Hosted by the Interactive AI CDT and the Intelligent Systems Laboratory
Event is 13.00-13.30 (pizza) followed by talks from 13:30–16.00.
Speaker:Professor Valero Laparra, Universitat de València
Talk Title: RBIG: an algorithm to convert any data distribution into a standard Gaussian.
Talk abstract: In this talk, I will present the Rotation-Based Iterative Gaussianization algorithm. It is a simple method that, given a data sample, finds a transformation that converts the data into a space where they have a Gaussian distribution. Since the transformation is invertible and differentiable, it allows multiple applications such as generating synthetic data or computing information theory measures in a simple way. The transformation is based on iteratively applying two simple steps: marginal Gaussianization + rotation. The algorithm for finding the appropriate steps has no parameter setting and is guaranteed to converge. We will see some applications and a simple practical demonstration.
Speaker: Alexander Hepburn
Talk Title: On the relation between statistical learning and perceptual distances.
Talk abstract: In this talk, we aim to unravel the non-trivial relationships between the probability distribution of the data, perceptual distances, and unsupervised machine learning. To this end, we show that perceptual sensitivity is correlated with the probability of an image in its close neighborhood. We also explore the relation between distances induced by autoencoders and the probability distribution of the training data, as well as how these induced distances are correlated with human perception. Finally, we find perceptual distances do not always lead to noticeable gains in performance over Euclidean distance in common image processing tasks, except when data is scarce and the perceptual distance provides regularization. We propose this may be due to a double-counting effect of the image statistics, once in the perceptual distance and once in the training procedure.