Beschreibung:
ABSTRACTWe examine the issue of bias in datasets designed to train visual question answering (VQA) algorithms. These datasets include a collection of natural language questions about images (aka ‐ visual questions). We consider three popular datasets that are captured by people with sight, people who are blind, and generated by computers. We first demonstrate that machine learning algorithms can be trained to recognize each dataset's bias, and so determine the source of a novel visual question. We then discuss potential risks and benefits of biased VQA datasets and corresponding machine learning algorithms that can identify the source of a visual question; e.g., whether it comes from a person with sight, a person who is blind, or bot (aka ‐ computer). Our ultimate aim is to inspire the development of more inclusive VQA systems.