Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster A in Poster Session A: Tuesday, August 4, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Prenatal retinal waves generate the proto-architecture of newborn visual systems

Lalit Pandey¹, Samantha Marie Waters Wood¹, Justin Newell Wood¹; ¹Indiana University

Presenter: Samantha Marie Waters Wood

How does the visual system acquire its structure? Newborn visual systems share three signatures of visual organization: orientation selectivity in early visual areas, shape selectivity in downstream regions, and progressively larger receptive fields across the visual hierarchy. Instructional theories attribute this proto-architecture to genetic design, whereas selectional theories predict the proto-architecture develops through domain-general adaptation to prenatal sensory experiences. Here we provide an image-computable test of these theories using transformers: a domain-general neural network model. We trained transformers on prenatal retinal waves using unsupervised temporal learning, and found that the models spontaneously developed all three signatures of newborn visual systems. These results show that the large-scale architecture of newborn visual systems emerges when domain-general neural networks adapt to prenatal sensory experiences, in the absence of genetically encoded design. The distinctive visual cortical architecture of newborn brains is a statistical consequence of high-dimensional systems adapting to prenatal experiences.

Topic Area: Computational Models of Vision & Visual Cortex