Cosmology with machine learning

Scientifically, the Lab is currently concentrating on probabilistic neural networks as a method for improving photometric redshifts to measure dark energy with the upcoming LSST survey with the Vera Rubin Observatory. The nature of dark energy is one of the biggest mysteries in physics and astrophysics as it represents 72% of the energy density of the Universe but we have very few constraints on its physical origin.

Because dark energy affects the fabric of the Universe, the LSST survey will image over 5 billion galaxies allowing us to see the effects of dark energy on the distribution of matter at unprecedented precision. For this project to be successful, we need to know the distances to all these galaxies and understand the uncertainty of our measurements.

Because galaxies are very complex objects, the most promising methods involve using machine learning to empirically learn about how galaxies vary across cosmic time. The biggest challenge for these machine learning models is to have accurate uncertainty estimates. Our novel approach to this problem is to use probabilistic neural networks, which are like more typical neural networks but can propagate uncertainties throughout the model.

Probabilistic deep learning such as Bayesian Neural Networks has many advantages compared to traditional neural networks, including: better uncertainty representations, better point predictions, and offers better interpretability of neural networks because they can be viewed through the lens of probability theory. In this way we can draw upon decades of development in Bayesian inference development. We think that probabilistic neural networks have the potential to be a transformative tool for cosmology by embedding both physics of the universe and telescope and instrumental characteristics into the same neural network.