Unsupervised representational learning: the Helmholtzian perspective

Unsupervised learning algorithms essentially transform input examples into neural representations aiming to reveal interesting aspects of data. Such useful representations may have different properties emphasizing distinct characteristics of the data considered. Despite their uniqueness, these algorithms share common ties. Our perspective on unsupervised learning is that of Helmholtz´s approach to vision, considering learning as a minimization problem solved in the presence of a generative model inverting the process of creating representations. We elucidate this point of view by comparing and reviewing three models performing representational learning.