Since the development of deep learning, some “big model” achievements have appeared in many fields such as language, vision, recommendation, code generation, etc., which constantly refreshes people’s cognition and imagination of AI. Deep learning relies on the training of a large amount of data, and the “big model” has more parameters and more complex functions. Such features make the results calculated by the model more accurate. With the further development of the Internet of Everything world, the expansion of data volume and data collection are no longer a problem. The new proposition that follows is how to deal with massive data and make better training.

Back in 2017,The Transformer structure was proposed, which made the deep learning model parameters exceed 100 million; in 2018, the BERT network model was proposed, making the number of parameters exceed 300 million for the first time; in 2020, GPT-3 with 175 billion parameters was born; 2021 launched in The ZionEX system, which supports recommendation models with a size of more than 10 trillion

With the exponential growth of data size, large models have gradually been recognized as a bridge to cognitive intelligence through deep learning.

However, the explosion of data volume has raised a new proposition – how to overcome bottlenecks such as communication and improve the training efficiency of large models? In order to support the training of large models, a large-scale distributed training framework is often required to train large models.

In this regard, Huawei’s answer sheet is the MindSpore AI framework, which natively supports large model training. Synthetic MindSpore has the industry’s leading fully automatic parallelism capability, providing 6-dimensional hybrid parallel algorithms, namely data parallelism, model parallelism, pipeline parallelism, optimizer parallelism, etc.; the ultimate global memory reuse capability, without the developer’s perception It can automatically realize the multi-level storage optimization of NPU memory / CPU memory / NVMe hard disk storage, which greatly reduces the cost of model training; the minimalist ability to resume training at breakpoints can solve the problem of task interruption caused by large cluster training failures… Through these Features, which can solve the problems of memory usage, communication bottlenecks, complex debugging, and difficult deployment encountered in the development of large models.

Focusing on the underlying capabilities, Shengsi MindSpore joins handspartnerbuild fourinnovationModel

It is worth noting that the MindSpore AI framework of Ascension focuses on the construction of underlying capabilities and provides the industry with the foundation for building large models.So far, Ascension MindSpore AI has joined hands withindustryLeading Research Institute, Launch Coveragenaturelanguagedeal with,remote sensingimage,biomedical, multimodalThe four models are widely used in finance, medical care, agriculture and forestry,manufacturewait for eachindustry.

In May 2021, at the Huawei Ecological Conference 2021 “Rising Miles and Winning the New Era of Smart Intelligence”, Pengcheng Lab launched the world’s first 200 billion-parameter Chinese NLP model Pengcheng. Pangu, based on Ascension MindSpore. An AI large model of human Chinese comprehension. Pengcheng.Pangu model has a wide range of application scenarios, and has outstanding performance in the fields of text generation such as knowledge question answering, knowledge retrieval, knowledge reasoning, and reading comprehension.

Two months later, the Institute of Automation of the Chinese Academy of Sciences and Huawei jointly built a platform based on the Ascend AI and Ascend MindSpore AI frameworks.worldwideThe first trimodal large model –“Zidong. Taichu” was officially launched.Zidong. TaichuIt can achieve efficient collaboration among the three modalities of vision, text, and voice, and has world-leading performance. It is an important achievement on the road to exploring general artificial intelligence. It will be widely used in industrial quality inspection, film and television creation, Internet recommendation, intelligent driving and other fields.At the same time, relying on itsTechnological innovation and industry influence, Zidong.too early to get this yearWAICThe highest award – the Super AI Leader Award (SAIL Award for short)

In addition to the base large model,Synthetic MindSpore AI framework has also supported the launch of two large industry models –Pengcheng. ShennongheWuhan. LuoJia.

Pengcheng Lab and Huawei are based on Ascend AI and Ascend MindSpore AI frameworkspanArtificial Intelligence Platform for Biomedical Field “Pengcheng. Shennong”.Pharmaceutical companies and medical research institutions use the AI ​​capabilities provided by “Pengcheng Shennong” to greatly accelerate the screening and development of new drugs.researchsystem, and let artificial intelligence escort human health.

Wuhan UniversityTogether with Huawei’s Ascend AI team, they jointly built the world’s first dedicated framework for intelligent interpretation of remote sensing images embedded with the advanced technology features of Ascend MindSpore.Digital village construction, food security protection, urban planning and construction, national economy and people’s livelihood application empowerment

Build an experience platform and open up large model capabilities

Usually, the human and resource costs of training a large model are very high, which makes it difficult for ordinary developers to get started. In order to allow more developers to experience the charm of large models,The MindSpore community has created a one-stop large model experience platform, which was officially launched on July 30.

  • Shengsi large model experience platform:

The Shengsi large model experience platform not only integrates model selection, online reasoning, and online training, but also supports Gradio project visual reasoning and online transfer learning. Developers can query online the models and datasets built based on MindSpore, and select the large models and related tasks they are interested in, such as Pengcheng. Searching for pictures with sounds, generating sounds with pictures, and generating pictures with sounds, etc.

Real-life heroes, Shengsi AI Challenge is now open!

The best way to get a new skill is to set a small goal and practice it yourself.

In the field of developers, the mastery of the underlying theory is difficult to represent the actual development effect.In order to allow more developers to have the opportunityLearn from Ascension MindSpore, explore model algorithms and improve algorithm capabilities, thereby reserve talents for the industry, and promote the prosperity and development of artificial intelligence software and hardware application ecology. Ascension MindSpore specially holds the Ascension AI Challenge.

Ascension MindSpore also prepared generous prizes for the contestants:

  • 1 first prize with a bonus of 5K yuan and an official certificate of honor
  • 2 second prizes with a bonus of 3K yuan and an official honorary certificate;
  • 3 third prizes with a bonus of 2 K yuan and an official certificate of honor;
  • Teams that are shortlisted for the semi-finals, pass the code review and successfully reason online will receive certificates, customized gift packs and other prizes;
  • Any team that provides a reasoning module can receive a participation award;

This AI Challenge is an event for global AI developers. It has three major tracks: multi-category image classification, text classification, and artist style transfer, covering the basic fields of AI.

Among them, image classification is the most basic task in computer vision, and the algorithm of image classification is still developing rapidly. The purpose of this competition is to familiarize the participants with SENS MindSpore and to exercise the participants’ ability to use MindSpore for image classification preprocessing and image classification. At the same time, in order to examine the ability of the participants to deal with a large amount of data, this competition uses the Celtech multi-category image data set.

Text classification research can realize various theories and methods for effective communication between humans and computers using natural language. Text classification has assumed an important role in the field of AI. The purpose of this competition is to familiarize the participants with SENS MindSpore and to exercise the participants’ ability to use MindSpore for NLP text processing and text classification. This competition uses the Amazon Review dataset, and the contestants need to predict the user’s rating (integer of 1-5 points) based on the user’s review text.

Artist style transfer track The development of image style transfer technology plays an invaluable role in image processing, computer vision, film and television production and other fields. The purpose of this question is to familiarize the participants with SENS MindSpore and to exercise the participants’ ability to use MindSpore for image style transfer. This competition uses Van Gogh paintings as the target style for style transfer.

Through the settings of the three major tracks, developers can freely choose the tracks of interest, realize the leap from theory to practice, understand the latest talent needs of the industry, and improve their own skills.

60,000 yuan bonus and many peripheral benefits, what are you waiting for? The registration channel has been opened, click the link to choose the track you like to register!

#wind #large #models #underlying #framework #land #News Fast Delivery

Leave a Comment

Your email address will not be published. Required fields are marked *