Markov Chain Monte Carlo Methods and Applications in Neuroscience

Abstract: An important task in brain modeling is that of estimating model parameters and quantifying their uncertainty. In this thesis we tackle this problem from a Bayesian perspective: we use experimental data to update the prior information about model parameters, in order to obtain their posterior distribution. Uncertainty quantification via a direct computation of the posterior has a prohibitive computational cost in high dimensions. An alternative to a direct computation is offered by Markov chain Monte Carlo (MCMC) methods.The aim of this project is to analyse some of the methods within this class and improve their convergence. In this thesis we describe the following MCMC methods: Metropolis-Hastings (MH) algorithm, Metropolis adjusted Langevin algorithm (MALA), simplified manifold MALA (smMALA) and Approximate Bayesian Computation MCMC (ABCMCMC).SmMALA is further analysed in Paper A, where we propose an algorithm to approximate a key component of this algorithm (the Fisher Information) when applied to ODE models, with the purpose of reducing the computational cost of the method.A theoretical analysis of MCMC methods is carried out in Paper B and relies on tools from the theory of large deviations. In particular, we analyse the convergence of the MH algorithm by stating and proving a large deviation principle (LDP) for the empirical measures produced by the algorithm.Some of the methods analysed in this thesis are implemented in an R package, available on GitHub as “icpm-kth/uqsa” and presented in Paper C, and are applied to subcellular pathway models within neurons in the context of uncertainty quantification of the model parameters.