Covariance-controlled adaptive Langevin thermostat for large-scale Bayesian sampling

摘要: Monte Carlo sampling for Bayesian posterior inference is a common approach used in machine learning. The Markov chain procedures that are often discrete-time analogues of associated stochastic differential equations (SDEs). These SDEs guaranteed to leave invariant the required distribution. An area current research addresses computational benefits gradient methods this setting. Existing techniques rely on estimating variance or covariance subsampling error, and typically assume constant variance. In article, we propose covariance-controlled adaptive Langevin thermostat can effectively dissipate parameter-dependent noise while maintaining desired target proposed method achieves substantial speedup over popular alternative schemes large-scale learning applications.