Dr Kate Storrs (University of Auckland)

Abstract

Computational visual neuroscience has come a long way in the past 10 years. Deep neural networks can recognise objects with near-human accuracy, and predict brain activity in the ventral visual cortex better than any previous models. However, vision is far from explained. Our most successful models have been supervised to recognise objects in images using ground-truth labels for millions of examples. Brains have no such access to the ground truth, and must instead learn directly from sensory data. Unsupervised deep learning, in which networks learn statistical regularities in their data by compressing, extrapolating or predicting images and videos, presents a more ecologically feasible alternative. We have been using unsupervised deep learning, combined with computer-rendered artificial environments and psychophysics experiments, as a framework to understand how brains learn rich scene representations without ground-truth information about the world. I will explore how unsupervised networks trained on environments of 3D rendered objects with varying shape, material and illumination, spontaneously come to encode these properties of the environment in their internal representations. More strikingly, they can predict, on an image-by-image basis, patterns of errors made by human observers. Unsupervised deep learning may provide a powerful framework for exploring how perceptual dimensions and categories arise.