A Minimax Approach to Supervised Learning