Shulou(Shulou.com)11/24 Report--

Thanks to CTOnews.com netizens who want to laugh, soft media new friends 1985234, carefree 1135, Mr. Aviation's clue delivery! CTOnews.com, June 4, according to titanium media, Huawei will release a multimodal 100 billion-level large model product called Pangu Chat, which is directly calibrated to ChatGPT.

According to reports, the Pangu model was successfully established in Huawei Cloud in November 2020. The "Pangu Chat" is expected to be released and tested internally at Huawei Cloud developer Conference (HDC.Cloud 2023) on July 7 this year, and the product is mainly aimed at To B / G government and enterprise customers.

According to data from a paper published by Huawei, the parameters of Huawei's Pangu PanGu- Σ large model are at most 1.085 trillion, based on the MindSpore framework developed by Huawei itself. On the whole, the PanGu- Σ large model may be close to the level of GPT-3.5 in terms of dialogue.

According to CTOnews.com, the Huawei Pangu model was officially released in April 2021, and then upgraded to version 2.0 in April 2022. At present, the AI large model NLP large model, CV large model and scientific computing large model (meteorological large model) have been marked as coming online.

According to reports, this is the first Chinese pre-training model with 100 billion parameters, while the CV model reaches 3 billion parameters for the first time. Pangu CV large model is the largest CV large model in the industry, the first time to achieve both discrimination and generation ability, small sample learning ability in the industry first; Pangu meteorological large model provides second weather forecast; Zidong. Taichu is the first picture, text and sound three-mode model in the world.

For the positioning of the Pangu model, the Huawei internal team established three key core design principles: first, the model should be large, which can absorb massive data; second, the network structure should be strong, which can really give play to the performance of the model; third, it should have excellent generalization ability and can really land on the work scene of various industries.

According to the PPT information of Huawei Cloud executives, at present, the basic layer of Huawei's "Pangu series AI model" mainly includes NLP model, CV model, and scientific computing model, while the upper layer is the Huawei industry model developed with partners.

According to Huawei's cloud official website, Pangu model is composed of NLP model, CV model, multimodal model, scientific computing model and other large models. Through model generalization, it can solve the problems of AI scale and industrialization that can not be solved under the traditional AI workshop development mode, and can support a variety of natural language processing tasks, including text generation, text classification, question and answer system, and so on.

Specifically, the Pangu NLP model uses the Encoder-Decoder architecture for the first time, taking into account the understanding and generation ability of the NLP model, which ensures the flexibility of the model embedded in different systems. In downstream applications, only a small number of samples and learnable parameters can complete the rapid fine-tuning and downstream adaptation of hundreds of billions of large-scale models. This model has a good performance in intelligent public opinion and intelligent marketing.

Pangu NLP model Pangu CV model is the largest CV model in the industry for the first time to realize model extraction on demand, taking into account both discrimination and generation ability for the first time. Based on the requirements of model size and running speed, different scale models are adaptively extracted, and AI application development is implemented quickly. By using hierarchical semantic alignment and semantic adjustment algorithm, better separability is obtained in shallow features, and the learning ability of small samples is significantly improved, reaching the first in the industry. this model has a good performance in intelligent inspection and intelligent logistics.

Pangu CV model Pangu meteorological model provides second weather forecast. With the help of innovative 3DEST network structure and hierarchical time aggregation algorithm, the accuracy of weather forecast is higher than that of the most advanced methods in terms of key elements and common time range, and the speed is more than 1000 times higher than that of traditional methods. At the same time, the Pangu meteorological model supports a wide range of downstream forecasting schemes, such as typhoon track prediction tasks, compared with the traditional numerical weather forecasting methods, the Pangu meteorological model can reduce the position error by more than 20%.

According to the information previously disclosed by Zeshang Securities, the Pangu meteorological model, Huawei used more than 2000 Pangu 910 chips when training the Pangu model with 100 billion parameters, and conducted more than two months of data training. According to Huawei, the annual large model training calls more than 4000 GPU / TPU cards, and the calculation cost of the three-year large model is as high as 960 million yuan.

According to the industry chain research report of Huawei Pangu model combed by Soochow Securities, the advantage of Huawei Pangu model lies in its independent control of talent reserve and computing power, so it is expected to become a leading large model in China, and its eco-industrial chain is expected to accelerate its development. including Tuo Wei Information, Sichuan Changhong, Kirin Software (Chinese software), Tongxin Software (Chengmai Technology), Kirin Xinan and other Huawei ecological companies. Guosheng Securities believes that Huawei Pangu is the first multi-modal 100 billion-level model, which is expected to endow all industries.

▲ source: Soochow Securities

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.