Abstract
Contrast is one of the most important dimensions of the human visual experience. Although defining contrast is relatively straightforward for simple stimuli such as gratings or lines, it is much more difficult for naturalistic scenes with complex and often nonlinear structure. Generative image models based on deep neural networks, known as Generative Adversarial Nets (GANs), represent an image as a high dimensional vector of nonlinear latent features. Here, we explore how contrast is represented in these nonlinear latent dimensions. Based on previous observations, we hypothesized that contrast is related to the length of the latent feature vectors. We measured “contrast” sensitivity functions from four observers along this hypothesized contrast dimension. In a two-alternative forced-choice task, observers selected the image that appeared higher in contrast, between pedestal and target images. As the length of the pedestal stimulus’ feature vector increased, just noticeable differences increased approximately linearly with an average slope of 0.012+/−0.0023 (mean+/−sem) units per unit pedestal increment (R-square of 0.61+/−0.071). This indicates that perceived contrast is related to vector length in a GAN’s latent space through a Weber’s law. A logistic regression model using the full latent space vectors did not predict single trials better than a model that only had access to the lengths of the latent vectors. This confirmed that the feature vectors’ length was the main determinant of observer responses in our task. We conclude that contrast in complex and naturalistic scenes can be approximated by image representations based on deep neural networks.