On Data Efficiency of Meta-learning