Towards Transparent AI Systems: Interpreting Visual Question Answering Models

Deep neural networks have shown striking progress and obtained-of-the-art results in many AI research fields in the recent years. , it is often unsatisfying to not know why they predict what they do. In paper, we address the problem of interpreting Visual Question Answering (VQA) models. Specifically, we are interested in finding what part of the input (pixels in images or words in questions) the VQA model focuses on while the question. To tackle this problem, we use two visualization -- guided backpropagation and occlusion -- to find important words the question and important regions in the image. We then present qualitative quantitative analyses of these importance maps. We found that even without attention mechanisms, VQA models may sometimes be implicitly attending relevant regions in the image, and often to appropriate words in the.