Technical Program

Paper Detail

Paper: PS-1B.36
Session: Poster Session 1B
Location: Symphony/Overture
Session Time: Thursday, September 6, 18:45 - 20:45
Presentation Time:Thursday, September 6, 18:45 - 20:45
Presentation: Poster
Paper Title: Unsupervised deep neural network for fMRI feedback modelling
Manuscript:  Click here to view manuscript
Authors: Michele Svanera, Andrew T. Morgan, Lucy S. Petro, Lars Muckli, University of Glasgow, United Kingdom
Abstract: The brain is constantly dealing with two streams of information: the feedforward stream which carries sensory inputs, and the feedback stream that contains predictions derived from internal models that the brain has about the world. During a brain imaging experiment, an effective method to study the feedback stream is to occlude a portion of the image presented, thereby isolating it from the feedforward signal. The predictive coding framework suggests that the brain is trying to reconstruct the image under the occlusion; this operation is conceptually similar to the image processing task of inpainting, in which an artificial model predicts the missing image part. Using an encoder/decoder network architecture, trained to fill occlusions and reconstruct an unseen image, we investigated similarities and contradictions between brain visual pathway and artificial neural networks. We will perform comparisons between brain data collected at 3T and 7T during the vision of static images and different layers of the encoder/decoder network. These analyses will be conducted using Representational Similarity Analysis (RSA) and by creating encoding models to investigate visual pathways and V1 layers. Understanding how information is integrated together in the early visual cortex will provide insight to fundamental neuroscientific questions about human vision, cognition, and perception.