Object Recognition Insights From Advances in fMRI Methods
چکیده
Recent advancements in imaging methods and analysis approaches have provided important insights about the neural bases of object recognition. We address the potential limitations of standard functional magnetic resonance imaging (fMRI) and discuss methodological advancements, including fMRI-adaptation, pattern analyses, and high-resolution fMRI, that may be more appropriate for studying object and face representations. fMRIadaptation and high-resolution fMRI measure responses of neural subpopulations within standard fMRI voxels, and pattern analyses examine the information in the distributed activations across voxels, which may differ from the mean response across these voxels. These methods have provided evidence for a multitude of representations across the human ventral stream that provide empirical constraints for cognitive theories of recognition. KEYWORDS—object recognition; face recognition; fMRIadaptation; pattern analyses; high-resolution fMRI. Object recognition is an amazing human feat. Humans effortlessly recognize objects within a fraction of a second. Variability in viewing conditions can cause changes in objects’ location, size, viewpoint, illumination, occlusion, etc., yet recognition is largely invariant to these dramatic changes in objects’ appearance. The underlying neural representations and computations that enable this to happen, however, are still mysterious. Neuroimaging techniques, especially functional magnetic resonance imaging (fMRI) offer an opportunity to investigate neural mechanisms underlying object recognition. Object recognition is thought to involve the visual ‘‘ventral stream,’’ a processing pathway extending from the occipital lobe to the temporal lobe. Indeed, fMRI studies have revealed a constellation of object-selective regions in the lateral and ventral occipito-temporal cortex (known as the lateral occipital complex, or LOC; Malach et al., 1995; Fig. 1a) that respond more strongly to objects than to noise patterns, textures, or scrambled objects (created by breaking object images into small squares and shuffling the locations of these squares) and several faceselective regions that respond more strongly to faces than to nonface objects (Fig. 1a). The most studied of these face-selective regions is the fusiform face area (FFA; Kanwisher, McDermott, & Chun, 1997). Activation in both the LOC and FFA correlates with subjects’ recognition performance (Fig. 1b). When stimuli are ambiguous or presented close to the perceptual threshold (i.e., the border between seeing and not seeing), activity in faceand object-selective cortical areas is higher when subjects recognize the stimuli than when they do not (GrillSpector, Knouf & Kanwisher, 2004; Grill-Spector, Kushnir, Hendler, & Malach, 2000; Tong, Nakayama, Vaughan, & Kanwisher, 1998). Further, prolonged experience with objects (Gauthier, Skudlarski, Gore, & Anderson, 2000; Gauthier, Tarr, Anderson, Skudlarski, & Gore, 1999) and faces (Golarai et al., 2007) modifies these representations. These studies provide fundamental knowledge about which brain regions are involved in object and face recognition. However, many questions about the nature of the underlying neural representations remain: How do these representations provide for invariant recognition without losing the acuity to distinguish between similar items? What mechanisms underlie rapid categorization (e.g., discrimination between a car and a face)? Are the same representations used for between-category discrimination (e.g., distinguishing between a car and a boat) and within-category discrimination (e.g., distinguishing between a sailboat and a motorboat)? Are representations of faces (or objects of expertise) fundamentally different from other object representations, and if so, how do they differ? How do experience and development shape these representations? In addressing these questions, the inherent limitations of standard-resolution fMRI (SR-fMRI) pose significant challenges. SR-fMRI measures the pooled neural responses across a voxel Address correspondence to Kalanit Grill-Spector, Department of Psychology, Jordan Hall, Stanford University, Stanford, CA 94305; e-mail: kalanit@psych.stanford.edu. CURRENT DIRECTIONS IN PSYCHOLOGICAL SCIENCE Volume 17—Number 2 73 Copyright r 2008 Association for Psychological Science (a three-dimensional volume element analogous to a pixel in a two-dimensional digital image) or many voxels that constitute a brain region. Since a cubic millimeter (mm) of cortex contains about 10,000 to 50,000 neurons and fMRI signals are typically measured across regions that are in the range of 50 to 5,000 cubic mm in size, those regions reflect the activity of about a Fig. 1. Activation in the human ventral stream. Colored patches on the right lateral (a, top) and ventral (a, bottom) views of the cortical surface of one representative subject show object-, face-, and place-selective regions. Red5 regions that respond more to faces than to nonface objects, green5 regions that respond more to places than to nonface objects, blue5 regions that respond to nonface objects more than to scrambled objects, pink 5 regions that respond both more to faces than to objects and more to objects than to scrambled objects, dark green 5 regions that respond more to places than to objects and more to objects than to scrambled objects. Boundaries of retinotopic visual areas are shown in black on the right hemisphere. The lateral occipital complex (LOC) includes object-selective regions (blue and pink regions) in the lateral and ventral occipital cortex. Face-selective regions include the fusiform face area (FFA, red and pink regions in the fusiform gyrus); a face-selective region (shown in red) in the posterior superior temporal sulcus (pSTS) and a face-selective region that overlaps the LOC (shown in pink). MT is a motion-selective region in the posterior end of the inferior temporal sulcus. Bar graphs on the left show the mean percentage signal from the LOC (excluding regions that respond more to faces than to objects; bottom bar graph) and FFA (top bar graph) from four kinds of stimuli, averaged across 10 subjects, relative to a scrambled-object baseline. Stimuli (indicated by icons) were faces, fourlegged mammals, cars, and abstract sculptures. Object-selective regions respond to both objects and faces more than to the scrambled baseline, but they do not respond more strongly to faces than to objects. fMRI signal and recognition performance as a function of image exposure is shown in the bottom graph (b). During scanning, subjects viewed photographs of faces, animals, and common objects (e.g., cars, boats), which were followed by scrambled images. Subjects were asked to name pictures covertly at the basic level (e.g., face, dog, car, boat). In different blocks, images appeared for a different duration (20, 40, 120, or 500 milliseconds, ms) followed by a scrambled image completing a 500-ms trial (480, 460, 380, or 0-ms mask duration). Immediately after fMRI scans (but while in the scanner), subjects saw the same images at the same duration and participated in a behavioral experiment in which recognition performance was recorded. Recognition performance (dashed black line) is correlated with LOC activation (red) but not with V1 (primary visual cortex) activation (blue). Adapted from ‘‘The Dynamics of Object-Selective Activation Correlate With Recognition Performance in Humans,’’ by K. Grill-Spector, T. Kushnir, T. Hendler, & R. Malach, 2000, Nature Neuroscience, 3, pp. 837–843, Figure 3. Copyright 2000, Nature Publishing Group. Adapted with permission. Fig. 2. fMRI methods used to study the properties of cortical regions involved in object recognition. The top (a) left graph shows the mean lateral occipital complex (LOC) response to animals presented in different degrees of rotation (vs. a scrambled baseline), averaged across 7 subjects, using standard fMRI. The top (a) right graph shows LOC responses, averaged across the same 7 subjects, during an fMRI-adaptation experiment in which we measured sensitivity of LOC responses to object rotation. In each fMRI-adaptation trial, two images of the same animal were shown either in the same view (0 rotation) or in different views that were rotated 60, 120, or 180 degrees away from the initial view. The black diamond indicates responses to trials containing two different animals presented in the same view (unadapted response); open circles indicate significant adaptation ( p< 0.05) when two images of the same animal were shown compared to when two different animals were shown. Note that there was adaptation when the same animal and view were repeated and also when the same animal was shown at a view that was rotated 60 degrees away from the initial view, but that there was no significant adaptation when the same animal was shown rotated 120 or 180 away from the initial view (Andresen & Grill-Spector, 2006). The middle graphs (b) show pattern analyses. The mean LOC response from a single subject to different object categories (shown in the icons), relative to a scrambled-object baseline, is shown in the middle (b) left graph. The distributed response across the LOC to these categories, from the same subject, is shown in the middle (b) right graph. The x-axis represents voxels (in this subject, the LOC contained about 200 voxels); the y-axis represents percent signal change relative to the scrambled-object baseline. Each line represents the distributed response to one category: red5 faces; yellow5 animals; light blue5 cars; and dark blue 5 abstract objects. Each category elicited a different pattern of response across the LOC. The bottom (c) images show face-selective activations in one subject on the cortical surface around the fusiform gyrus. The left panel shows activations measured with standard-resolution fMRI (voxels of 3 3 3 mm); the right panel shows the same activations as measured with high-resolution fMRI (voxels of 1 1 1 mm). The latter shows more localized faceselective activations. Activation maps were thresholded at a significance of p < 0.001. 74 Volume 17—Number 2 Object Recognition