People of ACM - Qiang Yang
June 28, 2018
To what avenue of research are you devoting most of your time right now?
I am working on a topic in machine learning known as transfer learning. Transfer learning aims to take an existing learned model in one domain or task and transfer that knowledge to a new domain. It can be an effective approach as long as the two domains are similar. Transfer learning emulates human problem-solving ability in learning by analogy. Humans attribute great wisdom to those who can easily transfer or adapt knowledge to different domains. In educational psychology, a measure of a good teacher is his/her ability to enable students to “learn how to learn.” This is known as “transfer of learning” in education. In AI, unless machines can acquire a similar ability to generalize and adapt in new situations, we won’t have true intelligence.
In practice, there are several advantages in transfer learning. First, if we only have “small data” in a new domain, we can find a “big data” domain and transfer the model from that domain to the new domain. This partially solves the problem that we often only have small data in practice. Second, if we take a model that is built in the cloud, and transfer that model to a client device, say a mobile phone, we can use transfer learning to adapt the general model to a specific personalized model for an individual. This “one-way” transfer of models protects user privacy, as there is no need for the individual to upload the personal data on the client device to the cloud. Third, transfer learning can potentially make a machine learning model more robust, as it allows the system to cater to variations in a trained model, even when the application environment is different from that in training.
One of the projects your team was working on at HKUST was a natural language dialog system. Where will we see the greatest advances in natural language processing in the next five to 10 years?
I believe that natural language processing (NLP) is a more difficult AI problem to solve than its computer vision counterpart, as humans are used to expressing a vast range of meanings using a few words. Natural-language dialog will be a challenging and useful area in which to do research. If you have watched the HBO TV series “Westworld,” you will recall that scientists who built the robots believed in a theory of intelligence known as “bicameral mind”—that intelligence emerges from the dialog of two faculties in the brain. In our research, if we program a computer system to have a conversation with another system or human, it is likely to spark something different. I believe that a dialog is like playing chess: the two parties who talk to each other negotiate their thoughts through a series of sentences following a “policy,” which can be learned using a reinforcement-learning algorithm. Recently, we have worked on the problem of transferring a general dialog policy to a personalized policy that can learn an individual’s preferences.
In an online coffee-ordering scenario, for example, one can take a coffee-ordering routine and adapt it to fit an individual’s taste or cup size preference. These task-oriented dialog systems can be very useful in practice, as they can amplify an organization’s ability to serve a huge number of people, for example in a bank or an airline’s call center. In a future call center, people and machines (powered by these task-oriented dialog systems) will be able to work in sync: as the machines serve the masses, people can teach machines how to improve their performance. This teacher-student relationship will be a realistic scenario, say in five to 10 years.
Your most-cited article, "A survey on transfer learning," co-authored with S.J. Pan in 2010, provided an overview of how transfer learning might boost machine learning performance. Will you tell us a little about transfer learning and how it could advance AI in the coming years?
I truly believe transfer learning is going to shape AI in years to come. When you look at how humans learn, you’ll realize that it is more important to “learn how to learn.” It is the ability to adapt that drives evolution. Machines today learn one thing at a time. However, if we endow machines with the ability to learn knowledge as well as to learn how to generalize the knowledge, we can enter a positive cycle of “lifelong machine learning,” whereby machines can continue to improve themselves.
Algorithmically, there have been several major recent developments. When we look at deep learning, we can see that knowledge is expressed in several layers of representations, from low to high. For a given domain, some of these layers encode more common knowledge between different domains. Some researchers have been successful in quantifying the knowledge embedded in each layer when transferring between image-classification tasks. It is also possible to use a deep neural architecture to jointly learn the common meanings of images and words, which makes computers realize the meaning of images through words. Many new cross-domain applications can thus be designed, such as cross-domain product recommendation and sentiment classification.
With Charles X. Ling, you co-authored the 2012 book Crafting Your Research Future: A Guide to Successful Master’s and Ph.D. Degrees in Science & Engineering. Why did you think the book was needed, and what is one important idea you hope prospective graduate students take away from it?
Over the years as a professor, I have had the chance to work with lots of PhD and Master’s students. I have observed many patterns of research that are common among many beginning students. Charles felt the same, and we decided to write it up to benefit future students.
One thing that we have emphasized in the book is that the ability to find a good topic for research is sometimes more important than working on a chosen topic. There are patterns for identifying good research subjects that are very similar to doing a market study when writing up a business plan: you’ll need to argue for significance, novelty, simplicity and decomposability. There is also the element of being able to find the right data when working in empirical science. A good test is to find someone outside your field, say your grandmother, and tell them what you are planning to work on. If you can excite these other people, you are more likely to have selected a good topic.
Another issue we cover in the book is how to write papers and treat reviews. Writing good papers is a skill that is acquired through practice, but it is as much of an art as it is a science—there are patterns that you can follow to make people understand you as well as appreciate your work. Finally, we try to paint students a realistic picture of life as a researcher and professor; it is indeed not that rosy if you consider all the work leading to a tenure position and beyond, but the “joy of finding things out” (to quote Richard Feynman) is certainly worth it!
Qiang Yang is Chair Professor and former Head of Department of Computer Science and Engineering at Hong Kong University of Science and Technology (HKUST). His research interests include transfer learning, machine learning, planning and data mining. He was also a founding head of the Hong Kong division of Huawei’s Noah’s Ark Lab. Recently Yang served as Chair of the ACM KDD 2017 Test of Time Award Committee, as well as the Chair of the Awards Committee for the International Joint Conference on Artificial Intelligence (IJCAI 2017). He is also serving as president of IJCAI from 2017 to 2019.
Yang has authored or co-authored 349 articles and three books. He has also held a number of editorial positions, including serving as founding editor of ACM Transactions on Intelligent Systems and Technology (TIST) from 2009 to 2015. In 2017, he was named an ACM Fellow for contributions to artificial intelligence and data mining.