Isabel Gauthier - US grants

Visual object recognition occurs at different levels of abstraction ranging from categorical levels, e.g., "dog," to the more specific individual level, e.g., "my English hound." Moreover, we can develop "expertise" at one of these levels for a given category; for instance, bird watchers are experts at the species level. This research will continue to investigate the roles of level of categorization and perceptual expertise in the development of cognitive and neural mechanisms selective for object categories (such as faces or birds). Because different methods offer different strengths and weaknesses, this research will involve converging evidence, including behavioral psychophysics, functional brain imaging (fMRI), and event related potentials (ERPs) in normal humans, as well as extending these techniques to brain-injured individuals. The research program is divided into four sections addressing different questions:

1) How do people become perceptual experts? A first set of experiments will manipulate whether subjects rely on their own observations or require feedback and supervision. A second set of studies will examine whether non-visual knowledge about objects contributes to the learning process and affects the organization of category-specific areas. Other experiments will test the plasticity of the brain regions, which support object recognition, investigating whether damage to one area can be compensated for by reorganization of other areas.

2) What are the computational roles of different brain areas within the network that mediates expertise with visually-similar objects? Experiments using a combination of fMRI, ERP, and behavioral measures will investigate how different category-selective brain areas support identification at the categorical, subordinate, and individual levels.

3) What is the capacity of perceptual expertise? Experiments will test whether one can become an expert with many different classes of objects (e.g., birds, dogs, cars, faces, flowers, etc.), as well as whether there is interference when objects from different expertise domains are processed at the same time.

4) Can perceptual expertise be acquired more easily with some object geometries? In particular, adaptive pressures for accurate face recognition may have "biased" the system to prefer face-like configurations. By manipulating the visual structure of stimulus objects, behavioral and fMRI experiments will investigate the geometric constraints on the acquisition of expertise.

Overall, these experiments should help us to better understand the nature of visual object recognition, elucidating how a single system can support the wide range of recognition tasks we are able to perform. The implications of these findings vary from possible protocols for the rehabilitation of brain-injured individuals to the better education of learning-impaired children (e.g., as in autism) to the development of more effective and robust machine vision systems for face and object recognition.

PROJECT SUMMARY The long-term goal of this line of research is to understand individual differences in our ability to recognize faces and objects. Prior research has suggested a genetic basis for individual differences in face recognition that is independent from our ability to recognize objects. However, evidence suggests that the ability to recognize objects is well measured by existing methods, and that the relationship between face and object recognition is mediated by experience with objects. In particular, this research seeks to investigate a possible link between the ability to recognize faces and the ability to individuate visually similar non-face objects, given a controlled amount of experience. This relationship is postulated on the basis of prior findings revealing that when people acquire experience individuating objects, they process them as wholes, or holistically, rather than as a collection of parts, and such holistic processing is known to be hallmark of face processing. We will seek evidence for an underlying domain-general ability that supports individuation of visually similar objects for which one has experience, by testing a model in which this ability is expressed across different measures of individuation and across different object categories (Aim 1). During training, participants will learn to individuate visually similar objects within each of four visually distinct categories. Using confirmatory factor analysis (CFA) we will specify and test two alternative models specifying that a superordinate and/or broad factor of individuation ability can be identified that accounts for substantial variance in performance across tasks and categories. Using the best-fitting model from step 1, we will then test whether individuating objects and recognizing faces are related abilities (Aim 2) by conducting a second CFA that links measures of face and object recognition. Understanding the nature of the abilities that are linked to object and face recognition will facilitate prediction in a number of domains where skilled perception is important, including medical diagnosis. It will further our understanding of the underlying mechanisms and neural substrates underlying these abilities and should also facilitate behavioral genetic studies of high-level vision, contributing to our understanding of a hereditary form of a visual face recognition deficit, developmental prosopagnosia, with a prevalence in the normal population estimated by one study to be as high as 2.5%.

This Science of Learning Collaborative Network brings together researchers from Vanderbilt University, Carnegie-Mellon University, and University of California-San Diego to investigate how and why people differ in their ability to recognize, remember, and categorize faces and objects. Many important real-world problems, such as forensics, medical imaging, and homeland security demand precise visual understanding from human experts. Understanding individual differences in high-level visual cognition has received little attention compared to other aspects of human performance. Recent studies indicate that there likely is far greater variability than commonly acknowledged in the ability to learn high-level visual skills and that such ability is poorly predicted by general intelligence. This project supports a collaborative interdisciplinary research network that aims to develop measures of individual differences in visual recognition, relate behavioral and neural markers of individual differences, develop models that explain individual differences, and relate models with neural data. Because outcomes in many real-world domains depend on decisions based on visual information, developing measures, markers, and models of individual differences can have broader impacts on identifying real-world visual talent and improving visual performance and training. Students and fellows conducting research as part of this collaborative network, including female scientists and underrepresented minorities, will be mentored by scientists from multiple disciplines, providing them with an understanding far deeper than that achievable by a single discipline.

The project will support the activities of a collaborative research network on the study of individual differences in visual recognition. The scientists involved in these interdisciplinary efforts include experts in brain imaging at ultra-high field strength, cutting-edge methods in the development of psychological tests, and cognitive and "deep" convolutional neural network models of high-level vision. The project will investigate how functional brain activity and anatomical brain structure can predict the quality and time-course of visual performance and visual learning. The team will develop and validate tests of visual ability that can be used to make precise predictions about brain activity and behavioral performance. These brain measures and behavioral tests will be related to deep convolutional neural network models; such models are the most successful computer vision models to date, and higher layers of these hierarchical networks provide outstanding models of brain areas critical to object recognition. So far these models have not been used to understand individual differences. Instead of the typical approach seeking to achieve the best performance possible, the collaborative team will seek models that can mirror human variability, making errors when people make errors, being slow when people are slow, and displaying a range of visual abilities and learning as observed in humans.

The award is from the Science of Learning-Collaborative Networks (SL-CN) Program, with funding from the SBE Division of Behavioral and Cognitive Sciences (BCS), the SBE Office of Multidisciplinary Activities (SMA), and the CISE Division of Computer and Network Systems (CNS).