In February, the Aalto University School of Science presented the 2020 Master’s Thesis Awards. About 400 students graduated from the School of Science with a master’s degree, and only about one percent of them were awarded the prize.
Of the four winners, two are from the Department of Computer Science, Silja Sormunen and Erik Härkönen.
The title of Silja Sormunen's thesis is ‘Distinguishing subsampled power laws from other heavytailed distributions.’ Her thesis was supervised by Professor Jari Saramäki.
Power law distributions have frequently been observed in both natural and artificial systems, and they play a prominent role in network science. However, distinguishing power law distributions from other heavy-tailed distributions is not straight-forward, and this task is further complicated if the data is subsampled.
'In this work, we analysed how well two commonly used methods for detecting power law distributions succeed in distinguishing subsampled power laws from other heavy-tailed distributions. The thesis showed that classifying the distribution’s type correctly is challenging, but that subsampling affects the two methods’ performance differently – in fact, one of the methods can in some cases perform better on subsampled data than on the original distribution,' Sormunen explains.
Sormunen is a doctoral track student who started working on her doctoral dissertation already at the beginning of her master's studies. Read more about the doctoral track here.
The title of Erik Härkönen's master's thesis is ‘Unsupervised Discovery of Interpretable GAN Controls.’ His work was supervised by Professor Jaakko Lehtinen. The work deals with the control of data-based generative GAN machine learning models in ways that are understandable and interpretable for humans. In their basic form, GAN models are so-called black boxes, the interpretable control of which requires significant effort.
'We found out that the models can be analyzed and controlled in a simpler way than previous methods, utilizing PCA, which is a basic method of statistical analysis. Due to its unsupervised nature, the presented method also makes it possible to find completely new, previously unknown interpretable control methods,' says Härkönen.
The research work that was done for the thesis was conducted in part at Adobe Research in Cambridge, Massachusetts. The research article written about it was approved to the world’s largest and most important machine learning conference, NeurIPS, in 2020.