This thesis presents an analysis of a novel, deep learning-based approach for user verification. Face verification is the task of comparing a candidate face to another and verifying whether it is a match. The traditional approach consists of relying on analytical metrics to shape the classification boundary. Instead of defining a metric, our approach allows the network to inherently learn it by mapping matching and non-matching face pairs onto different statistical distributions. Although any class of target distributions can be applied, using the Gaussians is a logical choice since the natural output of large enough fully-connected layers comes to be Gaussian.