Baidu Presents Top 10 Frontier Technology Inventions of 2022
Baidu yesterday announced its first list of Top 10 Frontier Technology Inventions of 2022, highlighting ten of Baidu's cutting-edge creations introduced this year with the potential to significantly shape the world we live in over the next decade.
With a focus on AI and autonomous driving, the list includes the world's first cross-modal AI-generated content (AIGC) model to unify visual-language understanding and generation, a multi-sensor fusion system to allow driverless cars such as Baidu's robotaxis to navigate complex urban streets, and a knowledge-enhanced large language model with exceptional understanding and creative writing skills.
Introducing the list, Baidu CTO Dr. Haifeng Wang said as a leading AI company, Baidu is committed to accelerating innovation in the frontier areas of technology. At the same time, Baidu seeks to empower partners, customers, and the wider ecosystem, harnessing innovation to promote industrial development and allow users to determine the course of their future development.
With over a decade of intensive investment in AI, Baidu's cumulative R&D spending totaled RMB 100 billion yuan as of this year. Baidu has ranked first in China in terms of AI patent applications and authorizations for four straight years. In 2021, Baidu led the world in the number of deep learning patent applications and the number of autonomous driving patent families. The company also won the only Chinese Patent Gold Award in the field of AI interaction, making Baidu the high-tech company that has won the most Chinese patent awards and the highest award level in the field of AI.
Zhixiang Liang, senior vice president of Baidu, said since day one, Baidu has attached great importance to independent innovation and patent protection, building a complete intellectual property protection mechanism to stimulate quality innovation.
Baidu's top ten frontier technology inventions of 2022 are:
1. General and controllable cross-modal AIGC
The invention proposes a unified model for visual-language understanding and generation for the first time in the industry. By enhancing content generation with knowledge integration, the invention overcomes the bottleneck of general and controllable content generation and paves the way for downstream tasks like text generation, image generation, video generation, and digital human generation. Baidu's AIGC model has been applied to a number of innovative products such as the first image-text-to-video production tool VidPress, high-precision digital human generation, text-to-image generation, ramping up the efficiency of content production, and opening a new chapter in AI content generation.
2. Multi-sensor fusion processing system for driverless vehicles
The multi-sensor fusion system improves the perceptual ability of LiDAR and builds an independent closed-loop ability to provide a surround view. This invention has been applied on a large scale in autonomous vehicles, reducing missed detection in real-world settings by 60% (83% for low obstacles), effectively strengthening the Baidu robotaxis' driverless capabilities on complex roads in various urban scenarios.
3. Knowledge-enhanced large AI model
The invention is a core technology of Baidu's large model family, "Wenxin (文心)," which integrates large-scale knowledge and massive data and enables exceptional understanding and generation capabilities. The model family includes large language models – represented by the world's first 100 billion-level knowledge-enhanced model PCL-Baidu Wenxin - computer vision models, cross-modal models, and other large models, as well as industry models fine-tuned for application in power grids, finance, and aerospace. These models have achieved state-of-the-art results in over 100 benchmark tasks and have been applied to various Baidu products on a large scale while also serving intelligent upgrades across a wide range of industries through PaddlePaddle and Baidu AI Cloud.
4. General-purpose heterogeneous parameter server architecture for deep learning
Through a scalable architecture design, the inventions can not only support the training of models on parameter server architecture of any type of hardware such as CPU, GPU, and XPU, but also assign the same deep learning model training task to different types of hardware of computing nodes, through reinforcement learning algorithms. The inventions perform hybrid hardware heterogeneous training, thus achieving optimal collocation of computing resources, which can reduce the training cost of deep learning models by more than 50% and effectively improve the training efficiency of deep learning models.
5. PaddleHelix, an AI-based computational biology platform
PaddleHelix encompasses a series of AI-based computational biology innovations, including the self-developed LinearDesign algorithm for efficient mRNA vaccine design, the world's first compound representation learning model based on geometric space conformation HelixGEM, and the end-to-end single-sequence protein structure prediction model HelixFold-Single. These have significantly improved the efficiency of new drug research and development and vaccine design and helped fight the Covid-19 pandemic.
6. Key V2X technologies for autonomous driving
Building a complex technology system and fusion cooperative mechanism for V2X-powered autonomous driving, the invention enables cooperative perception between vehicles and infrastructures, solving numerous perception long-tail problems such as dynamic and static blind spots, over-the-horizon, and occlusion. Through collaborative decision-making planning and collaborative control, the invention simplifies navigation of lane changes, complex interactions, congestion, and extreme corner cases in a mixed state, reducing the number of human interventions and the risk of safety accidents while ensuring the safe and continuous operation of autonomous vehicles.
7. All-platform quantum hardware-software integration
The invention is centered around Liang Xi, the world's first all-platform quantum software-hardware integration solution based on Qian Shi, Baidu's 1st superconducting quantum computer. Liang Xi offers versatile quantum services through private deployment, cloud services, and hardware access, which streamlines the process of deploying quantum hardware to quantum services. Liang Xi can connect with many types of mainstream quantum chips, such as superconducting quantum chips and ion traps, making quantum chips "plug and play."
8. Intelligent production of digital humans
Built on AI technologies such as voice, semantics, and vision, and using smart terminals as the carrier to achieve human-machine visual voice interaction services, Baidu AI Cloud's Xiling platform provides a one-stop service for digital humans, from production of digital humans to personality management, content generation and business operational process automation. Combined with the UNIT7.0 cross-modal dialogue engine and AIGC technology, the invention supports the efficient production and operation of service-oriented and performing digital humans of 2D, 3D, and cartoon types.
9. All-element dual-bus technology for smart cities
The all-element dual bus includes both the intelligence bus and the knowledge bus. The intelligence bus serves as a collaborative development and operation environment for all resources of urban services, and the knowledge bus relies on a multimodal Wenxin large model tailored for smart cities and an all-element fusion graph that integrates city data, knowledge, and algorithms. This invention can efficiently be used to build urban applications.
10. Multimodal pedestrian motion prediction for autonomous driving
The invention proposes a deep learning model incorporating multimodal input features and multidimensional interaction patterns to predict the motion trajectory of pedestrians in the next six seconds. It has been deployed to Baidu's robotaxis, with a 30% improvement in pedestrian prediction precision and recall and a 95% solving rate of VRU collision risk problems.
In the past ten years, Baidu's annual R&D expense rate has been above 15%, and the R&D expense of Baidu Core's revenue in 2021 reached 23.21%, leading among the world's large Internet companies. In the future, Baidu is dedicated to continuing to delve deep into technology innovation and enable intelligent transformation across industries.