Irving Biederman Homepage

In a fraction of a second -- from a single visual fixation -- humans are able to comprehend novel images of objects and scenes, often under highly degraded and novel viewing conditions. To account for this extraordinary capacity, we have proposed that objects are represented as an arrangement of simple, convex, viewpoint-invariant shape primitives, termed "geons," such as bricks, cylinders, wedges, and cones, that serve to distinguish visual entities at a basic (or entry) level, so that a given image can be determined to be that of a chair, fork, or penguin (Biederman, 1987). The geons can be distinguished by properties of edges that are invariant with orientation in depth (such as straight vs. curved contours) so representations distinguished by geons possess the same invariance. As long as two or three geons in their specified relations can be extracted from the image, entry-level classification will almost always be successful despite drastic variations in the object's silhouette, specific contours, and occlusion of large regions of the object. In a series of priming experiments, we have demonstrated that all the priming can be attributed to the activation of the geons (in their specified relations), rather than to the local image features (lines and vertices; spatial frequency components) that initially activated the geons or an overall model of the object. Other experiments showed that, indeed, the invariant properties of edges are far more important in object classification than metric properties, properties that do not remain invariant under rotation in depth, such as degree of curvature or aspect ratio.

A neural net implementation of this theory (Hummel & Biederman, 1992) provided an account, for the first time, of how the priming -- presumed to occur in the ventral pathway for shape recognition -- could be invariant with the position, size, and orientation in depth of the object (up to parts occlusion). A major contribution of the model is a scheme for the perceptual grouping of units activated by image contours from the same geon: Units that are to be grouped together fire together. The dorsal pathway is assumed to contribute to old-new recognition memory judgments, which provides an account of why such judgments show none of the invariances characteristic of object classification.

More recently, we have confirmed the neural basis of some of the major assumption of geon theory through single unit recording experiments in the inferior temporal region of the macaque (e.g., Kayaert, Biederman, & Vogels, 2005) as well as fMRI studies done at our new Dornsife Imaging Center adjacent to the Neuroscience Building. A population code of these neurons more readily distinguishes viewpoint invariant properties than metric properties and the tuning to viewpoint invariant properties is independent of the depth orientation of the object. We have also developed a theory of the distinction between face and object representation (Biederman & Kalocsai, 1987). This project explores the extent to which the similarity of faces, when they cannot be distinguished by easy features, can be modeled in terms of the pattern of activation over a lattice of spatial filters (Biederman & Kalocsai, 1997). Whereas an arrangement of geons appears to suffice for object recognition, a similarity space that preserves the metrics implicit in the original spatial filter activations may suffice for face recognition.

A new series of experiments in testing a theory that we have developed of perceptual and cognitive pleasure (Biederman & Vessel, 2006). People generally seek out, with every visual fixation, and decision as to what to read or whaat movie to see, new but interpretable information. What is the neural basis for such behavior? That is, what makes us infovores? Stay tuned.