VISUAL ATTENTION 213 214 DESIMONE DUNCAN etal 1993).lesions of the ventral prefrontal cortex impair working memory tendency of identical nontargets to group together and apart from the target for objects (Mishkin Manning 1978).Furthermore.Wilson et al (1993) Visual grouping determines which parts of the input belong together;subse- report that cells in the dorsolateral cortex have maintained activity for object quent competitive operations tend to respect,or preserve.these groupings location whereas cells in the ventral cortex have maintained activity for object (Duncan Humphreys 1989). identity.Indeed,the dorsal and ventral prefrontal cortices appear to be the Other than a general tendency for neuronal responses in the visual cortex frontal extensions of the dorsal and ventral processing streams,respectively to be influenced by the ove rall distribution of ite (Mishkin et al 1983,Wilson et al 1993). the neural mechanisms underlying object grouping are ssb Just as the posterior parietal cortex may work together with the dorsolateral grouping as a produc mechanism prefrontal cortex in generating spatial templates,the anterior IT cortex may us role with the inferior prefrontal cortex in generating object this res pl ates (desi monc 4.Fuster et al 1985).Both ar interconnecte y (or gerleid osely related to grouping is the binding problem,or the problem in a 1989). and ns In distributed representation of keeping together parts or attributes of the same d during identical working memory tasks for objects object or entity (Hinton Lang 1985).It is often presumed that there are separate representations for different features,such as color and onentation in the cortex.If so.then an obvious question is how the color red.say.becomes OBJECTS.GROUPING.AND THE BINDING PROBLEM bound to the bar of the appropnate orientation when there are multiple colored hars of different orientation in the visual field.a common view is that attention So far we have not dealt specifically with the representation of objects in the helps solve the binding problem by linking together different features at the cortex.Although this is a key issue for understanding attention,little is actually attended location (Treisman Gelade 1980,Treisman Schmidt 1982).One known about the neural representations of objects.We review just a few of problem with this view is that a complex object such as a face has many the relevant behavioral and neurophysiological facts. differ nt features that would need to be hound too cther, one at a time.Multi As described above.when human subjects divide attention between two uch as the h have h objects,the decre ent in performance isr ather inser sitive tospatial sep e bin to be iden of a hand (e. What does een as part the enti ongt o the s It is far targets may pop ou f a visual display before they are the focus object tha even wh n they are define by the conjun ncan Lappin 196 (Duncan 198 even when t two objects overlap attributes (Duncan Humphreys 1989,McLeod et al 1988,Wolfe et al 1989 4).Indee under simple onditions subjects can identify two This implies some type of solution to the binding problem that works in paralle properties of a single object just as casily as they can identify one(Duncan across the visual field. 1984,1993). At the neural level,the necessity to bind together the output of cells spe The operations that segment and group visual input into discrete objects or cialized for different elemental features may be overstated.To our knowledge, chunks are beyond the scope of this review.Many factors combine to determine no cortical cell has ever been reported that is influenced by only one stimulus which parts of the visual input belong together,including spatial proximity, feature.Neurons may convey more information about some features than shared motion or color.contour features such as local concavities and t-iunc others.but their resp nses often vary along many different feature dimensions tions,and long-term familiarity with the object (see e.g.Beck et al 1983, particularly in area V4 and IT cortex (Desimo ne et al 1984.Tanaka 1993) Hummel Bicderman 1992.Grossberg et al 1994.Palmer 1977).The data Some cells in nd objects with highly suggest.howe ver.that the obiects so constructed behave as wholes when they of f ch a anesthesia pete for and/o ontrol of beh elective mably absen one et al 1984).A isible the P d grouping between irr levan play ite n motio dis. y or ng e,giv mer &make s the Gray 1995).As w with grou much more 0 igno ver Baylis 1989 oson 1991).The known about object representations in the cortex before we understan Ithe role search in homogenous arrays (Figure 3a)partly reflects the of attention in binding. VISUAL ATTENTION 213 et al 1993), lesions of the ventral prefrontal cortex impair working memory for objects (Mishkin & Manning 1978). Furthermore, Wilson et al (1993) report that cells in the dorsolateral cortex have maintained activity for object location whereas cells in the ventral cortex have maintained activity for object identity. Indeed, the dorsal and ventral prefrontal cortices appear to be the frontal extensions of the dorsal and ventral processing streams, respectively (Mishkin et al 1983, Wilson et al 1993). Just as the posterior parietal cortex may work together with the dorsolateral prefrontal cortex in generating spatial templates, the anterior IT cortex may play an analogous role with the inferior prefrontal cortex in generating object and feature templates (Desimonet al 1994, Fuster et al 1985). Both are heavily interconnected anatomically (Ungerleider et al 1989), and neurons in both structures are activated during identical working memory tasks for objects (Chelazzi et al 1993b). OBJECTS, GROUPING, AND THE BINDING PROBLEM So far we have not dealt specifically with the representation of objects in the cortex. Although this is a key issue for understanding attention, little is actually known about the neural representations of objects. We review just a few of the relevant behavioral and neurophysiological facts. As described above, when human subjects divide attention between two objects, the decrement in performance is rather insensitive to spatial separation. What does matter in divided attention is whether two properties to be identified belong to the same or different objects. It is far easier to identify two properties (e.g. orientation and contrast) of one object than properties of two different objects (Duncan 1993, Lappin 1967), even when the two objects overlap (Duncan 1984). Indeed, under simple conditions subjects can identify two properties of a single object just as easily as they can identify one (Duncan 1984, 1993). The operations that segment and group visual input into discrete objects or chunks are beyond the scope of this review. Many factors combine to determine which parts of the visual input belong together, including spatial proximity, shared motion or color, contour features such as local concavities and T-junc￾tions, and long-term familiarity with the object (see e.g. Beck et al 1983, Hummel & Biederman 1992, Grossberg et al 1994, Palmer 1977). The data suggest, however, that the objects so constructed behave as wholes when they compete for visual representation and/or control of behavior. Strengthening the perceived grouping between irrelevant and relevant dis￾play items by, for example, giving them a common motion makes the irrelevant items harder to ignore (Driver & Baylis 1989, Kramer & Jacobson 1991). The ease of visual search in homogenous arrays (Figure 3a) partly reflects the www.annualreviews.org/aronline Annual Reviews Annu. Rev. Neurosci. 1995.18:193-222. Downloaded from arjournals.annualreviews.org by University of California - San Diego on 01/05/07. For personal use only. 214 DESIMONE & DUNCAN tendency of identical nontargets to group together and apart from the target. Visual grouping determines which parts of the input belong together; subse￾quent competitive operations tend to respect, or preserve, these groupings (Duncan & Humphreys 1989). Other than a general tendency for neuronal responses in the visual cortex to be influenced by the overall distribution of items within the receptive field, the neural mechanisms underlying object grouping are unknown. Grossberg et al (1994) have attempted to model grouping as a product of mechanisms for image segmentation. Given the importance of grouping for attentional control, this is a ripe area for future research. Closely related to grouping is the binding problem, or the problem in a distributed representation of keeping together parts or attributes of the same object or entity (Hinton & Lang 1985). It is often presumed that there are separate representations for different features, such as color and orientation, in the cortex. If so, then an obvious question is how the color red, say, becomes bound to the bar of the appropriate orientation when there are multiple colored bars of different orientation in the visual field. A common view is that attention helps solve the binding problem by linking together different features at the attended location (Treisman & Gelade 1980, Treisman & Schmidt 1982). One problem with this view is that a complex object such as a face has many different features that would need to be bound together, one at a time. Multi￾part objects, Such as the human body, may have hierarchical part-whole rela￾tionships that would require comparable binding hierarchies (e.g. a finger may be seen as part of a hand, a limb, or the entire body). Additionall’y~ as we halve - said, targets may pop out of a visual search display before they are the focus of attention, even when they are defined by the conjunction of elementary attributes (Duncan & Humphreys 1989, McLeod et al 1988, Wolfe et al 1989). This implies some type of solution to the binding problem that works in parallel across the visual field. At the neural level, the necessity to bind together the output of cells spe￾cialized for different elemental features may be overstated. To our knowledge, no cortical cell has ever been reported that is influenced by only one stimulus feature. Neurons may convey more information about some features than others, but their responses often vary along many different feature dimensions, particularly in area V4 and IT cortex (Desimone et al 1984, Tanaka 1993). Some cells in temporal cortex respond specifically to objects with highly complex conjunctions of features, such as faces, even under anesthesia when selective attention is presumably absent (Desimonet al 1984). A possible role for correlated activity of neurons in binding is considered elsewhere in this volume (Singer & Gray 1995). As with grouping, much more needs to known about object representations in the cortex before we understand the role of attention in binding. www.annualreviews.org/aronline Annual Reviews Annu. Rev. Neurosci. 1995.18:193-222. Downloaded from arjournals.annualreviews.org by University of California - San Diego on 01/05/07. For personal use only