It has been 13 years since Amazon released its first cloud service S3 in 2006. In recent years, with the rapid development of AI technology, the combination of cloud and AI is releasing new computing and intelligent dividends.

This is also the signal of Huawei's intelligent computing conference last week. At the conference, Hong Fangming, president of Huawei Cloud China, introduced Huawei's exploration in "Cloud+AI" and a series of progress. Hong Fangming revealed that in the past six months alone, the number of Huawei cloud customers has increased 33 times, and more than 1.7 million enterprise customers and developers have chosen Huawei Cloud.

In addition, according to public data, among the latest figures released by IDC, Huawei Cloud has entered the top five of China's cloud computing market with revenue exceeding growth by more than 300% and PaaS market share growing by nearly 700%. It is also the fastest growing cloud computing company in the global market.

The above achievements fully demonstrate the correctness of Huawei's "cloud+AI" strategy. Since the establishment of the AI strategy in 2018, Huawei has built an AI infrastructure from chip to server based on the Kunpeng ecology. The Shengteng 910 chip released recently leads the industry in both computing power and power consumption. At the same time, Huawei has also continued to build AI solutions and products with full stack and full scenario to meet the diversified needs of enterprises and developers for AI.

Among a series of AI products of Huawei Cloud, the importance of the AI development platform ModelArts is self-evident. On the one hand, ModelArts connects with AI development frameworks downward, such as Tensoflow and MindSpot, which Huawei is about to open source, and cooperates with AI general technology and industry solutions upward; On the other hand, as a one-stop AI development platform, the product experience, technology and integration capabilities of ModelArts, in part, determine the recognition of AI developers for Huawei's cloud AI capabilities.

What is the technical capability of Huawei Cloud ModelArts? In the growing global demand for cloud based machine learning and development, can ModelArts continue to promote Huawei's "cloud+AI" strategy and seize the commanding heights in the future "cloud+AI+5G" technology trend?

Before answering these questions, it is advisable to take a look at the current situation of machine learning development and the challenges and opportunities it faces from an industry perspective.

Pain Points of Machine Learning Development

Compared with the mature software development in the past decades, machine learning development has extremely distinctive characteristics. For example, it needs massive data as support. For most small and medium-sized enterprises and individual developers, multiple open source data sets in the industry may meet the basic needs.

Secondly, machine learning development has a great demand for computing power. If we use the GPU training model locally, let alone how expensive it is, the current speed of development of machine learning model computing power has already exceeded the industry rule of Moore's Law of "doubling in 18 months". According to the survey data of OpenAI in 2018, since 2012, The computing power demand for machine learning training will double every 3.43 months on average.

This means that using the local GPU to train machine learning models is a "bottomless hole" with high costs and low training efficiency. The rest of the options can only rely on the cloud, which is why Google, Amazon and Alibaba have launched cloud based machine learning platforms in the past few years.

But facing so many platform choices, enterprises and developers have to face a series of new problems. For example, how to solve the delay problem of machine learning model training and deployment when Google and other overseas cloud services have not landed in China? For another example, the general environment of AI development and entrepreneurship between China and the United States is very different. To what extent can the AI capabilities of these overseas cloud services meet the development needs of Chinese developers is a huge problem.

More importantly, machine learning development is a complex process, which involves multiple stages such as data collection, data annotation, model training, algorithm tuning and optimization, and model deployment. Different stages have different computing needs. For example, data scientists and machine learning researchers complete model construction and training, while software engineers Machine learning engineers and data engineers complete it. For another example, model training is usually completed by multiple people on multiple virtual servers, while the deployment model needs to be extensible and able to handle massive API requests.

The above points constitute several pain points in the current machine learning development field, which are opportunities for all cloud based machine learning platforms. Next, let's see how Huawei Cloud ModelArts can solve these pain points, so as to realize the vision of one-stop development of machine learning.

Three advantages enable machine learning development

The emergence of cloud computing has changed the billing model of the past technology infrastructure. Pay as you go or pay on time has become the most prominent business feature in the cloud era. As mentioned above, machine learning requires a lot of computing power, so how to shorten the time for cloud machine learning and speed up the process of model training and model deployment has become a difficult problem for all cloud machine learning platforms, including Huawei Cloud ModelArts.

Huawei Cloud engineers solved this problem through technological breakthroughs. In March 2019, Huawei Cloud ModelArts won the double titles of image recognition training and reasoning performance in the latest DAWN Bench list released by Stanford University, significantly reducing the model training time while achieving super reasoning performance.

Looking at a set of specific figures, in terms of training performance, the test results on ResNet50_on_ImageNet show that when 128 V100 are used, the model training time on Huawei Cloud ModelArts is only 4:08 seconds, twice as fast as the record of 9:22 seconds set in December 2018, and four times faster than the previous fast.ai training speed on the AWS platform; In terms of reasoning performance, Huawei Cloud ModelArts recognizes images 1.7 times faster than the second manufacturer, 4 times faster than Amazon, and 9.1 times faster than Google.

DAWNBench of Stanford University is one of the most authoritative competitions in the field of global artificial intelligence. It is an international authoritative benchmark test platform used to measure the end-to-end deep learning model training and reasoning performance, Huawei Cloud ModelArts has made such achievements, demonstrating its ability to optimize technology on the machine learning platform, reducing the use cost of the machine learning platform through technological innovation, and ultimately transferring the technical dividend to enterprises and developers 。

On the other hand, with the rapid development of AI, the technical threshold of AI must be lowered, especially for the training and deployment of machine learning models. Huawei Cloud ModelArts also practices Huawei's concept of "leaving complexity to oneself, and bringing simplicity to a customer". It has built-in automatic (machine) learning features. It realizes automatic selection of model training parameters and automatic tuning of models through algorithms, so that business developers with zero AI foundation can quickly complete the training and deployment of models, Even zero code development AI models can be implemented in some scenarios.

Secondly, if technological innovation reduces the technical threshold and use cost for developers, then Platform capability of ModelArts in AI whole process development , which solves many problems of developers in machine learning lengthy processes.

We have to mention the "past life" of ModelArts. As a product derived from Huawei, ModelArts is also a concentrated demonstration of Huawei's internal AI development capabilities. Huawei has many internal algorithm engineers, AI developers, AI engineers, etc. They can understand the key points of data annotation and preparation, model training, model optimization, model deployment and other processes in the AI development process. Therefore, the products finally provided to enterprises and developers also have one-stop AI development capabilities.

For example, in order to solve the industry problem of time-consuming and labor-intensive data annotation in machine learning, ModelArts uses the built-in AI data framework to govern the data with the AI mechanism, and then solves the problem of annotated data volume through iterative training, which can improve the efficiency of data annotation 100 times in scenarios with large data volume.

In the model training and deployment phase, in addition to shortening the deployment time mentioned above, ModelArts has realized one click push model to all edge and end devices. Deployment on the cloud also supports online and batch reasoning, which can meet the needs of multiple scenarios such as large concurrency and distribution.

Furthermore, ModelArts also has a visual management of the whole process, which can help developers quickly understand the progress of model training. ModelArts provides visual management of the whole process from data, algorithms, training, models, and services. Viewing this process through any object really achieves the visualization of workflow.

At the same time, developers can also get a visual model evaluation report after model training by using the confusion matrix and thermodynamic diagram to help enterprises and developers quickly evaluate models or optimize models.

Thirdly, ModelArts also has its own thinking on ecological enabling from technological innovation to the whole process of products. AI market is one example , which is a developer ecological community based on ModelArts, providing content sharing functions such as AI models, API transactions, data, competition cases, etc.

In this market, whether scientific research institutions, AI application developers, solution integrators, or enterprises of different types in different industries can quickly find technology or business opportunities that meet their own needs, effectively connect all participants in the AI development ecological chain, and accelerate the development and implementation of AI products, It also protects the commercial interests of all participants in the AI development ecological chain.

It is precisely in terms of technology, platform capabilities and ecological capabilities that Huawei Cloud ModelArts has quickly gained a large number of "fans" since its release, which has been applied in medical, intelligent manufacturing, autonomous driving, smart cities, buildings, parks and other scenarios, helping many enterprises and developers, including Jinyu Medicine, Guanglianda, Yunlu Technology, speed up AI training and deployment, Push these fields onto the fast track of AI.

At the end: the transformation of cloud+AI+5G has just begun

In the past few years, the industry generally believed that AI was a general-purpose technology. This means that all industries are likely to be restructured by the technology and ideas brought by AI, which is a new opportunity brought by technological progress.

After Huawei established the AI strategy of the whole stack and the whole scene in 2018, it has created a series of bottom products to platform capabilities focusing on AI, gradually covering the complete AI technology architecture such as AI chips, AI frameworks, AI development platforms, etc. Together with Huawei's technology and product advantages in the cloud and 5G, it enables all walks of life and promotes digital and intelligent changes in the industry. The latest news is that Huawei will take on the task of building the only national AI development and innovation platform for basic software and hardware, which is also the recognition of Huawei's AI technology capability at the national level.

As a development platform connecting AI infrastructure and AI basic capabilities, ModelArts also shoulders the mission of enabling AI development processes and ecological construction in various industries. At the moment when 5G technology dividends are gradually emerging, the technology, platform capabilities and ecological construction of ModelArts have become the pioneer force of Huawei Cloud's rapid development in the AI field. When ordinary users can train a model at a very low cost, when enterprise developers can complete the training and deployment of complex models in a one-stop manner, and when developers and enterprises realize the commercial value of AI development, it is appropriate to choose Huawei Cloud AI development.

On this track with clear goals, Huawei Cloud's ModelArts is running fast. (End)