Search Papers | Poster Sessions | All Posters
Poster B117 in Poster Session B - Thursday, August 8, 2024, 1:30 – 3:30 pm, Johnson Ice Rink
Deep Neural Network Models of Infant Visual Cortex
Cliona O'Doherty1,2 (), Áine T. Dineen1,2, Anna Truzzi1,2, Graham King1,2, Lorijn Zaadnoordijk1,2, Enna-Louise D'Arcy1,2, Jessica White1,2, Keelin Harrison1,2, Chiara Caldinelli1,2, Tamrin Holloway1,2, Anna Kravchenko1,2, Ailbhe Tarrant4,6, Angela T. Byrne5,7, Adrienne Foran4,6, Eleanor J. Molloy3,5,7, Rhodri Cusack1,2; 1Trinity College Institute of Neuroscience, Trinity College Dublin, 2School of Psychology, Trinity College Dublin, 3Paediatrics and Child Health, Trinity College Dublin, 4The Rotunda Hospital, Dublin, 5The Coombe Hospital, Dublin, 6Children's Health Ireland at Temple Street, 7Children's Health Ireland at Crumlin
Deep convolutional neural networks (DNNs) are now cemented as effective computational models in adult visual neuroscience. However, comparing the learning human brain to the learning models had not yet been possible due to the difficulty in collecting sufficient neuroimaging data from infants. To address this, we conducted longitudinal fMRI on 2-month-old infants (n=130), and again at 9-months-old (n=65), while they were awake and viewing a variety of visual stimuli. Multivariate pattern analysis (MVPA) revealed a complex representational structure in visual cortex already at 2-months. We show that fully-trained DNNs capture a significant proportion of this structure, and different learning algorithms can determine the developmental stage that a DNN best explains.
Keywords: deep neural networks infants ventral stream fMRI