The eighth event of the Zhejiang University Postdoctoral Academic Salon: NetEase Fuxi discusses new areas of intelligence

On June 5, the eighth session of the Zhejiang University Postdoctoral Academic Salon concluded successfully at the Yuquan Campus of Zhejiang University. This event is a joint academic salon organized by the Zhejiang University Postdoctoral Association and NetEase Fuxi Laboratory. The four keynote speakers, Ph.D. Fan Changjie, Ding Yu, Li Lincheng, and Zhao Zeng, are all from NetEase Fuxi Laboratory. The theme of this salon is the work results of NetEase Fuxi Laboratory in the field of artificial intelligence. Fan Changjie shared the research results on agent-oriented programming, Ding Yu introduced related research on digital virtual humans, Li Lincheng explained the progress in visual computing, and Zhao Zeng focused on sharing the results of AIGC system construction. Through this salon, participants had an in-depth understanding of the innovative achievements of NetEase Fuxi Lab in the field of artificial intelligence and had in-depth exchanges and discussions. This salon event provides a valuable academic exchange platform for Zhejiang University postdoctoral fellows and promotes the development of academic cooperation and innovative research.

In the sharing, Fan Changjie, head of NetEase Fuxi, said that as the aging of the nation deepens and the demographic dividend disappears, labor shortages and high labor costs will continue to plague all walks of life. . Therefore, letting artificial intelligence (AI) and robots handle repetitive, high-risk tasks and improve human value has become an inevitable trend in the development of intelligent society in today's society. As the top artificial intelligence laboratory in China, NetEase Fuxi is committed to exploring a path with human-machine collaboration as the core to solve the practical problems faced by enterprises such as high AI threshold, long research and development cycle, and difficulty in application implementation.

The Aura Robot Platform is a new programming paradigm based on this background and designed around ideas such as AOP. Its core value is to provide a unified interface for connecting machines (or AI) and humans. Standardize interfaces and services. At present, many businesses of NetEase Fuxi have been connected to this theoretical method. Take the construction of key infrastructure projects in Southwest China as an example: On the one hand, the platform allows traditional excavators to achieve intelligent production through human-machine collaboration. On the other hand, through online distribution of crowdsourcing tasks, traditional excavator masters no longer engage in repetitive, boring and dangerous tasks. sex work. In addition to mining robot application scenarios, the platform has also been implemented in game character face-making, image and text recognition detection, smart parks, art crowdsourcing and other scenarios, verifying the strong implementation, sustainability and replicability of the human-machine collaboration platform. sex.

In this exchange, Dr. Ding Yu from NetEase’s Fuxi digital virtual human department introduced the latest progress of Fuxi’s cutting-edge research and practical applications in the virtual human expression field. Regarding the three aspects of AI expression synthesis, capture and binding, he shared the in-depth exploration and practice of Fuxi team and demonstrated the practical effects of applying these technologies in many NetEase games. Among them, the application of Fuxi AOP programming paradigm provides an effective solution for modeling the algorithm process of AI expression capture and binding. This work has been successfully applied to many NetEase games, such as "Ni Shui Han", "Tian Yu", "Meet Ni Shui Han" and "Ni Shui Han (Mobile Game)". The above work is dedicated to realizing the automatic generation and cross-character reuse of 3D expression parameters. In addition, Fuxi has published more than 20 top conference and journal papers in the field of expressions, and took the lead in proposing 135 classification standards for expressions, further promoting industry progress. Not only that, Fuxi has won the championship and runner-up of the Expression Perception Challenge (ABAW) for three consecutive times in the top conferences in the field of image vision (ICCV2021, CVPR2022-2023), winning a total of six championships and two runner-ups.

Dr. Li Lincheng, who works in the visual computing direction of NetEase Fuxi, mainly shared the intelligent exploration of 3D character and scene creation for games and Metaverse applications. The previous manual modeling process required a heavy workload and was difficult to meet the diverse, personalized, and customized needs of users. In this regard, Dr. Li introduced in detail NetEase Fuxi’s research work on 3D character and scene intelligent creation (3D AIGC). Through an intelligent creation solution based on 3D skeletons and meshes, users only need to input images or text to create real three-dimensional characters, replicate digital twins of real 3D scenes, and even generate ideal 3D scenes, fully satisfying personalization and Customize needs and optimize user experience. The application of these results provides unlimited possibilities for the construction of the digital world and data representation, and injects new vitality into the industry.

At the end of the exchange, NetEase Fuxi artificial intelligence expert Dr. Zhao Zeng shared the content about "NetEase Fuxi AIGC System Construction and Practice". He pointed out that generative artificial intelligence technology continues to mature, driven by natural language, and continues to refresh the public's understanding of its generative effects on text, images, audio and video, and 3D content. However, when generative artificial intelligence technology enters industrial scenarios from the research field, it still faces multiple challenges, including issues in multiple dimensions such as controllability, adaptability, and cost. By solving these challenges, generative artificial intelligence technology can be better applied in actual scenarios.

NetEase Fuxi's self-developed graphic and text model-Danqing

NetEase Fuxi began to build generative artificial intelligence technology in 2018, and received approval from the Zhejiang Provincial Department of Science and Technology Support key R&D projects. The scale of self-developed models has grown from the earliest 100 million parameters to 100 billion parameters. The model field has expanded from text to graphics, music, behavioral sequences and other modalities, and it has accumulated rich experience in pre-training model training and engineering optimization. Relevant technologies have also been applied in business scenarios such as word games, intelligent NPCs, text-assisted creation, music-assisted creation, art design, and Internet search recommendations within NetEase Group, and have achieved significant business results.

NetEase Fuxi, as the top institution in China specializing in the research and application of AI in games and pan-entertainment, will continue to develop in reinforcement learning, image actions, virtual humans, natural language, user portraits, large-scale Continue to explore areas such as pre-training models, and combine the crowdsourcing capabilities of intelligent robots to introduce "human" decision-making and data closed loops to further enhance the capabilities of intelligent agents.

The above is the detailed content of The eighth event of the Zhejiang University Postdoctoral Academic Salon: NetEase Fuxi discusses new areas of intelligence. For more information, please follow other related articles on the PHP Chinese website!