Brain-optimized neural networks learn non-hierarchical models of representation in human visual cortex

Deep neural networks (DNNs) trained to perform visual tasks learn representations that align with the hierarchy of visual areas in the primate brain. This finding has been taken to imply that the primate visual system forms representations by passing them through a hierarchical sequence of brain areas, just as DNNs form representations by passing them through a hierarchical sequence of layers. To test the validity of this assumption, we optimized DNNs not to perform visual tasks but to directly predict brain activity in human visual areas V1-V4. Using a massive sampling of human brain activity, we constructed brain-optimized networks that predict brain activity even more accurately than task-optimized networks. We show that brain-optimized networks can learn representations that diverge from those formed in a strict hierarchy. Brain-optimized networks do not need to align representations in V1-V4 with layer depth; moreover, they are able to accurately model anterior brain areas (e.g., V4) without computing intermediary representations associated with posterior brain areas (e.g., V1). Our results challenge the view that human visual areas V1-V4 act--like the early layers of a DNN--as a serial pre-processing sequence for higher areas, and suggest they may subserve their own independent functions.