SQuINTing at VQA Models: Introspecting VQA Models With Sub-Questions

Selvaraju, Ramprasaath R.; Tendulkar, Purva; Parikh, Devi; Horvitz, Eric; Ribeiro, Marco Túlio; Nushi, Besmira; Kamar, Ece

doi:10.1109/cvpr42600.2020.01002

Cited by 31 publications

(30 citation statements)

References 15 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…A crucial challenge all these models have to face is the ability to generalize the knowledge learned to unseen data, which can be achieved only if the model is able to compositionally build the multimodal representations, a must for any model of human intelligence (Lake et al., 2017). Since neural‐based VQA models have been shown to produce inconsistent answers to questions that are either similar or mutually exclusive, approaches to mitigate this behaviour have been recently proposed (Ray et al., 2019; Selvaraju et al., 2020). Interestingly, Gandhi and Lake (2020) showed that whilst children are driven by the mutual exclusivity assumption in their learning process, neural networks are not, and set this as an open challenge.…”

Section: The Recent Revival Of Vqamentioning

confidence: 99%

Linguistic issues behind visual question answering

Bernardi

Pezzelle

2021

Language and Linguist. Compass

View full text Add to dashboard Cite

Answering a question that is grounded in an image is a crucial ability that requires understanding the question, the visual context, and their interaction at many linguistic levels: among others, semantics, syntax and pragmatics. As such, visually‐grounded questions have long been of interest to theoretical linguists and cognitive scientists. Moreover, they have inspired the first attempts to computationally model natural language understanding, where pioneering systems were faced with the highly challenging task—still unsolved—of jointly dealing with syntax, semantics and inference whilst understanding a visual context. Boosted by impressive advancements in machine learning, the task of answering visually‐grounded questions has experienced a renewed interest in recent years, to the point of becoming a research sub‐field at the intersection of computational linguistics and computer vision. In this paper, we review current approaches to the problem which encompass the development of datasets, models and frameworks. We conduct our investigation from the perspective of the theoretical linguists; we extract from pioneering computational linguistic work a list of desiderata that we use to review current computational achievements. We acknowledge that impressive progress has been made to reconcile the engineering with the theoretical view. At the same time, we claim that further research is needed to get to a unified approach which jointly encompasses all the underlying linguistic problems. We conclude the paper by sharing our own desiderata for the future.

show abstract

Section: The Recent Revival Of Vqamentioning

confidence: 99%

Linguistic issues behind visual question answering

Bernardi

Pezzelle

2021

Language and Linguist. Compass

View full text Add to dashboard Cite

show abstract

“…Using counterfactual images as explanations can also be thought of as the visual equivalent to observing VQA behavior by rephrasing the question and checking if the model responds consistently. [20][21][22] Hence, such counterfactual images hint at how consistent these models are to users, and that aids in their mental model improvement.…”

Section: Related Workmentioning

confidence: 99%

Improving users' mental model with attention‐directed counterfactual edits

et al. 2021

View full text Add to dashboard Cite

In the domain of visual question answering (VQA), studies have shown improvement in users' mental model of the VQA system when they are exposed to examples of how these systems answer certain image‐question (IQ) pairs. In this work, we show that showing controlled counterfactual IQ examples are more effective at improving the mental model of users as compared to simply showing random examples. We compare a generative approach and a retrieval‐based approach to show counterfactual examples. We use recent advances in generative adversarial networks to generate counterfactual images by deleting and inpainting certain regions of interest in the image. We then expose users to changes in the VQA system's answer on those altered images. To select the region of interest for inpainting, we experiment with using both human‐annotated attention maps and a fully automatic method that uses the VQA system's attention values. Finally, we test the user's mental model by asking them to predict the model's performance on a test counterfactual image. We note an overall improvement in users' accuracy to predict answer change when shown counterfactual explanations. While realistic retrieved counterfactuals obviously are the most effective at improving the mental model, we show that a generative approach can also be equally effective.

show abstract

“…They have also been used in an optimization process where 19 proposed a loss function to find the minimum changes in the input that results in a change in the output of a classifier. Using counterfactual images as explanations can also be thought of as the visual equivalent to observing VQA behavior by rephrasing the question and checking if the model responds consistently 20,21,22 . Hence, such counterfactual images hint at how consistent these models are to users, and that aids in their mental model improvement.…”

Section: Related Workmentioning

confidence: 99%

Improving Users' Mental Model with Attention-directed Counterfactual Edits

Alipour¹,

Ray²,

Lin³

et al. 2021

Preprint

View full text Add to dashboard Cite

In the domain of Visual Question Answering (VQA), studies have shown improvement in users' mental model of the VQA system when they are exposed to examples of how these systems answer certain Image-Question (IQ) pairs 1 . In this work, we show that showing controlled counterfactual image-question examples are more effective at improving the mental model of users as compared to simply showing random examples. We compare a generative approach and a retrieval-based approach to show counterfactual examples. We use recent advances in generative adversarial networks (GANs) to generate counterfactual images by deleting and inpainting certain regions of interest in the image. We then expose users to changes in the VQA sys-tem's answer on those altered images. To select the region of interest for inpainting, we experiment with using both human-annotated attention maps and a fully automatic method that uses the VQA system's attention values. Finally, we test the user's mental model by asking them to predict the model's performance on a test counterfactual image. We note an overall improvement in users' accuracy to predict answer change when shown counterfactual explanations. While realistic retrieved counterfactuals obviously are the most effective at improving the mental model, we show that a generative approach can also be equally effective.

show abstract

SQuINTing at VQA Models: Introspecting VQA Models With Sub-Questions

Cited by 31 publications

References 15 publications

Linguistic issues behind visual question answering

Linguistic issues behind visual question answering

Improving users' mental model with attention‐directed counterfactual edits

Improving Users' Mental Model with Attention-directed Counterfactual Edits

Contact Info

Product

Resources

About