Ingo Fruend

The paper uses deep generative models to study human target detection in complex images, finding that biologically inspired models better predict performance than standard CNNs.

The paper quantifies human perception invariance to image contrast and content using a generative model, finding a factorial structure in discrimination thresholds.

This study uses GANs to constrain image perturbations within a natural image manifold and finds humans are more sensitive to such perturbations than to pixel-space changes.