"Top Ten Science and Technology Frontier Invention" was released, cross -modular AIGC, intelligent production of digital people, and autonomous driving

Cover Journalist Fu Wenchao

On September 20th, Baidu first released the "2022 Top Ten Technology Frontier Inventions", including cross -modular universal control AIGC, unmanned vehicle multi -sensor fusion processing system, knowledge enhanced model, deep learning universal heterogeneous parameter parameter server architecture, based on based on AI's biological computing platform, key technologies for autonomous vehicle -oriented vehicle roads.

AIGC is considered a new generation of technology to subvert existing content production methods. It is understood that the cross -modular and controllable AIGC in the "Top Ten Science and Technology Frontier Invention" proposes a unified modeling technology of cross -modular and cross -tasks in terms of universality. Essence

Some experts said that in recent years, my country has regarded scientific and technological innovation as the first driving force to lead development, implemented the innovation -driven development strategy, and opened a new journey to build a world scientific and technological power. The key core technology patent is the heavy weapon of the country. The achievement of the "top ten technology frontiers" reflects the realization and breakthrough of a number of international frontier core technologies, which is of great significance to promote economic development and technological progress.

2022 Top Ten Technology Frontier Inventions, the specifics are as follows:

1. Cross -modal GM controlled AIGC: The invention of the industry proposed the unified understanding and generation model of language and visual unity in the industry, innovated the generation of integration of knowledge, broke through the technical bottleneck of general controlled content generation, realized text generation, image generation, image generation , Video generation, practical leaping of digital people.

2. Multi -sensor fusion processing system of unmanned vehicles: This invention takes multi -sensor fusion as its core, improves the ability of laser radar perception, and builds an independent closed -loop capability of visual vision. The total inspection rate was reduced by 60%, and the inspection rate of low obstacles was reduced by 83%. It effectively supported the construction of Baidu Robotaxi in various cities and unmanned driving capabilities on various cities.

3. Large model of knowledge enhancement: This invention is the core technology of Baidu Wenxin's large model. It integrates learning from large -scale knowledge and massive data, and has a leading understanding and generating ability. Including the world's first big model Pengcheng-Baidu Wenxin, the world's first 100 billion-level knowledge enhancement of big models, computer vision, cross-model and other large models, as well as large models such as power, finance, and aerospace.

4. Deep learning General heterogeneous parameter parameter server architecture: This invention can not only support separate training of parameter server architecture based on different types of hardware such as CPU, GPU, and XPU through scalable architecture. Learning model training tasks are allocated on the computing nodes of different types of hardware for mixed hardware heterogeneous training to achieve the optimal mix of computing resources.

5. AI -based biological computing platform PaddleHelix: This invention proposes a series of AI -based biological computing innovation technologies, including independent developing high -efficiency MRNA vaccine design LINEARDESIGN algorithms. The first compound characteristic model HELIXGEM based on geometric space conformity, end -to -end Single-serial protein structure prediction model HelixFold-Single, etc.

6. Key technology for autonomous driving -oriented vehicle and road synergy: This invention builds a complex technical system and integration and synergy mechanism for vehicle road collaborative autonomous driving. Through synergy perception, a series of series of autonomous driving static blind spots, overvasing distances, and blocking are solved Perceive long -tail problems, solve the problems of reality changes, games, blocking and extreme scenes in the mixed state through collaborative decision -making planning and collaborative control to reduce the number of autonomous driving takeover and safety accident risks, and ensure continuous safety operation of autonomous driving.

7. Full platform quantum soft and hard integration: The invention focuses on the layout of "quantity". A series of services such as cloud services and hardware access simplify the entire process of quantum hardware deployment to quantum services to the greatest extent.

8. Intelligent production of digital people: The invention is based on AI technologies such as voice, semantics, and vision, and uses intelligent terminals as a carrier to realize human -machine visual voice interaction services. With the help of Baidu Smart Yunxi Ling Platform, we can open the entire process of digital people's production, human settlement, and content creation to business arrangement.

9. Smart City Full Element Dual Bunch Technology: The full factor dual bus of this invention includes smart bus and knowledge bus. The smart bus builds the full resource coordinated development and operating environment for urban services, and uniformly disables all kinds of data and algorithms in the city. , Computing power, provide comprehensive technology supply for intelligent application development; the knowledge bus relying on the city's enhanced city multi -modal model, knowledge map, NLP and other AI technologies to build a city -based factor integration of knowledge map, support cross -department, cross -application, cross -application The knowledge system between across data is opened and resources are shared.

10. Autonomous driving multi -modal pedestrian movement prediction: The invention proposes a deep learning model that integrates multi -modal input features and multi -dimensional interactive mode, which is used to predict the sports trajectory of pedestrians in the future. Pedestrian forecasting was increased by 30%, and the solution of the VRU collision risk problem was 95%.

- END -