Each year, the NeurIPS conference hosts several competitions where researchers from all over the world can compete in challenges ranging from basic research to AI applications. Scientists from FIT CTU brought home wins in two competitions from NeurIPS 2022 - Weather4cast and Ariel Machine Learning Data Challenge.
By winning the Weather4cast competition, the faculty team capitalized on their collaborative research with Meteopress in improving AI algorithms for weather forecasting. The winning team of researchers from the Data Science Laboratory, consisting of Bc. Jiří Pihrt, Bc. Rudolf Raevskiy, Mgr. Petr Šimánek and Ing. Matej Choma had to predict the rainfall as accurately as possible based on satellite data from 7 different regions over a period of 2 years. The uniqueness of their winning project in the Weather4cast competition lies in the fact that it can very realistically simulate the forecast of extreme precipitation in places where weather radars are not available, for example in developing countries.
"Some neural networks have special architectures that help generate more physically faithful simulations. This makes the predictions not only more accurate, but also more realistic," says doc. Pavel Kordík, vice-dean for cooperation with industry at FIT CTU, and adds: "By combining the knowledge of scientists and students from the faculty with the Meteopress company, we are already able to forecast the weather very accurately for several tens of minutes in advance. This is one of the prime examples of how to effectively link theoretical research with practice."
The second big success at the NeurIPS 2022 conference is the victory of Ing. Ondřej Podsztávek in the Ariel Machine Learning Data Challenge focused on modelling the atmospheres of planets outside our solar system (exoplanets). To solve the challenge, Ondřej used his experience from long-term research cooperation with RNDr. Petr Škoda, CSc. from the Institute of Astronomy of CAS and FIT CTU and prof. Ing. Pavel Tvrdík, CSc. from the Department of Computer Systems, FIT CTU.
"The aim of the project was to design the most efficient method for detecting the temperature of exoplanet atmospheres and the amount of gases in them. I therefore proposed a so-called 'deep ensemble' algorithm, which consists of twenty convolutional neural networks adapted to process the spectra produced by the distribution of light passing through the atmospheres of exoplanets," explains Ondřej Podsztavek, adding: "Research on exoplanets helps us understand the uniqueness of our planet Earth."