Alan Blair (Senior Lecturer @CSE)

Autoencoders, Bidirectional GANs and Deep Principal Component Analysis using a Plum Pudding Loss Function

Abstract

There are many situations in deep learning where we wish to force a set of vectors or activations to approximate either a hyperspherical or multivariate Gaussian distribution. For example, deep autoencoders aim to map a set of real images or other data to latent variables, enabling the reconstruction of the original images as well as generation of similar images from random latent variables.

In this talk, we show how this type of regularization can be achieved using a novel loss function inspired by Thomson’s Plum Pudding model of the atom, which simulates a repulsive force between every pair of items, and an attractive force of each item toward the origin. We discuss how this loss function can be applied to Autoencoders and Bidirectional Generative Adversarial Networks, and how it may potentially provide a foundation for deep principal component analysis

Bio

Dr Alan Blair (B.Sc. USyd, Ph.D. Mathematics MIT) is a senior lecturer in the School of Computer Science and Engineering, UNSW. He has previously held research positions at the University of Melbourne, the University of Queensland and Brandeis University. Alan’s research interests include self-learning for strategic games, robot navigation, language processing, convolutional network architectures and training, adversarial and coevolutionary dynamics, multi-task learning, hierarchical evolution and computational creativity.