Human processes in artificial vision

Abstract

From the earliest experiments with artificial neurons to the development of convolutional neural networks and beyond, nature has proven to be an inexhaustible source of inspiration. Not only can the mechanisms of vision in nature provide a template for the construction of new artificial models, but the analyses and tools developed over centuries of biological and psychological study can help to illuminate the nature of learned functions. In this work, we start by studying the extent to which convolutional networks make use of shape and colour cues when classifying images by analysing collections of synthetic images that maximise the models prediction for a target class (super-stimuli). Next, we explore the impact of convolutional neural network architecture on learned colour processing via the concept of opponency from the vision science literature. We show that the distribution of opponent cells and an analysis of cell tuning can provide valuable insight regarding the nature of colour processing. Following on from these findings, we consider the impact of opponency on adversarial robustness, finding that an increase in the percentage of opponent cells is not sufficient to consistently improve robustness to a suite of attacks. In the penultimate chapter, we introduce a convolutional model of foveation. We show that the addition of this foveated convolution improves the localisation performance of a visual attention model. We study the distribution of opponent cells and stimulus preference in these foveated layers as a function of eccentricity and find strong connections with findings from the biology literature regarding the nature of peripheral colour vision. In our final chapter, we construct a model of visual attention with the capacity to sketch its input, inspired by psychological concepts surrounding the production of visual memories. We demonstrate several interesting properties of this model, particularly regarding the drawing policies it learns and their connection to a notion of object.

More information

Identifiers

ORCID for Ethan William Albert Harris: orcid.org/0000-0003-3545-1349

ORCID for Jonathon Hare: orcid.org/0000-0003-2921-4283

Catalogue record

Export record