Efficient sparse Bayesian learning using spike-and-slab priors