Predicting the association between fine-grained attributes and black-boxed perceptual performance

Per-image performance of a certain visual perception algorithm is the combination of per-task performances in the image. Based on the black box test, we address to discover and explain the potential shortness of the evaluated intelligent algorithms/systems at the fine-grained task-level by human knowledge. By assuming the domain knowledge in visual tasks could be represented by a latent vector which is a sparse embedding of the catenated object-level and image-level features, we propose a latent dictionary learning framework for joint latent knowledge representation and knowledge-output regression at task level. In this way, so we can use semantic concepts to explain the relationship between test cases and test results. The experiments validate the idea of task-level explainable AI evaluation initially as well as the effectiveness of proposed method.