New discoveries with statistical inference
FCAI postdoc Ulpu Remes develops ELFI, the open-source software for likelihood-free inference.
Photo: Jani Närhi/University of Helsinki
Ulpu Remes has a background in speech and language technology. Remes has used ELFI since it was released in 2017, and now she is preparing to take over the maintenance of this open-source package.
Her doctoral research at Aalto University focused on missing-data methods for noise-robust automatic speech recognition. “After that, I had a chance to work with simulator-based models for human-computer interaction, and this is when I learned about ELFI. We used it to learn individual user characteristics based on observed interaction data. From there I went on to participate in other ELFI work and later started to contribute as a developer. Meanwhile I also worked on multi-fidelity optimisation for atomistic structure search and contributed to BOSS,” says Remes, who is now a researcher at the University of Helsinki.
ELFI uses observed data to constrain simulator parameters
ELFI provides a way to conduct statistical inference with simulator-based models that do not have a computable likelihood. The project started as a collaboration between FCAI’s director Samuel Kaski and professor Jukka Corander from the University of Helsinki and University of Oslo. The methods implemented in ELFI represent so-called likelihood-free or simulator-based inference methods.
According to Remes, likelihood-free inference is needed when researchers observe a system in the real world and then want to fit an appropriate mechanistic model to the observed data.
“We do this to learn about unobservable system parameters or because we want to make predictions about system behaviour in unseen conditions. In both cases we want the model fit to capture both what we can and cannot learn based on the available observations. For example, in case the observations are not informative about some model parameters at all, we want to take that into account when we make conclusions or decisions based on the model, and we use likelihood-free inference to fit the model because it captures these uncertainties,” says Remes.
ELFI was planned so that it serves both the practitioners who need likelihood-free inference in their simulator-based studies, and the method developers who want to create and evaluate new methods for likelihood-free inference.
“I think this has been an important motivation behind ELFI work, and also these are the two audiences we have in mind when we now maintain and develop ELFI,” says Remes.
Since 2019, ELFI development has been coordinated and led by Henri Pesonen at the University of Oslo.
New users of ELFI are welcome
Several new methods have been contributed to ELFI since its initial release, and ELFI has been used in various simulator-based studies. Some examples are presented in a recent article that discusses likelihood-free inference for real applications.
“We appreciate feedback from ELFI users and hope that new users also find ELFI. New applications are a chance for us to learn about new problems and learn how we can make ELFI better,” says Remes.
Discover FCAI's other software contributions at https://fcai.fi/software