Our missionWe are an international collaboration of researchers interested in developing and applying cutting-edge statistical inference techniques to study the spatial distribution of matter in our Universe. We embrace the latest innovations in information theory and artificial intelligence to optimally extract physical information from data and use derived results to facilitate new discoveries.
Our latest results
Why neural networks don’t work and how to use them
Throughout the scientific community neural networks are being used for a variety of different tasks. Unfortunately, this is normally done without thought of the statistical implication. Here we lay down the statistical notions showing why neural networks cannot be used by themselves for scientific purposes. We then provide a suite of methods which allows them to be used safely within a statistical framework for parameter inference.
Neural physical engines for inferring the halo mass distribution function
The tracing of the dark matter distribution by halos is complex and requires the knowledge of unknown small scale astrophysics. We use physically motivated neural networks to agnostically probe this bias model. The tunable parameters of the neural network are inferred as part of the BORG algorithm, and provide an exceptional fit to the halo mass distribution function. No training data is necessary since the network is conditioned on the observed halo catalogue directly.
Algorithms for likelihood-free cosmological data analysis
Many numerical models in cosmology can only be simulated forward. We have developed two novel algorithms to perform rigorous statistical inference from these models, in two different scenarios. The first one, BOLFI, reduces the number of required simulations for physical parameter inference by several orders of magnitude. The second one, SELFI, allows a full reconstruction of the primordial matter power spectrum.