This paper introduces Deep InfoMax (DIM), a method for unsupervised learning of representations by maximizing mutual information (MI) between the input and the output of a deep neural network encoder. The authors emphasize the importance of incorporating structural information, such as locality in the input, into the objective function to improve the quality of learned representations for downstream tasks. DIM combines MI maximization with prior matching using adversarial learning to constrain the representation to have desired statistical properties. The method is evaluated on several classification tasks and is shown to outperform or compete with popular unsupervised learning methods and fully-supervised learning. DIM opens new avenues for unsupervised representation learning and is a significant step towards flexible formulations of representation learning objectives for specific end-goals.This paper introduces Deep InfoMax (DIM), a method for unsupervised learning of representations by maximizing mutual information (MI) between the input and the output of a deep neural network encoder. The authors emphasize the importance of incorporating structural information, such as locality in the input, into the objective function to improve the quality of learned representations for downstream tasks. DIM combines MI maximization with prior matching using adversarial learning to constrain the representation to have desired statistical properties. The method is evaluated on several classification tasks and is shown to outperform or compete with popular unsupervised learning methods and fully-supervised learning. DIM opens new avenues for unsupervised representation learning and is a significant step towards flexible formulations of representation learning objectives for specific end-goals.