MindSet: A Bias-Detection Interface Using a Visual Human-in-the-Loop Workflow

Handling data artifacts is a critical and unsolved challenge in deep learning. Disregarding such asymmetries may lead to biased and socially unfair predictions, prohibiting applications in high-stake scenarios. In the case of visual data, its inherently unstructured nature makes automated bias detection especially difficult. Thus, a promising remedy is to rely on human feedback. Hu et al. [14] introduced a three-stage theoretical study framework to use a human-in-the-loop approach for bias detection in visual datasets and ran a small-sample study. While showing encouraging results, no implementation is available to enable researchers and practitioners to study their image datasets. In this work, we present a dataset-agnostic implementation based on a highly flexible web app interface. With this implementation, we aim to bring this theoretical framework into practice by following a user-centric approach. We also extend the framework so that the workflow can be adjusted to the researcher’s needs in terms of the granularity of detected anomalies.