Learning deep representations by mutual information estimation and maximization
This work addresses the problem of learning effective representations without labeled data for machine learning practitioners, offering a novel approach that is not incremental but introduces a new paradigm.
The paper tackles unsupervised representation learning by maximizing mutual information between input and encoder output, incorporating locality knowledge and adversarial prior matching to improve downstream task performance. The method, Deep InfoMax (DIM), outperforms popular unsupervised methods and competes with fully-supervised learning on classification tasks.
In this work, we perform unsupervised learning of representations by maximizing mutual information between an input and the output of a deep neural network encoder. Importantly, we show that structure matters: incorporating knowledge about locality of the input to the objective can greatly influence a representation's suitability for downstream tasks. We further control characteristics of the representation by matching to a prior distribution adversarially. Our method, which we call Deep InfoMax (DIM), outperforms a number of popular unsupervised learning methods and competes with fully-supervised learning on several classification tasks. DIM opens new avenues for unsupervised learning of representations and is an important step towards flexible formulations of representation-learning objectives for specific end-goals.