Discovering Latent Groups for Robust Classification

Machine learning models often rely on spurious correlations, resulting in high average accuracy but poor performance on underrepresented subgroups. To address this issue, researchers have developed methods that adjust network parameters based on subgroup annotations or inferred pseudo-group labels. However, these methods only provide class predictions at inference time, without revealing the sample's latent subgroup membership. A new approach focuses on discovering latent groups to improve robust classification. By uncovering these hidden subgroups, models can better capture nuanced patterns in the data, leading to more accurate and fair predictions. This is particularly important for high-stakes applications where biased models can have significant consequences¹. The ability to identify and account for latent subgroups can help mitigate these risks, making machine learning models more reliable and trustworthy. This matters to practitioners because it can help them develop more robust and fair models that perform well across diverse populations.