Grant: Master's Award: Canada Graduate Scholarships

Humans have the ability to perform fine tuned discrimination between different natural images. Previous studies exploring visual discrimination often used low-level grating stimuli due to their ease of manipulation. However gratings are often quite different from real world images, that contain complex and nonlinear structure. Recent advances in artificial intelligence have given rise to generative image models based on deep neural networks, known as Generative Adversarial Nets (GANs). GANs provide an avenue for the parametric manipulation of naturalistic images, allowing for the parametric study of higher level human perception. My previous study focused on how algebraic properties of the GANs latent space are related to objective human discrimination performance. The proposed study will further this by examining the topological properties of the simulated images and understanding how the aspects learned by a GAN are related to subjective qualities of human experience. During the relatively short time GANs have been present, one of the main questions in the field has been how these parameterization can be evaluated against human perception. The proposed study is a novel attempt to draw a connection between GAN parameters and underlying human processes using a multidimensional psychophysical scaling method.