VisQA: X-raying Vision and Language Reasoning in Transformers.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Human-Computer Interaction

Background:

• Visual Question Answering (VQA) models are crucial for HCI tasks like assisting the visually impaired.
• Current VQA models often rely on data biases rather than genuine reasoning.
• This reliance on bias hinders reliable model performance and interpretability.

Purpose of the Study:

• To introduce VisQA, a visual analytics tool for exploring bias exploitation versus reasoning in VQA models.
• To investigate the role of attention maps in understanding neural model decision-making.
• To provide insights into improving VQA model design and training.

Main Methods:

• Developed VisQA, a visual analytics tool focusing on transformer attention maps.
• Motivated by known bias examples in deep learning and vision-language reasoning.
• Evaluated VisQA through interdisciplinary collaboration and expert analysis of model decision processes.

Main Results:

• VisQA effectively exposes attention distributions, correlating them with model predictions.
• The study led to a better understanding of bias exploitation in VQA.
• A method for transferring reasoning patterns from an oracle model was proposed.

Conclusions:

• VisQA makes the inner workings of VQA models more accessible to users.
• The tool aids in identifying and mitigating bias in VQA systems.
• Understanding attention mechanisms is key to developing more robust and trustworthy VQA models.