the fundamental incompatibility of HMC and data subsampling
Last week, Michael Betancourt, from Warwick, arXived a neat wee note on the fundamental difficulties in running HMC on a subsample of the original data. The core message is that using only one fraction of the data to run an HMC with the hope that it will preserve the stationary distribution does not work. The only way to recover from the bias is to use a Metropolis-Hastings step using the whole data, a step that both kills most of the computing gain and has very low acceptance probabilities. Even the strategy that subsamples for each step in a single trajectory fails: there cannot be a significant gain in time without a significant bias in the outcome. Too bad..! Now, there are ways of accelerating HMC, for instance by parallelising the computation of gradients but, just as in any other approach (?), the information provided by the whole data is only available when looking at the whole data.