Schedule of Events | Search Abstracts | Invited Symposia | Symposia | Poster Sessions | Data Blitz

Graduate Student Award Winner 

Awake Infant fMRI and Deep Neural Network Models Reveal Hierarchical Development of Object Feature Tuning in the Ventral Visual Stream

Poster Session D - Monday, March 9, 2026, 8:00 – 10:00 am PDT, Fairview/Kitsilano Ballroom
Also presenting in Data Blitz Session 2 - Saturday, March 7, 2026, 10:30 am – 12:00 pm PST, Salon D.

Áine T. Dineen¹ (), Clíona O'Doherty^1,2, Anna Truzzi^1,3, Anna Kravchenko¹, Graham King¹, Alex R. Wade⁴, Lorijn Zaadnoordijk¹, Enna-Louise D’Arcy¹, Tamrin Holloway¹, Jessica White¹, Chiara Caldinelli¹, Angela T. Byrne^5,6, Ailbhe Tarrant^7,8, Adrienne Foran^7,8, Eleanor J. Molloy^1,5, Rhodri Cusack¹; ¹Trinity College Dublin, ²Stanford University, ³Queen’s University Belfast, ⁴University of York, ⁵The Coombe Hospital, Dublin, Ireland, ⁶Children's Health Ireland at Crumlin, Dublin, Ireland, ⁷The Rotunda Hospital, Dublin, Ireland, ⁸Children's Health Ireland at Temple Street, Dublin, Ireland

When an adult views an object, the ventral visual stream (VVS) encodes features across spatial scales, from fine details to global shape, and a range of complexities, from low-level perceptual features to conceptual properties. Due to methodological challenges, when and how object feature tuning emerges during development remains unknown. Behavioral studies indicate that young infants attend to low-level visual features, but cannot reveal which features are represented along the VVS. Through the Foundations of Cognition Project, we therefore conducted the largest cross-sectional and first longitudinal awake infant fMRI study to date, acquiring object responses at 2-months (n=113) and 9-months (n=51), and from adults (n=17). To probe feature tuning, deep neural networks (DNNs) were trained on objects blurred at varying levels, suppressing fine detail and yielding models with graded sensitivity to features across spatial scales. Representational similarity analysis was used to characterise group-level responses along the developing VVS and compare them to model responses across blur levels. For each comparison, the model layer whose representations best matched the brain response was selected, capturing the appropriate level of complexity. By 2-months, V1 responses resembled adult-like tuning. In contrast, later VVS regions remained immature through 9-months, demonstrating a hierarchical trajectory of development. Across development, V1 responses were best modelled by intermediate layers, consistent with previous work suggesting that V1 is sensitive to more complex features in addition to low-level ones. That such complexity appears to be present by 2-months indicates surprisingly sophisticated early sensory representations and motivates further work to probe their content.

Topic Area: PERCEPTION & ACTION: Development & aging