Huawei Rotating Chairman Eric Xu announced Huawei’s AI strategy and its full-stack, all-scenario AI portfolio in his keynote speech at HUAWEI CONNECT 2018. He also announced the release of the company’s Ascend series of AI chips – the world’s first AI IP and chip series designed for a full range of scenarios.
Below is the full text of Xu’s keynote speech:
In 1956, the concept of artificial intelligence was proposed at the Dartmouth Workshop organized by John McCarthy, an assistant professor of mathematics at Dartmouth College. That was more than 60 years ago. Since that time, we’ve seen two AI winters, when funding and interest in AI research reduced sharply. Despite these setbacks, AI has never stopped advancing.
In 1971, Intel introduced the first microprocessor. That was almost 50 years ago, and Moore’s law has continued to hold true throughout the robust development of the ICT industry.
If we overlay graphs of AI and ICT development cycles over the past 60 years, it’s clear that advances in AI and ICT are closely related. Academic research findings and engineering advances go hand in hand.
In the past, we went through two AI winters because expectations for AI were way beyond the engineering capacity of the ICT industry at the time. The good news is that each winter eventually gave way to spring, marking a new beginning for AI.
Today, we again find ourselves in a season of harvest, one made possible through six decades of commitment and collaboration between academic and industry stakeholders in ICT domains.
Moving forward, we need to make full use of AI technology. We need to start reaping the benefits sooner rather than later, and work hard to expand its value. We have to do what we can to extend this harvest season. Let’s plant AI along the equator, so to speak, to create an environment where it can continue to blossom and thrive.
We can’t fully unleash the value of a technology unless we properly define its positioning. This is essential for us to truly understand and adopt AI technology.
At Huawei we recognize AI as a combination of technologies that, together, form a new general purpose technology (GPT). We have seen the rise of many general purpose technologies before AI. The wheel and the iron, which are both several thousand years old. Railways and electricity in the 19th century. And automobiles, computers, and the Internet in the 20th century.
In Economic Transformations: General Purpose Technologies and Long-Term Economic Growth, the Canadian academic Richard G. Lipsey noted that new GPTs are the drivers of sustainable socioeconomic growth. A GPT must have multiple uses across the economy, according to Lipsey, and have many technological complementarities, meaning when two or more different technologies strengthen and reinforce each other, and spillovers.
Economists have observed that, throughout human history, there have been 26 technologies that can be classified as true GPTs. AI is one of them.
By emphasizing AI’s role as a general purpose technology, my goal is to call your attention to how influential and valuable AI is to our future. As a GPT, AI will help us find more efficient solutions to problems we already know how to fix. It will also help us address problems that, to date, have remained unsolved.
As companies, if we want to stay ahead we need to adopt an AI mindset – use AI concepts and technologies to tackle both existing and future problems.
Inside Huawei, our experience shows that AI can replace humans in certain tasks, and also automate cost reduction throughout production cycles. This is the most valuable characteristic of AI, and what sets it apart from run-of-the-mill informatization, which can’t automatically reduce production costs.
AI will change all industries. We need to ask, “How will AI reshape or even disrupt the industry I’m working in?” Going forward, we need to think of new ways to prepare our businesses and industries for change.
The list goes on and on.
In just the past year since we launched Huawei Cloud EI and HiAI, we’ve already seen AI drive unprecedented momentum across all kinds of industries.
We’ve seen several technological revolutions since the 18th century. Each has had a huge impact on organizational structures, processes, and workforce skills. But AI will change jobs and skills in a way that’s quite different from the previous revolutions.
Previous revolutions led to huge demand for repetitive routine tasks, such as operating equipment in textile mills, and running car and phone assembly lines. AI will greatly boost automation in almost all aspects of an organization. This means there will be much less demand for jobs that handle repetitive, routine tasks.
Demand for data science jobs will keep rising, including those for data scientists and data science engineers with basic know-how in data science. The total number of these jobs will be much smaller than the number of jobs that handle repetitive, routine tasks.
It’s likely that organizations will become more diamond-shaped, with AI systems taking the place of the people at the bottom, where they handle huge volumes of repetitive and routine tasks.
Change can mean good news for some and bad news for others, especially when the changes first start to emerge.
Some people might get excited about the new, once-unimaginable functions that AI will make possible. These people will feel a strong urge to drive large-scale AI adoption. And there will also be those who feel anxious about underperforming AI projects, or who worry about the reliability and security of AI applications. These are the ones who will remain uncertain about how to best use AI in the future.
If we look at the history of all GPTs, these reactions to AI are all very natural.
There are four different phases along the AI productivity/adoption curve. We have just left the first phase, where exploration of AI technology and application takes place on a small scale.
Now we’re in the second phase, where new technology and society are colliding. From a tech perspective, more issues are emerging as AI technology continues to advance. If we look at things from an application perspective, however, the value of AI is seeing greater recognition as it comes into wider use.
That said, existing policies, corporate processes, and workforces are built around older technologies, such as those in the information and Internet eras. The broader social environment isn’t yet ready for the AI era. So in this phase we see a certain degree of collision – even conflict – between tech development and society.
However, AI will ultimately find itself in a social environment that’s more conducive to its development and application. When that happens, we will step into the third phase, where we’ll see rapid, comprehensive advances in AI adoption and productivity.
The fourth phase will be the golden era of AI, where humanity will benefit from a constant stream of new advances in technology until a new GPT emerges. Nevertheless, it’s important to keep in mind that AI isn’t a cure-all. No technology can solve every problem. We need to focus on areas where AI can create the most value, not on problems that AI isn’t equipped to solve. Finding the right problem is more important than devising a novel solution.
To get started, we need to take a look at where we are today with AI.
The world has seen significant achievements:
Despite these incredible achievements, we’ve also seen quite a few smaller figures that speak to lukewarm AI adoption in its early stages. For example:
The gaps between stellar achievements and lukewarm adoption are the driving forces that will push the industry forward. I find these gaps to be very inspiring.
To close these gaps, we need the right technology, the right talent, and the right industry ecosystem. Next, I’d like to discuss ten important changes that we have to work on together across all three of these elements.
With existing technology, training more complex models often takes days, if not months. Successful innovation only happens after several rounds of iteration. Slow model training seriously impedes application innovation. We believe that training should be completed in minutes or even seconds.
Computing power is the foundation of AI. Right now, it’s a costly and scarce resource. While growth in computing power has been a major driver behind progress in AI, a lack of readily available and affordable computing power is becoming a constraint that holds back broad-scale AI adoption.
We need to provide more abundant and affordable computing power in the future. We should take action now to meet this demand.
Hybrid clouds have become a major cloud service model for enterprise use. Right now, AI is deployed mostly in the cloud, with only a small portion at the edge. AI hasn’t yet been closely integrated into business environments.
AI should be pervasive. Furthermore, it should be adaptable to all scenarios, and in all cases, user privacy must be respected and protected.
Algorithms are another driver behind AI development. The majority of the basic algorithms we use today were invented before the 1980s. As AI comes into wider use, the weaknesses of existing algorithms are becoming more apparent.
Algorithms of the future should be data-efficient. That means they can deliver the same results with less data. Future algorithms should also be energy-efficient, producing the same results with less compute and less energy. Algorithms must be secure and explainable. Algorithms like these will set the stage for wide-scale AI development.
At present, AI projects are labor-intensive, especially during the data labeling process. This requires so much labor, in fact, that specialized “data labeler” jobs have begun to emerge. There’s even a running joke in the industry: “no labor, no intelligence”.
Moving forward, we must greatly increase AI automation to achieve automated or semi-automated operations, especially during processes like data labeling, data collection, feature extraction, model design, and training.
In June 2018, Benjamin Recht, an associate professor at UC Berkeley, released a paper with a perplexing title: “Do CIFAR-10 Classifiers Generalize to CIFAR-10?” According to the paper, models that perform with high accuracy in one test set of CIFAR-10 classifiers are 5 percent to 15 percent less accurate in another test set that closely resembles CIFAR-10, which Recht himself developed. This means a large drop in the practical application of a given model.
It’s clear that many high-performing models and algorithms perform better in tests than in real-world execution.
Industrial-grade AI models of the future must be able to meet the needs of real-world execution. It’s not enough to perform well in test sets alone.
The accuracy of any given model shouldn’t be static, as accuracy changes with data distribution, application environments, and hardware environments. Keeping accuracy numbers within an acceptable scope is necessary for enterprise applications. Existing model updates, however, are not done in real time. They rely on human input at fixed intervals. It’s a semi-open loop system.
We believe that the models of the future need to be adaptive to changes and updated in real time. This represents a real-time, closed-loop system that helps enterprise AI applications continue to operate in an optimal state.
Every GPT delivers maximum economic value only when it’s combined with other technologies. AI is no exception. But current discussions on AI more often than not focus entirely on AI, with no mention of other technologies.
In the future, we need to promote greater synergy between AI and other technologies, including cloud, IoT, edge computing, Blockchain, big data, and databases. This is the only way to fully unleash the value of AI.
At present, AI is a job that can only be done by highly skilled experts. There aren’t enough mature, stable, and extensive automation tools. Producing AI models is complex work that takes a lot of time and effort.
Moving forward, we need a one-stop platform that provides the necessary automation tools, making it easier and faster to develop AI applications. When this platform is in place, AI will become a basic skill of all application developers, even all ICT workers.
Lack of AI talent, especially data scientists, has long been seen as a major obstacle to AI progress. Data scientists are scarce and will remain so in the future.
Addressing this challenge requires an AI mindset. That means providing intelligent, automated, and easy-to-use AI platforms, tools, services, and training and education programs to foster a huge number of data science engineers. These people must be equipped with the ability to deal with massive volumes of basic data science tasks.
The AI workforce will be organized in a pyramid-like structure, with a large number of data science engineers working with data scientists and subject matter experts. This is how we can help resolve the scarcity of AI talent.
These ten changes don’t represent the full picture of AI technology, talent, and industry development. But if we can drive these changes, they will lay a solid foundation for future AI growth.
These ten changes are what Huawei expects to see in the AI industry. They are also the inspiration behind Huawei’s AI strategy.
To drive these ten changes, our AI strategy includes the following five priorities:
Invest in AI research: Develop basic capabilities in data and power-efficiency, for example, using less data, computing resources, and power; build secure and trusted platforms; and develop automated and autonomous machine learning for computer vision, natural language processing, decision and inference, and so on.
Build a full-stack AI portfolio:
Develop an open ecosystem and talent: Collaborate widely with global academia, industries, and partners.
Strengthen our existing portfolio: Bring an AI mindset and techniques into existing products and solutions to create greater value and enhance competitive strengths.
Drive operational efficiency at Huawei: Apply AI to massive volumes of routine business activities for better efficiency and quality.
By “all-scenario”, we mean different deployment scenarios for AI, including public clouds, private clouds, edge computing in all forms, industrial IoT devices, and consumer devices.
“Full stack” is about the functionality of our technology. Our full-stack portfolio includes chips, chip enablement, a training and inference framework, and application enablement.
Specifically, our full-stack portfolio includes the following:
Ascend: An AI IP and chip series based on a unified, scalable architecture. In this series, we have Ascend Max, Mini, Lite, Tiny, and Nano. Ascend 910 has the world’s greatest computing density in a single chip. Ascend 310 is the most efficient AI SoC for low-power computing.
CANN (Compute Architecture for Neural Networks): A chip operators library and highly automated operators development toolkit
MindSpore: A unified training and inference framework for device, edge, and cloud (both standalone and cooperative)
Application enablement: Full-pipeline services (ModelArts), hierarchical APIs, and pre-integrated solutions
In September 2017, we released Huawei Cloud EI, an AI service platform for enterprises and governments.
In April 2018, Huawei announced HiAI, our AI engine for smart devices.
Our full-stack, all-scenario AI portfolio is designed to provide powerful support for Huawei Cloud EI and HiAI.
Backed by our AI portfolio, Huawei Cloud EI will be able to deliver a full-stack portfolio for enterprise and government customers, and HiAI will provide a full-stack portfolio for smart devices. HiAI services are deployed on Huawei Cloud EI.
To sum up, our AI strategy is to invest in basic research and talent development, build a full-stack, all-scenario AI portfolio, and foster an open global ecosystem.
Within Huawei, we will continue to explore to improve management and efficiency with AI.
In the telecom sector, we aim to adopt SoftCOM AI to make network O&M more efficient.
In the consumer market, HiAI will infuse intelligence into consumer devices, making them smarter than ever.
Huawei Cloud EI public cloud services and FusionMind private cloud solutions will provide abundant and affordable computing power for all organizations – especially businesses and governments – and help them use AI with greater ease.
Our portfolio will also include AI acceleration card, AI server, AI appliance, and many other products.
“All-scenario” means that Huawei is able to deliver pervasive intelligence for a fully connected, intelligent world.
“Full stack” means that Huawei is able to provide AI application developers with unparalleled computing power and a strong application development platform. We are working towards making AI more inclusive – making it affordable, effective, and reliable for all. And we have what it takes to make that happen.
I hope we can all work together to turn AI into a practical reality, making it inclusive and available for all. Huawei is committed to working closely with our customers, partners, and academia to grow together, promote pervasive AI, and ultimately build a fully connected, intelligent world.