Skip to content

Implement VSBC diagnostic from Yao et al. #60

Open
@martinmodrak

Description

@martinmodrak

The Yes, but Did It Work?: Evaluating Variational Inference paper by Yao et al. proposes a less stringent diagnostic for variational inference than full SBC (as we know the approximation is almost always imperfect and SBC is thus eventually bound to fail). The main point is that we only test for bias, i.e. whether P(true_value < fitted_mean) = P(true_value > fitted_mean) across multiple simulations.

The implementation proposed in the paper seems specific to methods where we get exact mean, but the sample-based version should also work.

It might be neat to support this bias test, as it could be more generally useful in understanding if approximations are a problem or not.

Metadata

Metadata

Assignees

No one assigned

    Labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions