Representation Learning: A Review and New Perspectives

This paper reviews recent work in unsupervised feature learning and deep learning, covering advances in probabilistic models, autoencoders, manifold learning, and deep networks. It starts from the hypothesis that the success of machine learning algorithms generally depends on data representation because different representations can entangle and hide more or less the different explanatory factors of variation behind the data, and it notes that learning with generic priors can complement specific domain knowledge in designing representations.

Beyond surveying the field, the review motivates longer-term unanswered questions about the appropriate objectives for learning good representations, about computing representations (that is, inference), and about the geometrical connections between representation learning, density estimation, and manifold learning, framing the quest for AI as motivation for designing more powerful representation-learning algorithms.