This paper introduces Deep InfoMax (DIM), a method for unsupervised learning of representations by maximizing mutual information (MI) between input and encoder output. The authors show that incorporating knowledge about locality in the input can significantly improve the quality of representations for downstream tasks. DIM also controls representation characteristics by matching to a prior distribution adversarially. The method outperforms several popular unsupervised learning methods and performs competitively with fully-supervised learning on classification tasks. DIM enables new avenues for unsupervised representation learning and is an important step towards flexible formulations of representation learning objectives for specific end-goals. The paper explores the use of mutual information maximization for representation learning, showing that maximizing MI between the complete input and encoder output (global MI) is often insufficient. Instead, maximizing the average MI between the representation and local regions of the input (e.g., patches) can greatly improve the representation's quality for tasks like classification. DIM also incorporates prior matching to constrain representations according to desired statistical properties, similar to adversarial autoencoders. The authors propose DIM, which combines MI maximization with prior matching. They use adversarial learning to constrain the representation to have desired statistical characteristics specific to a prior. DIM is evaluated on several image datasets, including CIFAR10, CIFAR100, Tiny ImageNet, and STL-10. The results show that DIM outperforms other unsupervised methods on these datasets, particularly when using a local MI objective. DIM also performs competitively with fully-supervised learning on some tasks. The paper also discusses related work, including other methods for representation learning such as variational autoencoders, adversarial autoencoders, and contrastive predictive coding. The authors show that DIM can leverage local structure in the input to improve the suitability of representations for classification. They also introduce two new measures of representation quality, one based on mutual information neural estimation and a neural dependency measure, to evaluate DIM against other unsupervised methods.This paper introduces Deep InfoMax (DIM), a method for unsupervised learning of representations by maximizing mutual information (MI) between input and encoder output. The authors show that incorporating knowledge about locality in the input can significantly improve the quality of representations for downstream tasks. DIM also controls representation characteristics by matching to a prior distribution adversarially. The method outperforms several popular unsupervised learning methods and performs competitively with fully-supervised learning on classification tasks. DIM enables new avenues for unsupervised representation learning and is an important step towards flexible formulations of representation learning objectives for specific end-goals. The paper explores the use of mutual information maximization for representation learning, showing that maximizing MI between the complete input and encoder output (global MI) is often insufficient. Instead, maximizing the average MI between the representation and local regions of the input (e.g., patches) can greatly improve the representation's quality for tasks like classification. DIM also incorporates prior matching to constrain representations according to desired statistical properties, similar to adversarial autoencoders. The authors propose DIM, which combines MI maximization with prior matching. They use adversarial learning to constrain the representation to have desired statistical characteristics specific to a prior. DIM is evaluated on several image datasets, including CIFAR10, CIFAR100, Tiny ImageNet, and STL-10. The results show that DIM outperforms other unsupervised methods on these datasets, particularly when using a local MI objective. DIM also performs competitively with fully-supervised learning on some tasks. The paper also discusses related work, including other methods for representation learning such as variational autoencoders, adversarial autoencoders, and contrastive predictive coding. The authors show that DIM can leverage local structure in the input to improve the suitability of representations for classification. They also introduce two new measures of representation quality, one based on mutual information neural estimation and a neural dependency measure, to evaluate DIM against other unsupervised methods.