Uroš Seljak
Speaking Sessions
Solving inverse problems in science with AI
Scientific inverse problems in science are some of the most difficult problems characterized by high dimensionality (often 2d or 3d images), and expensive forward models. In recent years there has been a lot of development of AI powered solutions to inverse problems. Two of the most common approaches are explicit and implicit likelihood analyses. In explicit likelihood analysis one performs optimization or Monte Carlo Markov Chain (MCMC) sampling, and there has been significant progress in methods development, including novel gradient based MCMC methods such as MicroCanonical Langevin Monte Carlo. Implicit likelihood analysis, also called Simulation Based Inference, uses forward model simulations and Neural Networks to learn the likelihoods, and this approach has been shown to be able to solve problems that were previously not feasible. I will present examples of both approaches to data analysis in cosmology.
Biography
Seljak is a cosmologist who is particularly well-known for his research on cosmic microwave background radiation, galaxy clustering and weak gravitational lensing, and the implications of these observations for the large scale structure of the universe. In 1997, Seljak predicted the existence of B-modes in CMB polarization that are a tracer of primordial gravitational waves from inflation. Cosmologists are racing to detect this signal, which would prove inflation.
Seljak is actively developing methods for accelerated approximate Bayesian methodologies, and applying them to cosmology, astronomy, and other sciences. Examples of this work are the MicroCanonical Hamiltonian and Langevin Monte Carlo and Deterministic Langevin Monte Carlo samplers. Seljak is also developing machine learning methods with applications to cosmology, astronomy, and other sciences.