Modeling the Representation of Object Boundary Contours in Human fMRI Data
The human visual system consists of a hierarchy of areas, each of which represents different features of the visual world. Recent studies have revealed that most brain areas—and even many individual neurons—represent information about multiple visual features. Thus, a complete model of the brain must specify the relative importance of multiple visual features across the visual hierarchy. This talk will describe our work to estimate the importance of object boundary contours relative to other features.
Boundary contours define the edges of figural objects in scenes, and figure/ground segmentation has long been held to be a critical process in human vision. However, the relative importance of boundary contours compared to both lower- and higher-level features (e.g. motion energy and visual categories) remains unknown. To address this issue, we measured fMRI responses while human subjects viewed two sets of movies that varied in many feature dimensions: rendered movies of artificial scenes and cinematic movies. We modeled responses to both sets of movies independently using the same three models: models of motion energy, object boundary contours, and visual categories. We used the encoding models to predict withheld fMRI data, and used variance partitioning to determine whether the various models explained unique or shared variance in each dataset. We found that the pattern of unique variance explained by the three models was qualitatively consistent across both datasets, with unique variance explained by boundary contours in Lateral Occipital cortex and other areas. However, the three models also shared substantially more variance in the cinematic movies, likely due to correlations between model features. For example, much of the motion energy in the cinematic movies was a result of people moving. The shared variance between all three models in the cinematic movies in particular highlights the need for complex stimulus sets in which features in different models are de-correlated from each other.
Access is free to subscribers of the IEEE Brain Community.
Dr. Mark Lescroart
Mark Lescroart joined the Cognitive & Brain Sciences group in the department of psychology at UN Reno in the spring of 2018. Dr. Lescroart attended the University of Southern California and graduated with a B.S. in Psychobiology and a minor in Japanese. He taught English for a year in Japan, and attended USC for graduate school as well. He got his PhD in 2011, working with Irving Biederman to study shape recognition. He did postdoctoral work in computational neuroscience with Jack Gallant at UC Berkeley, and received a Ruth L. Kirschstein National Research Service Award Fellowship to support this work in 2012. Dr. Lescroart’s research seeks to understand how our brains transform patterns of light on our retinas into useful information about the world. His lab is specifically interested how the brain provides information about the identity and shape of objects, the structure of the space around us, and the types of actions (such as running, jumping, and throwing) performed by humans in our environment. The lab pursues these questions by making computational models of brain responses measured by fMRI.
Publication Year: 2019