1:10 pm (Cosmology/ BCCP)
Justin Alsing, CCA
"Towards scalable likelihood-free inference for cosmology"
Many statistical models for cosmological data analysis can be simulated forwards but have intractable likelihoods, for example because of complicated reduction pipelines and systematics, or non-linear structure formation and astrophysics impacting small scales. Likelihood-free inference provides an alternative paradigm for doing Bayesian inference using forward simulations only, eliminating the need to make uncomfortable likelihood approximations and allowing us to extract information from previously inaccessible statistics and scales. Traditional Approximate Bayesian Computation (ABC) methods involve drawing parameters and forward simulating mock data, accepting the parameters if the simulated data are within some small distance \epsilon of the real data (recovering the true posterior in the limit \epsilon -> 0). These methods scale poorly with the size of the dataset and critically slow down as \epsilon -> 0, often leading to expensive analyses with overly conservative error bars. I will introduce a new approach to likelihood-free inference that is “epsilon-free”, bypassing the limitations of ABC, and use massive lossless data compression to dramatically reduce the size of the data-space; together, these advances provide a framework for performing scalable likelihood-free inference from cosmological surveys, using reasonable numbers of forward simulations.