Explicit Bias Discovery in Visual Question Answering Models

By Varun Manjunatha, Nirat Saini, Larry S. Davis. 2018

99 citations recorded by the source corpus.

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.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Multimodal Learning, Question Answering & Reading Comprehension

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Multimodal Learning

In Question Answering & Reading Comprehension

Citations in the corpus

Benefits of Intermediate Annotations in Reading Comprehension (2020, background): “Unanticipated biases (Min et al., 2019; Manjunatha et al., 2019) are often introduced during dataset collection due to many reasons (eg., domain-speciﬁc contexts, crowd-workers distributions, etc.).”
Question and Answer Test-Train Overlap in Open-Domain Question Answering Datasets (2020, background): “…perform an analysis of machine comprehension dataset dif-ﬁculty, Kaushik and Lipton (2018) analyse the dif-ﬁculty of various machine reading datasets, and Manjunatha et al. (2018) show that visual question answering models memorize common question-answer relationships present in training data.”
Unbiased Scene Graph Generation From Biased Training (2020, neutral): “We didn’t use proxy vision-language tasks like captioning [65, 67] and VQA [54, 15] as the diagnosis, because their implementations have too many components unrelated to SGG and their datasets are challenged by their own biases [1, 14, 34].”
Commonsense Reasoning for Natural Language Understanding: A Survey of Benchmarks, Resources, and Approaches (2019, background): “Manjunatha et al. (2018) suggest that biases can further be avoided in VQA benchmarks by forcing questions to require a particular skill (e.g., telling time) to be answered, and this rule of thumb can be applicable for textual benchmarks as well.”
Counterfactual VQA: A Cause-Effect Look at Language Bias (2020, background): “Recent studies found that dataset bias, especially language bias, widely exists in vision-and-language tasks [14, 15, 24, 25, 26, 27, 28].”
MuMuQA: Multimedia Multi-Hop News Question Answering via Cross-Media Knowledge Extraction and Grounding (2021, unknown): “Simple baseline methods that only use question understanding (Kazemi and Elqursh 2017) or answer option sentiment analysis (Manjunatha, Saini, and Davis 2019) have been proven to perform surprisingly well on datasets such as VQA (Antol et al. 2015) and VQA2.”
Trends in Integration of Vision and Language Research: A Survey of Tasks, Datasets, and Methods (2019, background): “…works aim at interpretability/explainability by overcoming priors (Agrawal et al., 2018), concentrate better on the image to extract relevant information (Goyal et al., 2019), human-interpretable rules that gives better insights (Manjunatha et al., 2018), and cycleconsistency (Shah et al., 2019a).”
iPerceive: Applying Common-Sense Reasoning to Multi-Modal Dense Video Captioning and Video Question Answering (2020, methodology): “Prior work [12, 37] has unraveled the spurious observational bias that models fall prey to.”
RUBi: Reducing Unimodal Biases in Visual Question Answering (2019, background): “As a result, machine learning models tend to reﬂect these biases because they capture often undesirable correlations between the inputs and the ground truth annotations [30–32].”