Information Maximization for Few-Shot Learning