Litchi News (Shanghai Work Department reporter/Chang Ying, Hu Yiqiao, Warren Wang)
At the 2022 World Artificial Intelligence Conference, which opened on September 1st, the awarding ceremony of the highest award "SAIL Award" ("Outstanding Artificial Intelligence Leader") was held. The world’s first three-mode large model "Zidong Taichu" based on Ascending AI by Institute of Automation, Chinese Academy of Sciences has become one of the projects that won this honor.
Wang Jinqiao, a researcher at Institute of Automation, Chinese Academy of Sciences and president of Wuhan Institute of Artificial Intelligence, SHOWed the world’s first three-mode large model "Zidong Taichu" based on the basic software and hardware platform of Ascending AI in his speech on the show column. "Zidong Taichu" realized the unified representation and mutual generation of different modal data such as images, characters, voices and videos through cross-modal multi-task self-supervised learning, and formed a complete intelligent representation, reasoning and generation ability.
Dean Wang told the guests: "In traditional artificial intelligence learning, face recognition can be realized through visual models, but we don’t know how the machine represents everyone’s feature differences. We can only explain the process of machine learning from the feature map of the middle result of the image, and through the" Zidong Taichu "three-mode model, images and sounds can be unified across modes to the dimension of human language, which is closer to the way of human understanding and thinking."
Counting Four Breakthroughs of "Zidong Taichu"
1. Multi-task and multi-level cross-modal self-supervised learning.
A multi-task and multi-level training framework of cross-modal self-supervised learning is proposed, which supports the training of entry level, modal level and sample level, and realizes the unified modeling of cross-modal understanding and generation.
2. For the first time, "making sounds by pictures" and "making pictures by sounds" have become a reality.
"Zidong Taichu" is the first time to penetrate several kinds of information, such as voice, image and text, and has formed a complete ability of intelligent representation, reasoning and generation. It is the latest development trend in the field of data intelligence and provides an excellent platform for exploring the essence of human intelligence.
3. The first multi-modal pre-training model with 100 billion parameters.
"Zidong Taichu", as the world’s first three-mode large model with hundreds of billions of parameters, marks an important exploration of China’s path from weak artificial intelligence in limited fields to general artificial intelligence.
4. Breakthrough from "one specialization" to "multi-specialization and multi-function"
The performance of many algorithm indexes of "Zidong Taichu" ranks first in the world. Let AI move from "one specific function" to "multi-specialty and multi-function", and at the same time surpass the best performance of the industry in a number of downstream tasks, and build a fully autonomous artificial intelligence technology system.
In addition to showing the technical advantages of "Zidong Taichu" to the online and offline participants, Dean Wang also said that the basic model has been open source and the large model service has been opened, and a new version of the "Zidong Taichu" service platform will be opened in the near future, supporting the training, fine-tuning and deployment of low-code artificial intelligence models, automatically uploading data and automatically labeling models through the mode called by API, automatically forming a tool for reasoning and deployment, and automatically experiencing the effect.
