Search Papers | Poster Sessions | All Posters
Poster A109 in Poster Session A - Tuesday, August 6, 2024, 4:15 – 6:15 pm, Johnson Ice Rink
Learning trajectories of deep neural networks during self-supervised visual representation learning
Ehsan Tousi1 (), Chelsea Kim1, Marieke Mur1; 1Western University, London ON, Canada
Recent work has started exploring the possibility of using self-supervised deep learning as a framework for modeling human visual development. Here, we provide a first step in that direction, by examining learning trajectories of a deep feedforward neural network, ResNet50, as it is trained on ImageNet using self-supervised contrastive learning. We ask if the learning trajectories show developmental signatures similar to those observed in the primate visual system. We show that representations change rapidly during the first few training epochs, and then stabilize. Like in the primate visual system, visual representations stabilize faster in early than in deep layers. Within- and between-category information emerge simultaneously, consistent with the notion that self-supervised contrastive learning promotes both. Our work provides preliminary support for using self-supervised deep learning to model human visual development, which opens up the possibility of systematically testing how developmental constraints shape visual representations.
Keywords: visual representation learning self-supervised learning deep neural networks