Reducing blind spots
Associate Professor at the Multimedia Computing Group of Delft University of Technology Cynthia Liem has always shaped her own career path, guided by her passions for classical music and for knowledge engineering. The fact that she is both a certified computer scientist and a qualified pianist helps her to broaden the perspective on how to filter information in multimodal archives and use artificial intelligence wisely.
How did you end up in computer science?
‘As a happy accident. In high school, I was preparing to enroll at the conservatory to become a professional piano player. But people around me thought it would be a waste of my aptitude for the sciences if I would focus on music alone. So, I started to look out for a study I could do “on the side”. At first, I considered going to med school. But because of the numerus fixus, that would mean me taking up a spot of someone who might be more intrinsically dedicated to becoming a medical doctor. In my search for something else, I accidentally ended up at an orientation event for mathematics at TU Delft. They had a demonstration of Media and Knowledge Engineering, about building an accessible interface for some medical application. Even though up until then, I had never programmed anything myself, this was a field I thought I could make a meaningful contribution to.’
You simultaneously obtained two master’s degrees: in classical piano performance and in computer science. How did you manage to combine these two very different full-time studies?
‘That was a challenge indeed, since neither of the programs offered any standard opportunities to combine multiple studies. So, for each term, I would devise my own timetable, combining the courses in Delft with those at the Royal Conservatoire in The Hague. I was lucky enough to encounter some lecturers at the conservatory who were willing to give me private lessons during the evenings, allowing me to attend mandatory courses at TU Delft during the day.
After I obtained both my bachelor’s degrees, I wanted to proceed with both programs. At the conservatory, a master’s degree is not for everyone. But I was selected for this, which did mean the first years of my professional life as a graduated computer scientist would need to take place close to the city of The Hague.
You hold a PhD in Music Information Retrieval. How did you arrive at this field?
‘During my studies I had first stumbled upon this field of research, and immediately saw a niche for myself. I both have the required technical expertise and the domain knowledge: coming from a solid background in classical music, I can identify problems in the music sector, and with my computer science background, I also have the tools to come up with possible solutions.
Classical music is an important musical genre. But most music services are designed for pop music, while not fitting the ontology of classical music. For example, in pop music, you can describe a song by an artist, title and album. However, in classical music, interpreters typically perform multi-movement works that were previously written by a composer. If this type of music should be described in terms of an artist, title, and album, it will be awkward how to describe the composer and performer. In addition, these different interpretations will to a pop music listener be considered as near-duplicates, where one interpretation will be sufficiently representative for the work in general. However, when professionally studying these, it can be meaningful to know all available interpretations of the same work, as they are meaningfully different. Such use cases are not typically accommodated, I generally have been missing this degree of nuance in the way music is digitally treated. As another example, the fact that a certain piece is written in a major key does not necessarily make it happy, while many automated music classification methods tend to be trained on annotations suggesting this.
When I set out to obtain my PhD in this field, it turned out to be somewhat of a challenge. At that time, Music Information Retrieval was no topic of interest in Delft, and finding funding in the cultural sector was virtually impossible. In the end I managed to obtain a scholarship from Google, which I could spend on a topic of my own choice. Through the scholarship, I also visited Google multiple times as an intern. That is where I learned the ways of industry and confirmed my interest in making information more universally accessible, with attention to our blind spots. However, while industry clearly had the infrastructure for this, there were few incentives for industry to have attention to blind spots. Therefore, I felt it was important to work on complementary research in a not-for-profit environment, informing my choice to do this in academia.
“When we as computer scientists talk about data, models, predictions, or outcomes, that has a vastly different meaning to us than to colleagues in other sectors. From my own interdisciplinary experience, I can help bridging these types of barriers and reducing translational issues.”
What is your current research about?
‘It still is about making information accessible while minding our blind spots, although over the years, beyond music, I have been expanding to the responsible use of artificial intelligence broadly. I look at aspects of data quality and validity, annotation practices, and how machine learning tools decide upon a certain answer. What is it that the system is optimizing, is the pattern it has found indeed a meaningful one, and what counterexamples should we use in testing? In broadening my scope beyond music, I have been considering other societal relevant application areas like the library domain, where I have been looking into pluriform and accessible collection development, as well as questions of fairer data-driven hiring procedures. In my research, I prefer to work with data that has societal relevance. What is always hard when connecting to a new domain is that you need to learn to speak each other’s language. When we as computer scientists talk about data, models, predictions, or outcomes, that has a vastly different meaning to us than to colleagues in other sectors. From my own interdisciplinary experience, I can help bridging these types of barriers and reducing translational issues.’
Besides a scientist, you are also an active performer in the musical arts. Why do you want to be an artist as well?
‘Everything I do is about discovering new things. That also goes for my music. I like studying and performing the work of lesser-known composers and introducing them to a wider audience. There is so much more to classical music than Bach and Beethoven alone, and as an artist I can be a pro-active champion for these other composers’ work.’
You also quite often engage in public debate about computer science. How did that come about?
‘Thanks to my interdisciplinary work, I became known as a person capable of translating technical problems in accessible ways. From this, I was the first technical expert consulted to publicly explain the working of the algorithms that contributed to the childcare benefits scandal in a major newspaper. This was quite confrontational: I realized that from my computer science background, I might have made the same choices for the underlying risk analysis model, given the information that was available to the technologist. But clearly, these choices exacerbated unrightful accusations of innocent people, with major societal damage. I thus see it as my professional duty to reflect and teach on topics like algorithmic bias and amplification. Overall, I feel the need to inform society and make people, including my students in Delft, aware of what we do not see or what we do not know.’
In 2015, you were one of the founders of ICT Next Generation (ICTng), creating a network of junior ICT academics. What is the importance of this organization nowadays?
‘All around me, I see how industry dominates in ICT developments, both when it comes to the technology itself and the talent working on it. I see it as our duty as publicly funded academics to broaden and complement the perspective on ICT development and its role in society. It is imperative to involve multiple generations in these types of discussions, and that is where ICTng comes into play.
We first started out as a discussion and networking group that organized events during ICT.OPEN. Over the years, we claimed a role going far beyond that. For example, we distributed a questionnaire to gain an impression of what ICT academics consider important in their careers, which was met with a great response. We then became a Special Interest Group of IPN, joining the table in discussions about nationwide ICT developments. Today, IPN has a ‘next generation’ junior researcher seat in its board, and also actively involves junior researchers in its working groups. This is definitely a move in the right direction.
It would be good for IPN to initiate discussions about who we are as a Dutch computer science community, where we think our field should be heading, and what we think our role should be in society. As the computer science field, we used to consider ourselves to be Calimero – the little bird complaining to be overlooked because of being small. But now digitalization and AI became big societal discussion topics, and we suddenly have turned into Superman. That comes with the responsibility to use our powers well, and be mindful of hubris.’
Photo: Sjoerd van der Hucht