- Abstract: One important aspect of generalization in machine learning involves reasoning about previously seen data in new settings. Such reasoning requires learning disentangled representations of data which are interpretable in isolation, but can also be combined in a new, unseen scenario. To this end, we introduce the context-aware learner, a model based on the variational autoencoding framework, which can learn such representations across data sets exhibiting a number of distinct contexts. Moreover, it is successfully able to combine these representations to generate data not seen at training time. The model enjoys an exponential increase in representational ability for a linear increase in context count. We demonstrate that the theory readily extends to a meta-learning setting such as this, and describe a fully unsupervised model in complete generality. Finally, we validate our approach using an adaptation with weak supervision.