Visual Interaction with Deep Learning Models through Collaborative Semantic Inference