Stan and federated learning

Hey Stan community,

I am wondering whether there is an existing application of Stan in federated learning (for example with vertically-partitioned data). Is there a related publication or specific application or something in development?


As currently coded, Stan assumes you have the full data set at model construction time before any sampling or optimization starts. You might want to look at some of the subsampling approaches to HMC for ideas, but we don’t have any plans of which I’m aware to implement any of these.

There’s nothing like this in Stan, but you can find SMC implementation in Turing.jl (SMC only requires a subset of the data at any given time), as well as piecewise-deterministic Monte Carlo (which can be used for error-free subsampling).