AI pioneer Wu Enda received an exclusive interview and talked about his outlook on the general trend of AI in the next 10 years. He believes that in the future, the focus of technology implementation will shift from hardware to data, forming a "data centric" AI. Have you ever felt that you have had enough of your current job and want to change direction? If you have, you are definitely not alone. However, in addition to participating in the dictionary, there are some less radical methods, such as Wu Enda's method. Wu Enda is one of the most outstanding figures in the field of artificial intelligence. He is landing AI and deep learning Founder of AI, co-chairman and co-founder of coursera and adjunct professor at Stanford University. Previously, he was also the chief scientist of Baidu and one of the founders of Google brain project. However, according to him, his current focus has shifted from the digital world to the real world, which is the so-called "from bits to things". In 2017, Wu Enda founded landing AI, a start-up company dedicated to promoting the application of artificial intelligence in manufacturing. We interviewed Wu Enda and discussed what he called "the data centric approach of AI" and its relationship with his work in landing AI and the general background of AI today. From digitization to landing Wu Enda said that his motivation is industry-oriented. He believes that manufacturing is "one of the great industries that have a great impact on everyone's life, but it is so invisible to many of us." Many countries, including the United States, are saddened by the decline of manufacturing. Wu Enda hopes to "adopt AI technology that has changed Internet enterprises and use it to help people working in manufacturing." This is a growing trend. According to a 2021 survey, 65% of manufacturing leaders are trying to pilot AI. It is expected to achieve a compound annual growth rate of 57.2% in the next five years. Although AI is increasingly used in manufacturing, this process is much more difficult than Wu Enda imagined. He admitted that when landing AI started, it mainly focused on consulting. However, after participating in many customer projects, Wu Enda and landing AI developed a new toolkit and game manual to make AI play a role in manufacturing and industrial automation. Landing lens is committed to enabling customers in the field of manufacturing and industrial automation to quickly and easily establish and deploy visual inspection systems. Wu Xiaobo had to adjust his work on consumer software to focus on artificial intelligence in manufacturing. For example, computer vision driven by artificial intelligence can help manufacturers complete tasks such as identifying defects in production lines. But it's not easy, he explained. "In consumer software, you can build a single AI system to provide services to 100 million or one billion users, and really get a lot of value in this way, but in the manufacturing industry, what each factory makes is different. Therefore, each manufacturer needs a customized AI system to train according to their own data." Wu Enda said that the challenge faced by many companies in the AI field is how to help 10000 manufacturing plants establish 10000 customer systems. The data centric approach believes that AI has reached the point where data is more important than models. If AI is regarded as a system with moving parts, it should keep the model relatively fixed and focus on high-quality data to fine tune the model, rather than continue to promote the marginal improvement of the model. Not many people have this idea. Chris R é, who leads the hazy research group at Stanford University, is another advocate of a data centric approach. Of course, as mentioned earlier, the importance of data is not new. There are mature mathematics, algorithms and system technologies to process data, which have been developed for decades. However, how to establish and re-examine these technologies on the basis of modern AI models and methods is the new requirement. Just a few years ago, we didn't have a long-lived AI system or a powerful depth model on this scale. Wu Enda pointed out that since he started talking about data centric AI in March 2021, the response he got reminded him of the scene when he and others began to discuss deep learning about 15 years ago. Wu Enda said, "today people's reaction is:" I always know this, there is nothing new "to" this can't succeed "." but some people say 'yes, I always think this industry needs this thing, this is a great direction'. " Data centric AI and basic model If artificial intelligence with data as the core is the right direction, how should it operate in the real world? Wu Enda pointed out that it is unrealistic to expect institutions to train their own customized AI models. The only way out of this dilemma is to design a tool to enable customers to design their own models, collect data and express their knowledge in their respective fields. Wu Enda and landing AI will achieve this through landing lens, giving experts in various fields the ability to convey knowledge through data markers. Wu Enda pointed out that in the field of production, there is generally no large amount of data for reference. For example, if the goal is to identify the wrong products, a fairly good production line will not have so many waste pictures to refer to. In the field of production, sometimes there are only 50 pictures in the world for reference. This is simply not enough for existing AI. This is why the focus should now shift to allowing experts to record their knowledge by collecting data. Wu Enda said that the landing AI platform is doing this. The platform can help users find the most useful cases to build the most consistent tags, and improve the quality of images and tags input into the algorithm. The key here is "consistency". Wu Enda and others before him have found that expertise cannot be defined by a single expert. What is defective to one expert may be valued by another expert. This phenomenon is not unique, but it will surface only when you have to generate data sets with the same annotation. Wu Enda said, "That's why we need good tools and workflow to enable experts to reach agreement quickly. There is no need to spend time where consensus has been reached. On the contrary, our goal is to focus on the parts where experts have not reached agreement, so that they can solve the defective parts through discussion. It has proved that it is very important to achieve the consistency of the whole data if we want the AI system to achieve good performance quickly." This method is not only meaningful, but also has some similarities. The process described by Wu Enda obviously deviates from the method of "putting more data" often used by AI today, but points more to the method based on management, metadata and semantic coordination. In fact, people like David Talbot, a former head of machine translation at Google, have been conveying the idea that in addition to learning from data, applying knowledge in various fields also makes sense for machine translation. In the application of machine translation and natural language processing (NLP), the knowledge in the field refers to linguistics. We have now reached a new stage. We have the so-called NLP basic model: for example, a huge model like gpt3. After a lot of data training, people can use these models to fine tune for specific applications or fields. However, this kind of NLP basic model does not really use the knowledge of various fields. Can the basic model of computer vision do this? If so, how and when can we achieve it? What will the implementation bring? According to Wu Enda, the basic model is both a scale problem and a traditional problem. He believes that this can be achieved because many research groups are trying to establish the basic model of computer vision. Wu Enda said, "this does not mean that it is not the basic model the first day, but the next day. In the case of NLP, we see that the model is developing, from Google's Bert model, transformer model, gpt2 to gpt3. This is a series of models with larger and larger scale, which are trained on more and more data, and then some of them are called basic models. "I'm sure we'll see something similar in computer vision. Many people have been pre training on Imagenet for many years. I think the trend will gradually be pre training on larger and larger data sets, more and more pre training on unmarked data sets, and more and more pre training on video," Wu said The next 10 years of AI As an insider of computer vision, Wu Enda is well aware of the steady progress being made in artificial intelligence. He believes that at some time in the future, the media and the public will announce that the computer vision model belongs to the basic model. However, whether it can accurately predict when it will come true is another matter. For applications with large amounts of data, such as NLP, the amount of domain knowledge input into the system decreases over time. Wu Enda explained that in the early days of deep learning (including computer vision and NLP), people usually train a small deep learning model and then combine it with more traditional methods of domain knowledge base, because the effect of deep learning is not good. However, as the scale of the model becomes larger and larger, there are more and more data, and less knowledge in various fields is injected. According to Wu Enda, people tend to think that a large amount of data is a learning algorithm. This is why machine translation has finally proved that the end-to-end purity of learning methods can perform well. But this only applies to problems that require learning a lot of data. Domain knowledge does become important when you have a relatively small data set. Wu Enda believes that AI systems provide two sources of knowledge - data and human experience. When we have a lot of data, AI will rely more on data than human knowledge. However, in areas where data is scarce, such as manufacturing, we can only rely on human knowledge. The technical approach is to build tools that allow experts to express their knowledge. This seems to point to methods such as robust artificial intelligence, hybrid artificial intelligence or neural symbolic artificial intelligence, as well as technologies such as knowledge atlas for expressing domain knowledge. However, although Wu Enda knows these technologies and finds them interesting, landing AI does not cooperate with them. Wu Enda also found that the so-called multimodal AI or the combination of different forms of input (such as text and image) has development prospects. The first mock exam in the past ten years is to build and perfect single modal algorithms. Now that the artificial intelligence community has become larger and made progress, it is meaningful to pursue this direction. Although Wu Enda was one of the first people to use GPU for machine learning, now he doesn't pay much attention to hardware. Although it is a good thing to have a booming artificial intelligence chip ecosystem, including established enterprises such as NVIDIA, AMD and Intel, as well as upstarts with novel architectures, this is not the end. In the past decade, artificial intelligence