Abstract
Visual recognition is achieved in the brain by a series of hierarchical processing stages across areas of the ventral stream. Decoding object identity in this hierarchy is partially a bottom-up process, which accounts well for the recognition of easy, canonical appearances. Most natural settings, however, are visually complex, and present challenging viewing conditions. In the presence of visual uncertainty, the recognition of object identity is hence a hard decoding problem that cannot be solved by a purely bottom-up process. The human brain, however, is highly non-feedforward, and each area sends its connections downward or horizontally in the hierarchy. This non bottom-up, recurrent connectivity is thought to provide the necessary computational power to solve object recognition in the face of visual complexity. While evidence for this role is accumulating, the exact mechanisms of recurrent processing remain obscure. In this study, we investigated how recurrence in the visual cortex serves object recognition. A stimulus set was built that replicates many of the visual complexities found in the natural world. We carefully implemented several manipulations (occlusion, phase scrambling, clutter) linked to recurrent processing within a single experimental paradigm. We collected psychophysical data on a large sample (n=200) of participants in an online study. Results from a classification task confirm the effect of our manipulations. Additionally, using backward masking, we were able to determine among visual challenges which ones require more recurrence to be resolved. These results correlate with the behaviour of recurrent artificial neural networks, providing strong evidence that the manipulations we implemented are linked to recurrent processing. Overall, our data support the idea that different recurrent processes in the brain have different functions in object recognition.