Deep neural networks are susceptible to learn biased models with entangled\nfeature representations, which may lead to subpar performances on various\ndownstream tasks. This is particularly true for under-represented classes,\nwhere a lack of diversity in the data exacerbates the tendency. This limitation\nhas been addressed mostly in classification tasks, but there is little study on\nadditional challenges that may appear in more complex dense prediction problems\nincluding semantic segmentation. To this end, we propose a model-agnostic and\nstochastic training scheme for semantic segmentation, which facilitates the\nlearning of debiased and disentangled representations. For each class, we first\nextract class-specific information from the highly entangled feature map. Then,\ninformation related to a randomly sampled class is suppressed by a feature\nselection process in the feature space. By randomly eliminating certain class\ninformation in each training iteration, we effectively reduce feature\ndependencies among classes, and the model is able to learn more debiased and\ndisentangled feature representations. Models trained with our approach\ndemonstrate strong results on multiple semantic segmentation benchmarks, with\nespecially notable performance gains on under-represented classes.\n