Grounding computational models of vision with infant brain data

American University, Washington DC

Investigators

Abstract

How do children learn to recognize visual objects? They soon learn to recognize a cat or a cup or the faces of their family very quickly, often with no apparent effort. Human vision is one of the most complex systems in our body. Almost half of our brains are devoted to the visual system. Neuroscientists have a deep interest in understanding visual perception as an important sensory window to the world and as a key domain of human intelligence. Computer scientists are also interested in understanding human vision to gain insights that may help develop better computer vision systems. Artificial neural networks (ANNs) - a form of artificial intelligence - can be trained to recognize visual objects as well. Thus, while understanding human vision may have implications for computer vision, the converse is also possible and hence ANNs have been proposed as a model to understand human vision. However, ANN models do not yet fully match or explain human vision. For example, ANN models are trained in fundamentally different ways from how humans learn from infancy and beyond. One way to train a computer to recognize visual objects, such as cats, is to give it millions of images of cats that have been labeled. However, children can learn to recognize cats after only a few short encounters. Thus, to better model human vision, we need to know how vision develops in human infants. The goal of the research in this project will be to study how the infant brain represents visual objects. Data and tools developed as part of the project will be shared with other scientists to stimulate further research in this field. The project will also involve undergraduate and graduate students as well as outreach to K-12 students, STEM teachers, and local families. This project aims to jointly inform artificial models of human vision, enhance computer vision and advance our understanding of the infant brain. Researchers will use the safe and non-invasive technique of EEG (electro-encephalography) to measure brain activity in infants (12-15 months old). The first aim of the research will be to compare how infants and ANNs represent visual objects. The EEG data will be analyzed using multivariate pattern analysis (MVPA) and representational similarity analysis (RSA) in relation to predictions from different artificial neural network (ANN) models. A second aim will be to explore how infants represent visual objects in the context of partly hidden objects. A related aim is to discover how spoken words, and linguistic representation, can influence object recognition. Infants will be presented with whole or partially occluded images of objects, after hearing a congruent spoken cue or an incongruent cue. Congruent words are predicted to enhance visual processing for partial images, an indication of recurrent or top-down processing. This will test the hypothesis that top-down factors shape how the infant brain represents visual objects. Findings from this research may ultimately contribute to the design of more human-like artificial intelligence in the domain of visual perception. The studies in this project will also build bridges between computational and developmental neuroscience and machine learning and computer vision. For developmental researchers, the proposed methods (MVPA) provide a new and promising tool for the analysis of infant EEG data and results will offer a fresh angle for understanding infants’ visual processing. For researchers in computational neuroscience, results from the project offer an exciting opportunity to align a wide-spread goal, developing neural networks that learn like the human brain learns, to actual data from the learning infant brain. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.