Author(s)

Hao Ma, Xuefeng Zhu and Yifeng Zheng

Corresponding Author

Hao Ma

ABSTRACT

Nowadays, the visual reasoning model has been able to answer complex questions that cannot be answered by the visual question answering model on the CLEVR dataset. However, most visual reasoning models require a large amount of data for strongly supervised learning, which is easy to increase the cost of data labeling. In addition, these approaches will lead to over-fitting, thereby reducing the generalization performance of the model. To solve the above problem, in this paper, we propose a novel model combined with active learning. It utilizes active learning to select the most informative and representative sample as the training data efficiently and accurately. Therefore, fewer samples can be employed to train a visual inference model to obtain higher accuracy and better generalization ability. The experimental results from three aspects show the effectiveness of the proposed approach with active learning for the visual reasoning model.

KEYWORDS

Visual reasoning; active learning; overfitting; generalization ability; labeling effort