The Annual Conference on Neural Information Processing Systems (NIPS) is a 40-year-old conference, considered nowadays as the largest and most influential in Artificial Intelligence, with usual participation numbers in the orders of the thousands.
With 5,000 registered participants in 2016 and 8,000 in 2017, the exponential interest towards the field is making it a very popular event, with tickets for the 2018 edition selling faster than those for the World Cup final. In a culmination of months of hard work, the article “Adversarial Multiple Source Domain Adaptation” was submitted by ISR PhD Shanghang Zhang, competing with 4,856 other articles (the previous year had an acceptance rate of 21%) for publication in NIPS 2018. “It was a great feat to have this publication accepted. In the deep learning and motion learning field, NIPS is the best conference.” Explained Professor João Costeira, who co-oriented the work.
While doing research for her PhD, Shanghang Zhang focused on analyzing large-scale traffic data captured from city cameras. “When you are dealing with huge amounts of footage it’s expensive and difficult to label all the data. Even with limited amounts, it’s hard to have guaranteed good performance, so with such large-scale testing many problems arise that need advanced solutions, typically in motion learning.” The researcher explained that the NIPS article is, in fact, a representation of an important subtopic on domain adaptation, that is part of Shanghang’s work (more about her PhD thesis here). Among her publications, Shanghang managed several top-ranked papers. “We tried to solve the problem of labelling big amounts of data. Our algorithm performed very well in trimming the amount of data labelling
so from this point of view, its important work for the community since the model can be used on several fronts.”
In order to build a good system to process the data, there are several steps involved. There’s a need to collect, label, clean and pre-process the data. “Real world data is very noisy and much more challenging than simulation data. Nowadays many cities are equipped with hundreds of surveillance cameras that capture real-time video, 24/7. If you treat that information you can have applications in many fields, from traffic management, smartly controlled traffic lights, to information transfer for autonomous cars. These images are a big source of info but cannot be processed only by government personnel.”
Shanghang’s work focused on exploring different methods and advanced measures in order to solve a real-world problem. “When I looked into a university to apply for, I tried to find directions that where a combination of science and real-world problems. My motivation for the PhD was to advance the barrier of technology, even if in a small way, and contribute to improving people’s lives.” In fact, the Signal and Image Processing group aims at going from real-world issues to developing research at a personable level.
“We want to translate science into impact.”
SIPg researcher Cláudia Soares explained that there are many examples where this kind of research seems more than necessary. “If you think that at a call centre the calls are completely random, this translates into a waste of time for both the company and the person that is being called. With a more intelligent use, for instance, from demographic studies, it’s possible to focus on people who are actual targets and interested in your service.” Cláudia explained.
Smart cities are becoming a reality, and integrated systems being used more than ever. “The idea is to use the people themselves as sensors and receivers of information. It’s including tech to the service of the city management.” Using a router on a city for gathering information, for instance on the profile of users and their mobility, is already a reality, being tested by a strict but very mobile group of people. Work on big data in partnership with real-world sources allows for another level of contribution, be that from international partners or even companies. These projects may imply working with several partners and multidisciplinary teams, which involves a difficulty in management but also a potential for great contributions. “The goal is to translate the pulse of the city. The great challenge is to process and estimate data from diverse natures and heterogeneous supports. It can be a difficult process because it depends on factors of a social nature, but that’s what we enjoy.”
