Mining gold from implicit models to improve likelihood-free inference