Limitations and Scope

VBI is designed for amortized simulation-based inference (SBI) on whole-brain network models. It is most useful when you need full posterior distributions over a modest number of biophysically meaningful control parameters and can afford an upfront simulation and training budget in exchange for fast, reusable inference on new observations.

When to use VBI

VBI is a good fit when:

  • You work with whole-brain network models (e.g. Jansen-Rit, Wilson-Cowan, Wong-Wang, Montbrió/MPR, Stuart-Landau, Epileptor/VEP) and want to invert them against empirical neuroimaging data (fMRI, EEG, MEG, SEEG).

  • You want a full posterior distribution over model parameters, including uncertainty, rather than a single point estimate.

  • You plan to perform inference repeatedly on many subjects or observations, so the one-time cost of amortized training pays off.

  • You can run a sufficiently large simulation budget on CPU or GPU.

Limitations

Users should keep the following limitations in mind:

Parameter identifiability. Some parameters are only jointly (not individually) identifiable. For example, in the Wong-Wang model the global coupling G and synaptic coupling J are structurally non-identifiable and yield a curved, degenerate posterior. This is a property of the model and the data rather than of VBI itself, but it limits how uniquely individual parameters can be recovered.

Choice of data features. Inference quality depends strongly on selecting low-dimensional data features that are informative about the target parameters. Functional connectivity alone (FC/FCD) is often insufficient for estimating regional parameters; spatio-temporal features are usually required.

Noise sensitivity. High observational or dynamical noise can corrupt feature estimation and degrade the resulting posterior.

Simulation budget. There is no principled rule for the number of simulations required to train a reliable estimator. Adequacy must be checked empirically, for example using posterior z-scores and shrinkage or simulation-based calibration (SBC).

Upfront computational cost. Single-round amortized training requires a substantial one-time compute investment (roughly tens of minutes to several hours, depending on the model and simulation budget) before inference becomes inexpensive.

Preprocessing. Some statistical information (e.g. signal mean and variance) can be lost during feature preprocessing, reducing the informativeness of the extracted features.

Further reading

For a detailed validation study and an in-depth discussion of these points, see Ziaeemehr et al., eLife 2025.