Common Sense Bias Modeling for Classification Tasks

Machine learning model bias can arise from dataset composition: correlated sensitive features can disturb the downstream classification model's decision boundary and lead to performance differences along these features. Existing de-biasing works tackle most prominent bias features, like colors of digits or background of animals. However, a realworld dataset often includes a large number of feature correlations, that manifest intrinsically in the data as common sense information. Such spurious visual cues can further reduce model robustness. Thus, practitioners desire the whole picture of correlations and flexibility to treat concerned bias for specific domain tasks. With this goal, we propose a novel framework to extract comprehensive bias information in image datasets based on textual descriptions, a common sense-rich modality. Specifically, features are constructed by clustering noun phrase embeddings of similar semantics. Each feature's appearance across a dataset is inferred and their co-occurrence statistics are measured, with spurious correlations optionally examined by a human-in-the-loop interface. Downstream experiments show that our method discovers novel model biases on multiple image benchmark datasets. Furthermore, the discovered bias can be mitigated by a simple data re-weighting strategy that de-correlates the features, and outperforms state-of-the-art unsupervised bias mitigation methods. 92.4% images with "person" label Accuracy: 0.979 7.6% images without "person" Accuracy: 0.947