Human vision remains remarkably robust even under challenging perceptual conditions, such as when object categories are ambiguous. However, the precise role of top-down feedback in shaping neural representations along the visual ventral pathway, and the nature of the information it conveys, remains poorly understood. In this study, we combine functional magnetic resonance imaging (fMRI) with artificial neural networks (ANNs) to investigate how feedback refines visual representations.

Participants viewed 180 images (repeated 4 times) selected from ImageNet while undergoing fMRI scanning. Half of the images were adversarially perturbed to challenge visual recognition. To manipulate feedback processing, each image was followed by either an immediate or delayed backward mask, allowing or disrupting feedback signals, respectively. We used layer-wise features from a convolutional neural network (CNN) to model voxel-wise neural responses. Preliminary results from representational similarity analysis (RSA) and encoding models reveal that allowing feedback (via delayed masking) enhances alignment between CNN features and neural representations, not only in early visual areas but also in higher regions of the ventral stream.