everyone
since 13 Oct 2023">EveryoneRevisionsBibTeX
The ability to properly benchmark model performance in the face of spurious correlation is important to both build better predictors and increase confidence that models are operating as intended. We demonstrate that characterizing (as opposed to simply quantifying) model mistakes across subgroups is pivotal to properly reflect model biases, which are ignored by standard metrics such as worst-group accuracy or accuracy gap. Inspired by the hypothesis testing framework, we introduce SkewSize, a flexible metric that captures bias from mistakes in a model’s predictions. It can be used in multi-class settings or generalised to the open vocabulary setting of generative models. SkewSize is an aggregation of the effect size of the interaction between two categorical variables: the independent variable, representing the bias attribute (i.e. subgroup), and the dependent variable,representing the model’s prediction. We demonstrate the utility of SkewSize in multiple settings including: standard vision models trained on synthetic data, vision models trained on ImageNet as well as the DomainNet distribution shift benchmark, and large scale vision-language models from the BLIP-2 family. In each case, the proposed SkewSize is able to highlight biases not captured by other metrics, while also providing insights on the impact of recently proposed techniques, such as instruction tuning.