HINTS: An alternative to NUTS that doesn't require gradients

[Bob_Carpenter]

@Malcolm_S

In deep learning there’s been a lot of success using gradients of minibatches and I suspect the same will be true in a sampling context

Maybe, but see:

The Fundamental Incompatibility of Hamiltonian Monte Carlo and Data Subsampling

and

Subsampling in parallel and MCMC

I am working on lots of simulations and here’s a recent plot I came up with. [subsampling] Subsampling data and fitting each subsample separately leaves you with draws that are far away from the data resulting from the posterior that combines all the data. (First 15 plots) Posteriors for 15 simulated data sets of size N = 50 for a logistic regression model with intercept (alpha) and single predictor (beta) with weakly informative priors. (Final plot) Posterior resulting from combining all …