Improving the Minimax Rate of Active Learning