Automatic Reparameterisation of Probabilistic Programs

Abstract from Maria: Markov chain Monte Carlo (MCMC) algorithms can be used to approximate a probability distribution by continuously sampling from it. Some MCMC strategies, such as Hamiltonian Monte Carlo (HMC), use the gradient of the unnormalised density function to increase the quality of the obtained samples and the speed of convergence of the algorithm. But such algorithms can fail, and often silently, if the curvature of this target density function varies. One way to work around this problem is to reparameterise the distribution of interest, meaning to express it in terms of different parameters. In this talk, I will describe the practical challenges of finding a suitable reparameterisation, and demonstrate how we can use mechanisms available in recent probabilistic programming languages to implement a family of parameterisations often used in practice. In particular, we will look at a continuous relaxation of the question of what parameterisation to use and combine it with variational inference to obtain robust and efficient MCMC samplers. https://compcalc.github.io/public/gorinova/autoreparam.pdf