Good, Better, Best: Textual Distractors Generation for Multi-Choice VQA via Policy Gradient.

Textual distractors in current multi-choice VQA datasets are not challenging enough for state-of-the-art neural models. To better assess whether well-trained VQA models are vulnerable to potential attack such as more challenging distractors, we introduce a novel task called \textit{textual Distractors Generation for VQA} (DG-VQA). The goal of DG-VQA is to generate the most confusing distractors in multi-choice VQA tasks represented as a tuple of image, question, and the correct answer. Consequently, such distractors expose the vulnerability of neural models. We show that distractor generation can be formulated as a Markov Decision Process, and present a reinforcement learning solution to unsupervised produce distractors. Our solution addresses the lack of large annotated corpus issue in classical distractor generation methods. Our proposed model receives reward signals from well-trained multi-choice VQA models and updates its parameters via policy gradient. The empirical results show that the generated textual distractors can successfully confuse several cutting-edge models with an average 20% accuracy drop from around 64%. Furthermore, we conduct extra adversarial training to improve the robustness of VQA models by incorporating the generated distractors. The experiment validates the effectiveness of adversarial training by showing a performance improvement of 27% for the multi-choice VQA task