Existing models often leverage co-occurrences between objects and their\ncontext to improve recognition accuracy. However, strongly relying on context\nrisks a model's generalizability, especially when typical co-occurrence\npatterns are absent. This work focuses on addressing such contextual biases to\nimprove the robustness of the learnt feature representations. Our goal is to\naccurately recognize a category in the absence of its context, without\ncompromising on performance when it co-occurs with context. Our key idea is to\ndecorrelate feature representations of a category from its co-occurring\ncontext. We achieve this by learning a feature subspace that explicitly\nrepresents categories occurring in the absence of context along side a joint\nfeature subspace that represents both categories and context. Our very simple\nyet effective method is extensible to two multi-label tasks -- object and\nattribute classification. On 4 challenging datasets, we demonstrate the\neffectiveness of our method in reducing contextual bias.\n