How to choose an inference method

How to choose an inference method#

sbi implements three sets of methods for inferring the posterior:

  • NPE (Neural Posterior Estimation),

  • NLE (Neural Likelihood Estimation), and

  • NRE (Neural Ratio Estimation).

For each of these methods, sbi implements an amortized version and a sequential version. Which algorithm should you choose?

Explicit recommendations#

  • If have high-dimensional simulation-outputs (images, time-series,…), use NPE or NRE

  • If you want fast sampling, use NPE

  • If you have iid observations across many trials (and simulations are relatively expensive), use NRE or NLE

  • If you want to perform inference for many observations, use amortized methods (which is the default of the sbi package).

  • If you want to perform inference for few observations (<10), then start with amortized methods and switch to sequential methods only if the simulation cost is too high.

  • Do not use SNPE_A or SNPE_B unless you are an expert user and know what you are doing.

Additional explanation#

NPE vs NLE vs NRE#

NPE is the only method which estimates the posterior directly. This means that it will not require further sampling steps (such as MCMC) after training. This makes NPE the fastest to .sample() from the posterior estimate after training. In addition, it can use an embedding network to learn summary statistics (just like NRE, but unlike NLE), see the tutorial here.

NLE learns the likelihood. This allows NLE to emulate the simulator after training. To sample from the posterior, NLE is combined with MCMC or variational inference (see here), which makes it slower to .sample(). An advantage of NLE is that, if you have multiple iid observations, then it can be much more simulation efficient than NPE (see tutorial here).

NRE learns the likelihood-to-evidence ratio. It trains only a classifier (unlike NPE and NLE, which train a generative model), which makes it fastest to train. Like NLE, NRE has to be combined MCMC or variational inference to run .sample(). Like NPE, NRE can use an embedding network to learn summary statistics.

Amortized or sequential#

Amortized methods allow you to perform inference for any observation without running new simulations or retraining. In contrast, sequential methods tailor inference to a single observation, which can make them more simulation efficient.

If you want to perform inference for many observations (>10), use amortized methods. If you have only few observations (or just a single on), then we still recommend you start with amortized methods. Only if simulation cost becomes too large, switch to sequential methods.

The default behavior of sbi is to run each method in its amortized form. To run sequential methods, you have to build follow the tutorial here.

Navigation