Figure, ground, common ground

Matching human performance on real-world object recognition has been a longstanding goal in computer vision. Progress in computer vision models intrigued cognitive neuroscientists, who discovered that the representations in vision models and human visual cortex corresponded hierarchically. However, it's unclear which visual components are represented in this correspondence. In this dissertation, I use human EEG responses together with networks varying in architecture and training data to isolate the visual components shared by both systems. Brains and neural networks are representationally aligned in the early stages of visual processes but diverge at mid-level stages. The alignment of both systems concentrates largely on the global summaries of local image statistics: edges, lines, textures, and background structure. The common ground comes largely from the literal (back)ground of the image. Thus, both brains and neural networks may see the same fur, though not necessarily the same cat.