Robust uncertainty estimates with out-of-distribution pseudo-inputs training