Near-Optimal Bayesian Active Learning with Noisy Observations