“Collaboration is a feature of the Finnish research community”

Academy Research Fellow and ELLIS Member Ayush Bharti sees the strong collegial network as an asset to his research on making machine learning methods faster and more robust.

Scientists study complicated physical and biological systems, such as how the universe expands or how diseases spread, by making mathematical replicas, or models, of natural phenomena. To be useful, the input to these models, called parameters, need to be ‘tuned’ using data collected in the real world, to figure out which parameter setting led to that data. As an example, if you’re studying the spread of infectious disease, your data could be how many people are infected each day, recorded over weeks or years. The parameters in this case are unknown quantities you want to ‘guess’ from the data, for example the rate of which the disease spreads or how fast people recover.

Another way to think about this is the analogy of an old radio. Tuning the knobs, or parameters, of the radio (the model) will hopefully lead to the best sound, or a match to the data and an accurate reflection of the natural phenomenon. However, the machine learning methods used to tune the parameters can break down and give misleading answers when there is a mismatch between what a model can produce and the data we observe. This is quite common, since models rarely capture everything in the real world. My research aims to address this problem by developing machine learning methods that are robust to the mismatch between model and data, meaning that they work reliably even when there are discrepancies.

When I did my PhD in wireless communication, I was essentially trying to find the knob settings for models that described how radio waves bounce around in different environments. Now, as an Academy Research Fellow, I develop machine learning methods that are more broadly applicable to many different fields, from cosmology to epidemiology to the original radio propagation problems I used to study.

Another issue that I am interested in is that of computational cost. With widespread access to computing resources, these models are becoming increasingly complex and can take a long time to run. This makes it challenging to tune their parameters accurately within a reasonable time. My recent work with colleagues from Aalto University and University College London addresses this computational bottleneck, which could take days or weeks, by using a cheap but less accurate approximation of the model to reduce the total compute needed. This idea is reflected in the name of my newly founded research group, the Approximate Inference Lab.

My move from the applied field of wireless communications to the computer science department at Aalto has been quite smooth mainly because I had access to the strong collaborative network in Finland. I first noticed this when I joined the Finnish Center for Artificial Intelligence FCAI as a postdoc in 2021. The FCAI network, teams and joint meetings really encourage collaboration, which allowed me to work with researchers from other groups like Luigi Acerbi and Vikas Garg.

Finland has been able to attract talented people in this field and has established itself on the map in terms of machine learning research in Europe. Now with the resources, visibility and new faculty of the ELLIS Institute, I see the situation as only getting better. I think we enjoy a seamless collaboration culture here, compared to internal competition or isolation that I’ve heard about from colleagues elsewhere. The ELLIS brand can attract talented people, which is the main driver of impactful, quality research.