What do AI Models Know? A Case Study on Visual Question Answering
A major challenge in recent AI literature is understanding why state-of-the-art deep learning models show great success on a range of datasets while they severely degrade in performance when presented with examples which slightly vary from their training distribution. In this proposal, we will examine this question in the context of visual question-answering, a challenging task which requires models to jointly reason over images and text. We will start our exploration with the GQA dataset, which, along with images and text, also includes a rich semantic scene graph, representing the spatial relations between objects in the image, and thus lends itself to probing through high-quality automatic manipulation. In prelminary work we have augmented GQA with examples that vary slightly from the original questions, and shown that here too high-performing models perform much worse on the augmented questions compared to the original ones. Our proposal will analyze our results, exploring the reasons for the drop in performance, and what makes our new questions more challenging. We also plan to to generalize the reasons we find to other datasets of visual question answering, and more broadly to other AI datasets. Given those insights, we will augment the training set with instances that capture model “blind spots”, in an attempt to improve the model’s generalization ability. Our results will improve our understanding of what state-of-the-art AI models know, what they are still missing, and how can we improve them based on this new understanding.
To achieve a better understanding of the limitations of current state-of-the-art models and datasets, as well as ways to improve them.
What we did