Conference on Cognitive Computational Neuroscience

Visual illusions are the subjective experiences that deviate from physical features of stimuli, which are vital for understanding perception and brain. Although neural units that correspond similarly to illusory and real lines were discovered, there remains a gap between local neural activations and global percepts. To advance our understanding of how visual areas represent line illusions, we sought to reconstruct illusory percepts as images from brain activity. We leveraged deep neural network (DNN), a powerful tool for extracting image features represented by DNN unit activations, and trained linear decoders to predict DNN features of natural images using brain responses to the same images, assuming that stimulus features unambiguously correspond to the perceptual features encoded in the brain activity. The trained decoders were applied to brain responses to illusion-inducing abutting lines to obtain brain-decoded features, which were then passed to an image generator for reconstruction. Both inducer and illusory lines were reconstructed from brain-decoded features, whereas only inducer lines were reconstructed from stimulus features. The proportions of principal orientations in reconstructions that matched illusory orientation were highest in V2/V3, followed by V1/V4, and then higher areas. These results demonstrate the success of illusory line reconstruction and region-specific representations of line illusion.