Thoughts on the Era of Artificial IntelligenceDec 21, 2018
There remain many different views about how Artificial Intelligence (AI) will be applied across industry. In this interview, Professor Jiebo Luo from the Department of Computer Science at the University of Rochester answers questions about the rapid development and practical uses of AI across industry, including best-practice principles for the companies that operate AI applications.
Professor Jiebo Luo, Department of Computer Science, University of Rochester, USA
Introduction to Professor Jiebo Luo
Jiebo Luo joined the University of Rochester in Fall 2011 after over fifteen prolific years at Kodak Research Laboratories, where he was a Senior Principal Scientist leading research and advanced development. He has been involved in numerous technical conferences, including serving as the program co-chair of ACM Multimedia 2010, IEEE CVPR 2012 and IEEE ICIP 2017. He has served on the editorial boards of the IEEE Transactions on Pattern Analysis and Machine Intelligence, IEEE Transactions on Multimedia, IEEE Transactions on Circuits and Systems for Video Technology, ACM Transactions on Intelligent Systems and Technology, Pattern Recognition, Machine Vision and Applications, and Journal of Electronic Imaging. Dr. Luo is a Fellow of the SPIE, IEEE, and IAPR, and a member of ACM, AAAI, and AAAS. In addition, he is a Board Member of the Greater Rochester Data Science Industry Consortium. Professor Luo has published more than 350 academic papers and holds more than 90 US patents.
Q: Why has AI developed so rapidly in recent years?
Professor Luo: I think there are four reasons
First, the development of sensor technology makes image collection much more convenient than before and helps gather excessive amounts of raw data for image processing.
Second, the explosion of Big Data makes data-driven modeling possible. In the past, image processing was model-driven rather than data-driven. We had huge amounts of data in the model-driven era, but limited computing resources and capabilities.
Third, rapid development of technologies in graphics processing and cloud computing has tremendously improved computing capabilities. Now, everybody can conduct image processing. When we started image processing in the early years, memory was only a few megabytes, and images were read at a very low speed. But now the situation has changed.
Fourth, the human resource factor. Some say that the development of AI is determined by the human resources we invest in. This is true to the extent that an increase in the number of highly-skilled researchers has driven the increase in data volume, and specifically, because many researchers are committed to facial recognition using open data sets. More human resources directly lead to frequent updates of the data sets, adding more data samples to research and model-testing. Similarly, the number of human resources for data labeling has increased. For example, if you want to develop a gesture recognition application, there will be dozens of developers participating in data labeling for massive volumes of data sets, training the data sets, all with the goal of making this available within one week. With the availability of large numbers of highly skilled staff, gesture recognition and even speech recognition applications can be developed very quickly.
Q: Do you think AI technology is mature enough to be put into commercial use across industry?
Professor Luo: I would like to talk about visual cognition first. Regardless of whether we are in China or the USA, the industries' most mature technology mainly focuses on facial recognition involving detection and recognition. Why is facial recognition technology "mature"? Because facial recognition technology is not a rigid body, but is close to the rigid body. That is to say our faces are all very similar but with slight differences, and can therefore be recognized even with facial expressions. This is similar to vehicle recognition, another AI technology that I think is relatively mature. Although vehicles are updated every year, the appearances are similar. This ability to focus on core technology and improve these small individual traits, or facial expressions, has helped the industry develop AI.
Since Stanford University released ImageNet, deep learning was unveiled. The identification accuracy of general objects has been greatly improved with deep learning. Take a chair as an example. We used to think that the chair was inexplicable and could not be found in different scenes. Because "chair" is not a visual concept but a functional concept (an object on which people can sit). In addition, because there are different shapes of chair, it makes it difficult for the technology to correctly identify the object. But through AI algorithms, we identify chairs in scenes now because we have huge amounts of data that can be used to train models for identifying chairs from various angles. As a result, the identification accuracy of chairs is improved.
There are several important points in the application of AI across different industries. As for the commercial use of AI technology, I have an empirical formula: 70-90-99.5. The formula divides the process into three phases:
The first phase is described as 70%. It means that if the accuracy of data set training of technical research in the lab reaches 70%, the research is proved to be feasible.
The second phase is turning academic research into practical use and the accuracy must reach 90%. When the accuracy of data set training reaches 90%, it must be verified in a vertical scenario. In this scenario, researchers must prove that the remaining 10% errors are not catastrophic, or make the training more accurate than 90% in a limited range. After all these verifications, a specific product can be put into commercial use.
I consider the threshold of the third phase to be 99.5%. I don't think accuracy can ever reach 100%. However, if the accuracy of machine training can hit 99.5%, machines are more accurate than humans, making at-scale application possible.
In fact, we can put a specified AI technology into commercial use when its accuracy reaches 90%. We do not have to wait for the 99.5%. During the process when you put a specified AI technology into practical use, you can also pay attention to other technologies related to your products. By combining study and practice, AI is embracing a promising future.
Q: In your opinion, what are the common failings in the applications of AI?
Professor Luo: When I was at Kodak, there was consensus: If a product was not perfect, it could not be put on the market. This same principle does not apply to AI. We can launch an AI product when the product is 90 percent mature. In this way, we will find the defects in practice. Another point I want to share is that, in China an enterprise may promote a product before the product physically exists, and when the product is finally released, it will fail to live up to expectations. In much the same way, many people think that AI will change their lives, but it may not. We should be vigilant about over-commitments and try to be modest so that people do not lose their confidence in AI. We must remain realistic and not be either too aggressive or too conservative.
Q: In recent years Big Data has been a buzzword used in many industries. And now, AI is the new buzzword. How do you explain the relationship between Big Data and AI?
Professor Luo: I find that some companies will split Big Data and AI by setting up separate departments to manage them. Personally I think this is wrong because if we do not combine Big Data with AI the value of Big Data will never be completely mined. As the volume of data increases rapidly people are incapable of analyzing without the support of AI. AI is uniquely capable of discovering subtleties that humans cannot easily see. This is especially true with huge amount of data that crosses multiple dimensions. Therefore, I don't think AI can be separated from Big Data, especially for data-driven modeling.
I'm not saying that AI can do nothing without Big Data. What I am stressing is that by using AI, the value of Big Data can be mined to a higher level of completeness. While maybe feasible to study Big Data or AI independently in an academic environment, from an industrial perspective they should never be isolated. AI must be combined with Big Data to maximize the value of the data.
Q: Finally, could you please share your opinion on the development trends of AI technologies in the next few years?
Professor Luo: : I'm going to talk about something I'm not too familiar with. I think while AI will be the focus of much research, hardware will continue developing also. In particular, NVIDIA, the GPU vendor, is making an attempt to move computing from cloud to mobile and edge devices, and has developed a product line of high-quality mobile chips. I think, as a future trend, device-cloud synergy will continue to develop.
Zhu Songchun, professor of statistics and computer science at UCLA, talks about the concept of ‘big AI' in the article An Exploration into Artificial Intelligence. At the time of its conception, AI was considered to be a broad theory. As we have moved forward in time, AI has been divided into several fields, such as visual cognition, speech recognition, text understanding, machine learning, and robotics. This is because researchers in each field believe that they can make breakthroughs.
As of today, major breakthroughs have been made in many fields. Communication between researchers happens frequently. In my opinion, all sub-fields of AI should be converged to achieve greater success. So-called ‘big AI' will be achieved through the communication and shared experience across various fields.
Let's take robots as an example. The most common cognition capability of humans and robots is visual perception. Humans can send commands to robots in a specific programming language instead of manually pressing a key.
Schedule planning, a sub-field of AI usually ignored by people, is incredibly important for those wishing to implement reliable AI. A good example of schedule planning is GPS navigation. Finding the best route from A to B is a planning problem. Though this sub-field is not so popular now, its future is completely reliant on robots. Assume that a robot must move an object from one place to another. To do so it must consider how to bypass any obstacles along its path. This is a typical schedule-planning problem because the robot must know its location and adapt to the real-time surrounding environment as well as to the map.
Robotics is a sophisticated field that requires various AI technologies, such as visual cognition, speech recognition, natural language processing, schedule planning, and machine learning. Collaboration between various AI sub-fields to achieve greater goals is the future trend of AI in the next several years.
To be honest, especially from the perspective of industrial applications, any combination of the latest AI technologies in a single application scenario is bound to see great success. Customers do not care what technologies are used, or whether it is visual cognition or speech recognition. With this flexibility, various approaches can be applied to make user and customer systems more robust.