EECS students won the SAIS Best AI Master's Thesis Award 2020

We caught up with Elias and Ulme and asked them some questions about the award.

Congratulations! How does it feel to receive this award, and what does it mean to you?

“I am of course extremely proud and touched that some of Sweden's leading researchers in the AI space are finding my master's thesis exciting!” says Ulme Wennberg. “I have been putting in a lot of long hours into this project together with the brilliant people in my lab and my exceptional thesis supervisor - so it has been an exciting journey for all of us and it could not have ended in a better way.”

Tell us briefly about your thesis?

Ulme Wennberg:”My thesis “An evaluation of BERT for a Span-based Approach for Jointly Predicting Entities, Coreference Clusters and Relations Between Entities” is in the area of Natural Language Processing, a core subfield of AI, and evaluates the use of recent language models for important NLP tasks.”

Elias Lousseief: “My thesis, entitled “MahlerNet: Unbounded Orchestral Music with Neural Networks”, treats the interesting open problem of music composition by machines, based on machine learning models. It was motivated both from an artistic point of view, since I am also a composer and have studied composition, as well as from a scientific perspective and treats orchestral composition; a focus that, at the time, was fairly unexplored. Feel free to check out the project website at www.mahlernet.se.”

What excites you the most about your area of research?

UW: “One thing that struck me early on with the NLP-space is that while many models can learn to excel in specific tasks, I believe there has been a fundamental cornerstone missing in that algorithms have been struggling to find efficient ways to encode real world knowledge for reasoning tasks. Using NLP-based information extraction allows us to address this by facilitating a way to automatically generate knowledge graphs (similar to mind-maps where entities are represented as nodes in a graph and relations form connections between entities). This way we can encode real-world knowledge which can then be used for downstream reasoning tasks.”

EL:"To me it is very exciting to think that machines, with their "thinking capabilities" that largely differ from ours, could be able to either build onto existing music in new directions or develop entirely new music that doesn't sound like anything else. Another very inspiring idea is to let computers track down characteristics of existing music and then fuse these in amazing ways so that we could finally hear what it would sound like if Backstreet Boys and Mozart had written a hit together. This particular area has advanced a lot since I wrote my thesis, even though some music lovers would probably say that music is more complicated than this and doesn't lend itself to such simple operations."

In what way is your research important for society and what use or problem solving do you see in the future?

UW: “This research addresses how to aggregate information from large unstructured text masses and convert it into a structured form. I would argue that we are in a time where this type of unstructured text information is more readily available than ever before, thanks to the internet. We also make our multitask model DYGIE++ publicly available on GitHub, that sets new state-of-the art in entity recognition, relation extraction and event extraction on all datasets we tested it on. So if you are currently working on any of these NLP-tasks - you may want to check out our model! :).“

What are your future plans?

EL: “I have just started a new job where I have mostly worked from home due to the ongoing pandemic. It is a bizarre situation to get to know new colleagues and routines from home, but it has worked out so far. Working surely involves a lot of knowledge that is not taught at school and currently, I both enjoy and struggle to learn as much as possible in this new world. My current job does not involve AI (but programming of course), but I look forward to working more with machine learning in the future and I plan to spend time improving MahlerNet and contribute with a piece of music for Bob Sturm's AI Music concert 2021."

You seem to have a good connection with your supervisors. Anything in particular you would like to thank them for?

EL: ”When I started writing my thesis, I had another supervisor and while doing my literature survey, the name of my supervisor, Bob Sturm, came up several times. It was therefore amazing that he soon after moved to Sweden and joined KTH and agreed to be my supervisor. It truly gave me the perspective of someone "in the business" which motivated me a lot. Bob has also utterly been an academic inspiration in general and a motivator of high class. Thanks to him, we have written an article based on my thesis, presented it at a conference and now this prize... What I want to say is that Bob has been both a good supervisor, in particular, but also a good academic mentor, in general, and I wish to thank him very much for this."

UW:"I had the pleasure to get to know Gustav through reading groups and hallway-discussions the year before the project started, and I always found it fascinating to discuss with him and hear him explain his thoughts and ideas.
When I started working on my thesis at University of Washington in Seattle, USA, we set up weekly Skype sessions to discuss ideas and touch on last week's progress - despite being 9 time zones away. Throughout this, Gustav has been both a great mentor to me, and a good friend of mine. Since the project ended we have touch-based every few months to discuss some exciting advancements and keep in touch. Gustav is one of the people who I have learned the most from when it comes to machine learning as a field and as a research discipline. I want to thank him for all this."

What do you believe is the success formula for a good thesis?

EL: “Write about something that you're really interested in OR make sure to get really interested in what you're writing about!”

Read more about the thesis and the prize at the website for Swedish AI Society.