Explicit Bias Discovery in Visual Question Answering Models
This work addresses the need for understanding and debugging biases in VQA models, which is important for researchers and practitioners in AI and computer vision, though it is incremental as it applies existing rule mining methods to a known issue.
The paper tackles the problem of statistical biases in Visual Question Answering (VQA) models by explicitly discovering these biases using rule mining on a database of question, answer, and visual words, revealing unique insights into model behavior and unusual learned behaviors.
Researchers have observed that Visual Question Answering (VQA) models tend to answer questions by learning statistical biases in the data. For example, their answer to the question "What is the color of the grass?" is usually "Green", whereas a question like "What is the title of the book?" cannot be answered by inferring statistical biases. It is of interest to the community to explicitly discover such biases, both for understanding the behavior of such models, and towards debugging them. Our work address this problem. In a database, we store the words of the question, answer and visual words corresponding to regions of interest in attention maps. By running simple rule mining algorithms on this database, we discover human-interpretable rules which give us unique insight into the behavior of such models. Our results also show examples of unusual behaviors learned by models in attempting VQA tasks.