Convergent Expectation Propagation in Linear Models with Spike-and-slab Priors

摘要: Exact inference in the linear regression model with spike and slab priors is often intractable. Expectation propagation (EP) can be used for approximate inference. However, regular sequential form of EP (R-EP) may fail to converge this when size training set very small. As an alternative, we propose a provably convergent algorithm (PC-EP). PC-EP proved minimize energy function which, under some constraints, bounded from below whose stationary points coincide solution R-EP. Experiments synthetic data indicate that R-EP does not converge, approximation generated by better. By contrast, converges, both methods perform similarly.