Learning from discriminative feature feedback