Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session B: Wednesday, August 13, 1:00 – 4:00 pm, de Brug & E‑Hall

Recurrent processing strengthens feature representations throughout visual cortex

Lea-Maria Schmitt¹, Floris P De Lange²; ¹Donders Institute , ²Radboud University

Presenter: Floris P De Lange

Visual processing begins with a feedforward sweep that creates an initial perceptual representation, which is then refined through recurrent (lateral and feedback) processing. While recurrent signals in visual regions are abundant, their functional role in perceptual inference remains largely unclear. Using functional magnetic resonance imaging (fMRI) and artificial neural network (ANN) modeling, we aimed to examine whether recurrence modulates how images are represented in early and late visual regions. Participants were briefly presented with novel ambiguous images that were challenging to categorize. A visual mask followed these images either immediately or after a delay, thereby blocking or allowing for recurrent processing. We found that activity in early visual regions was best encoded by early layers of a convolutional neural network. These representations could no longer be observed when images were immediately masked. Conversely, activity in later ventral and dorsal visual regions was best encoded by later layers, and remained robust in the immediate masking condition. Comparing the brain alignment of ANNs with different recurrent dynamics revealed that activity in later ventral regions under delayed masking was best explained by a model with both lateral and feedback recurrence, suggestive of a role for recurrence in perceptual inference. The general strengthening of feature-specific representations with delayed masking likely reflects an interplay between early and late visual cortex, where lateral recurrence may help denoise low-level features to form more accurate high-level interpretations, which in turn may help disambiguate low-level features through feedback.

Topic Area: Visual Processing & Computational Vision

proceeding: Full Text on OpenReview