Search Papers | Poster Sessions | All Posters
Poster C5 in Poster Session C - Friday, August 9, 2024, 11:15 am – 1:15 pm, Johnson Ice Rink
Fast and robust visual recognition young children
Vladislav Ayzenberg1 (), Stella Lourenco2; 1University of Pennsylvania, 2Emory University
By adulthood, humans can rapidly identify objects from sparse visual displays and do so across disruptions to the object’s appearance. However, little is known about the development of these abilities. Here, we examined the robustness of children’s (3 to 5 years) recognition abilities using a challenging object recognition task which required them to identify rapidly presented objects (100 - 300 ms; forward and backward masked) that had complete, perturbed, or deleted contours. To shed light on the mechanisms underlying their recognition abilities, we compared their performance to biologically plausible deep neural networks (DNNs) with feedforward or recurrent architectures which were trained with either curated or variable image sets. We also characterized the gaps between child and machine vision by comparing children to performance optimized models. We found that even the youngest children could identify objects at high speeds when object contours were perturbed or deleted. Analyses of DNN performance revealed that both recurrence and variable visual experience were crucial for improving recognition accuracy, though they generally performed worse than children. These findings suggest that young children’s visual recognition abilities are fast and robust, but the mechanisms underlying these abilities are not understood well enough to implement into current models.
Keywords: Development Deep Learning Object Classification Recurrence