Characterizing and Modeling Infants' Self-Generated Object Views: Implications for Object Recognition and Language Learning

Trustees Of Indiana University, Bloomington IN

Investigators

Linked publications & trials

Abstract

PROJECT SUMMARY/ABSTRACT Human visual object recognition is foundational to many achievements?from object name learning to tool use to real world problem solving. Understanding the developmental processes that underlie visual object recognition is of critical importance because the individual differences that characterize early visual object recognition have clinical, educational, and societal implications. For instance, toddlers with poor visual object recognition skills are more likely to have below-average vocabulary sizes. Toddlers with smaller vocabularies have a greater likelihood of developing language impairments and are more likely to lag in pre-literacy and literacy skills. Deficits in visual object recognition and word learning have also been exhibited by individuals with Autism Spectrum Disorders (ADSs). The overarching goal of the proposed project is to better understand the sensory-motor mechanisms that support visual object recognition. Considerable evidence suggests that active object manipulation relates to better visual object recognition, however little is known about the mechanisms through which object manipulation connects to visual object processing during development. The proposed research tests the hypothesis that one major route through which object manipulation matters is that it generates many different views of the same object, and that the variation within multiple visual instances of the same object facilitates visual object recognition by building more generalizable representations for recognizing unseen instances. This hypothesis is tested by (1) characterizing the properties of object information generated by infants during free play and by (2) evaluating the information in those generated visual streams by feeding them to convolutional neural networks (CNNs) ? the first computational models of vision capable of human-like visual recognition. Two additional lines of research motivate the approach. First is evidence showing infants learn from statistical regularities in visual inputs presented briefly in a laboratory setting. Second is research using head-mounted cameras suggesting that object views generated by infant manipulation have unique properties, including views dominated by a single object. What we do not yet know are the visual statistics of the views infants generate in everyday toy play or their value for a statistical learner such as CNNs. The proposed research will address these gaps in the literature by characterizing the visual object inputs infants generate and how these inputs may facilitate visual object recognition. The proposed research will also determine how differences in visual inputs may be linked with individual differences in infant object name learning. This research will lead to a deeper understanding of the early development of visual object recognition, and may also provide a crucial missing link in our understanding of the developmental trajectory of other cognitive functions, including object name learning. Moreover, the knowledge to be gained from the proposed research has the potential to inform (1) individual differences in learning, (2) strategies for identifying learning delays, and (3) construction of interventions to remediate learning delays.