway section, developed a set of "automatic detection system for the body character and logo of railway freight cars" by teaching himself a flying paddle development kit to realize the automatic identification of railway freight car numbers. The whole project,
He alone. The work that used to require manual verification for several hours now takes only 3 minutes to complete, saving more than 200,000 yuan in costs for the Xiangyang depot.
On the flying paddle platform, there are 4.06 million developers like Gao Hongzhi and Li Sangyu. Feipao is the first open source, open, fully functional industry-level deep learning platform in China, it solidifies a large number of basic and repetitive code into models, and uses pre-built and optimized component sets to define models, deep learning researchers do not need to deeply understand the underlying algorithm, they can independently and quickly realize the entire process from building data sets, to model training, to model deployment, thereby greatly reducing the threshold for the use of artificial intelligence.
At the summit site, Xin Zhou, director of Baidu's AI product research and development department, demonstrated a 5-minute demonstration of a robot dog. Flying Propeller's newly released intelligent edge console, which only takes 5 minutes, allows a robot dog to learn a new skill of recognizing gestures. For Gao Hongzhi, the time for a more humane machine guide dog to "go out on the street" can be shortened.
It is worth mentioning that through the multi-level, low-cost hardware adaptation scheme, the adaptation cost of the framework and the chip is greatly reduced. Taking cambrian MLU adaptation as an example, compared with the original plan, the number of code lines is reduced by 69.4%, the modified code file is reduced by 62.3%, and the labor input cost is reduced by 60%. Up to now, Feipao and 22 domestic and foreign hardware manufacturers have completed the adaptation and joint optimization of 31 chips.
AI has entered a period of industrialized large-scale production
At the first WAVE SUMMIT+2021 Summit in April 2019, Wang Haifeng defined the deep learning framework as an "operating system in the intelligent era", which is connected to the chip and applied, especially in the large-scale production stage, which can output AI technology to all walks of life in a standardized, automated and modular manner to achieve large-scale application; at the same time, based on the platform to promote integration and innovation, common development, cohesion of all parties, by empowering the majority of developers, strongly support the AI industrial production, Promote technological innovation and industrial intelligent upgrading.
Originating in July 2018, the open source framework v0.14 was released, and Baidu officially opened 10 models such as CV/NLP/voice/reinforcement learning for the first time, providing support for the underlying capabilities of the whole process of deep learning, from data preprocessing to model deployment.
After three years of development, the flying propeller continues to break through the innovation of the core framework. The newly released open source framework v2.2, a large number of new scientific computing APIs, end-to-end adaptive large-scale distributed training technology that efficiently supports super model training, the whole process accelerates text tasks, solves the pain point problem of text field development in terms of performance and training and pushing, and in the flying paddle industry-level model library, the new knowledge enhancement wenxin big model can make the big model really enter the industrial application.
At the summit site, Xiao Fei, deputy director of the dispatching center of the State Grid Shanghai Electric Power Company, talked about the State Grid's desire for "artificial intelligence". Based on the goal of double carbon, the proportion of renewable energy sources such as wind and solar energy in the grid is gradually increasing, but it poses a very big challenge to the system management of the entire power grid, especially the uncertainty caused by changes in natural conditions, which may have a catastrophic impact on the entire grid. This requires the use of deep learning, supervised learning, unsupervised and other models to accurately predict the proportion of renewable energy consumption, load and other resources.
Xiao Fei introduced that through the Baidu flying paddle platform, the new energy prediction accuracy of the State Grid has increased by 85%, and the intelligent arrangement has been improved from the minute level to the second level, laying a very good foundation for the operation of the entire power grid.
Wu Tian, vice president of Baidu Group and deputy director of the National Engineering Laboratory for Deep Learning Technology and Application, pointed out that in recent years, the scale of AI developers in cities across the country has increased year by year, and the number of enterprises applying artificial intelligence has shown a scene of blossoming in many places and prospering in multiple industries. From general scenarios such as recommendations, to industry derivative scenarios such as intelligent dispatch of customer service systems, to key scenarios in the industry such as power generation forecasting... Similar to this technology and industry integration innovation cases in the flying paddle platform more and more, but also more and more professional, as of now, the flying propeller to serve 157,000 enterprises and institutions, effectively promote the industrial intelligent upgrading.
Cultivate compound AI talents
For the artificial intelligence industry, talent is a topic that will never go out of style. With the deep integration of AI and industry, more and more composite AI talents who understand AI and have industrial experience are needed.
"Universities should not use yesterday's knowledge to teach today's students to face tomorrow's needs." Xie Shaorong, dean of the School of Computer Engineering and Science at Shanghai University, has been thinking about how to incorporate the latest products developed by the industry into the professional training of talents in a timely manner. In his view, for undergraduates, what needs to be done in the lower grades is the cultivation of system knowledge, but in the senior grades, after having a certain common professional foundation, you can cooperate with Internet companies such as Baidu to introduce advanced algorithm models and open source platforms into colleges and universities to enhance the linkage effect of talent training.
Similar thinking exists in most domestic colleges and universities, and different colleges and universities are also exploring the cultivation of a new generation of AI talents.
Xi'an Jiaotong University is a well-known engineering college in China, as early as ten years ago, it began to explore the "elite class" model, and technology and industry frontier leading enterprises cooperation, by the school and enterprises to jointly develop training programs, jointly build a curriculum system, the tutor of the enterprise to participate in the school's theoretical course teaching, guide students' project design, science and technology planning, the current cooperation of enterprises include Baidu, Huawei, ZTE, 360 and so on.
Through the use of multidisciplinary and interdisciplinary selection methods, Xi'an Jiaotong University selects some good students with spare learning, conducts small class teaching after the establishment of the "elite class", and organically integrates the students' original majors and the frontier direction of the "elite class" for training. Up to now, 23 "elite classes" have been established in the whole school. Judging from the results, the fresh graduates who participated in the "elite class" generally believe that their awareness of innovation, engineering design and development practice ability have been greatly improved.
Internet companies are also extremely eager for such compound AI talents. All along, Feipao has been committed to the cultivation of compound AI talents, and has reached a number of talent training programs with many well-known colleges and universities. Since the "National Artificial Intelligence Teacher Training Class" was held in 2018, Baidu Feipao has trained more than 2,900 college teachers, covering more than 690 colleges and universities. In 2021, the "AI Talent Industry-Education Integration Training Program" was launched, which provides a reference industry-university-research practice plan for AI talent training from the content, special cooperation and service levels.