Posterior covariance information criterion for arbitrary loss functions
share
We propose a novel computationally low-cost method for estimating the predictive risks of Bayesian methods for arbitrary loss functions. The proposed method utilises posterior covariance and provides estimators of the Gibbs and the plugin generalization errors. We present theoretical guarantees of the proposed method, clarifying the connection between the widely applicable information criterion, the Bayesian sensitivity analysis, and the infinitesimal jackknife approximation of Bayesian leave-one-out cross validation. An application to differentially-private learning is also discussed.
READ FULL TEXT