Helmholtz Lecture Floris de Lange: Recurrent visual processing in the service of cognition

Donders Institute for Brain, Cognition and Behaviour, Radboud University, Nijmegen, NL

Within the first ~150 ms, a single feedforward sweep through the ventral stream can support rapid object recognition, and feedforward deep networks capture much of this. Yet the visual system is densely recurrent, raising a simple question: what is recurrent processing actually good for? In this talk I will present three new studies that each link recurrent visual processing to a different cognitive function.
First, using MEG and deep network encoding models, I will show that prior expectations speed up and sharpen neural representations of ambiguous objects, consistent with the idea that prior knowledge can shortcut the recurrent inference otherwise needed to resolve ambiguity. Second, using fMRI and backward masking, I will show that an image’s intrinsic memorability is reflected in ventral stream activity only when masking is delayed and recurrent processing can unfold, and that memorability likewise shapes recognition only under these conditions. Third, I will show that when objects are presented in the periphery, foveal early visual cortex comes to represent their task-relevant, but not task-irrelevant, features, in a format shared with directly viewed objects. Together, these findings suggest that recurrence contributes to cognition by transforming a transient sensory signal into a representation that is disambiguated by prior knowledge, encoded into memory, and available to a task-relevant visual workspace.