Multi-modal Active Learning From Human Data: A Deep Reinforcement Learning Approach