VISUAL ATTENTION 215 216 DESIMONE DUNCAN SERIAL AND PARALLEL MODELS for another,or the attentional dwell time,is a critical issue in comparing the different models.To make such a measurement,a brief temporal interval is When targets are selected by spatial location,all models of the underlying introduced between two targets,e.g.two letters to be identified.If the time mechanism posit some type of spatial gating mechanism.It is when the target's between pre of the ter than the nal dwvell location i and it must be d on the basis of its identits (e.B shonld be interfe both ta r proce ng c g of modelsd et al 1993,Schneide ning of a scer ne,each object co sumes processing capacity for only by a few dozen milliseconds:thus,the attentional dwell time is short and inter Shiftnn 1977,Treisman Gelade 1980 target pops out from the ference should be eliminated with correspondingly short interstimulus inter- background on the basis of an elemental feature difference.As each elemen vals.Attentional dwell times can be much greater in parallel models because is selected in turn by attention.it is evaluated by a recognition memory process more than one object in a scene is processed at once(with increasing interfer- and the scan of the array is terminated when the target is found.As morc and ence as the number of objects is increased).Thus,interference may last for more nontarget elements are added to the scene,it takes longer and longer to far longer periods of time.A recent study using this method of sequential target scan the array to find the target;50 ms per item is typical (although as we presentation found interference lasting for several hundred milliseconds,con. have said,this time varies continuously over a large range of values). sistent with parallel models (Duncan et al 1994;see also Pashler Badgio In the other major class of models,all elements of the visual input compete 1987. in parallel for visual processing (Atkinson et al 1969,Bundesen 1990,Du &Humphreys 1967).Thisclass includes the biased competition nt that has been the the me s far CONCLUSIONS s review dis een pa serial model is ofr on the reaction d tic By way of contrast,it would be useful to consider again the standard model of sele tive use recent ser e come hybrid i oth seri paralle this vie tion widely accepte attention fo nt process To explain pop-out effects with targets defined by th 1 by a system o ncneregionofthew I enna njunction of several features.for example.both Guided Search (Wolfe et a spatially mappe e proc ions and reduce it at un d ones.The 989)and Feature Integration Theory (Treisman Sato 1990)incorporatc parallel top-down processes to identify all regions in the visual field that share components of this system are revealed by neglect and extinction syndromes target features.Another interesting hybrid is the spatial and object search model following lesions.Attention is unnecessary for simple feature discri of Grossberg et al (1994).which explains both easy and difficult search on the but resolves the binding problem by linking together the output of cells coding different elemental features of the attended object.It is a serial,high-speed basis of grouping and recognition operations recursively applied in parallel across the visual field. scanning mechanism moving from one location to the next in around 50 ms. The data we have reviewed cast doubt over many of the postulates of the Unfortunately.the physiological data on obiect search in IT cortex (Chelazz standard view.Instead,they suggest the following conclusions: et al 1993)described ab e do not allow us to distinguish conclusively betweer serial and parallel mechanisms.The fact that se acti ate 1.At several points between input and resp se ohieets in the visual field of the and cont sistent with the b -up neura isms tha Ho it is seems to figures part by n their space and ame)and on,with senal scanne n mech sms tha Thesrongesiargumcn itching between e s at a rate to d to discem in the neural dat ch bias can control ed by many stimulus atribu nst the serial model is that known memory mech selection by spatial location,by simple object features,and by complex sms in I cortex are sufficient to explain the results without invoking a conjunctions of features. cess 3.Within the ventral stream,which underlies object recognition.top-down The time it takes to recognize one object and release processing capacity biasing inputs resolve competition mainly between objects located withinVISUAL ATIENTION 215 SERIAL AND PARALLEL MODELS When targets are selected by spatial location, all models of the underlying mechanism posit some type of spatial gating mechanism. It is when the target’s location is unknown and it must be found on the basis of its identity (e.g. searching for a face in a crowd) that different classes of models diverge significantly. According to serial search accounts, scenes are searched element by element by a spotlight of attention (Olhausen et al 1993, Schneider Shiffrin 1977, Treisman & Gelade 1980), unless the target pops out from the background on the basis of an elemental feature difference. As each element is selected in turn by attention, it is evaluated by a recognition memory process, and the scan of the array is terminated when the target is found. As more and more nontarget elements are added to the scene, it takes longer and longer to scan the array to find the target; 50 ms per item is typical (although as we have said, this time varies continuously over a large range of values). In the other major class of models, all elements of the visual input compete in parallel for visual processing (Atkinson et al 1969, Bundesen 1990, Duncan & Humphreys 1989, Sperling 1967). This class includes the biased competition account that has been the theme of this review so far. The difficulties of distinguishing between parallel and serial models on the basis of reaction time data are well known (Townsend 1971), particularly because recent serial models have become hybrids with both serial and parallel component processes. To explain pop-out effects with targets defined by the conjunction of several features, for example, both Guided Search (Wolfe et al 1989) and Feature Integration Theory (Treisman & Sato 1990) incorporate parallel top-down processes to identify all regions in the visual field that share target features. Another interesting hybrid is the spatial and object search model of Grossberg et al (1994), which explains both easy and difficult search on the basis of grouping and recognition operations recursively applied in parallel across the visual field. Unfortunately, the physiological data on object search in IT cortex (Chelazzi et al 1993) described above do not allow us to distinguish conclusively between serial and parallel mechanisms. The fact that search arrays initially activate cells selective for any of the component elements, targets or nontargets, is consistent with the biased competition model in which all objects are prcessed in parallel. However, it is possible that what seems to be an initial parallel activation lasting 200 ms is actually a serial activation, with the serial scanner switching between elements at a rate too rapid to discern in the neural data. The strongest argument against the serial model is that known memory mech￾anisms in IT cortex are sufficient to explain the results without invoking a hidden serial process. The time it takes to recognize one object and release processing capacity 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. 216 DESIMONE & DUNCAN for another, or the attentional dwell time, is a critical issue in comparing the different models. To make such a measurement, a brief temporal interval is introduced between two targets, e.g. two letters to be identified. If the time between presentation of the targets is shorter than the attentional dwell time, there should be interference, as both targets will compete for processing ca￾pacity. According to typical serial models, which posit rapid attentional scan￾ning of objects in a scene, each object consumes processing capacity for only a few dozen milliseconds; thus, the attentional dwell time is short and inter￾ference should be eliminated with correspondingly short interstimulus inter￾vals. Attentional dwell times can be much greater in parallel models because more than one object in a scene is processed at once (with increasing interfer￾ence as the number of objects is increased). Thus, interference may last for far longer periods of time. A recent study using this method of sequential target presentation found interference lasting for several hundred milliseconds, con￾sistent with parallel models (Duncan et al 1994; see also Pashler & Badgio 1987). CONCLUSIONS By way of contrast, it would be useful to consider again the standard model of selective visual attention widely accepted in neuroscience. According to this view, attention focuses on one region of the visual field at a time. It is mediated by a system of spatially mapped structures that enhance processing in visual cortex at attended locations and reduce it at unattended ones. The components of this system are revealed by neglect and extinction syndromes following lesions. Attention is unnecessary for simple feature discriminations but resolves the binding problem by linking together the output of cells coding different elemental features of the attended object. It is a serial, high-speed scanning mechanism moving from one location to the next in around 50 ms. The data we have reviewed cast doubt over many of the postulates of the standard view. Instead, they suggest the following conclusions: 1. At several points between input and response, objects in the visual field compete for limited processing capacity and control of behavior. 2. This competition is biased in part by bottom-up neural mechanisms that separate figures from their background (in both space and time) and in part by top-down mechanisms that select objects of relevance to current behav￾ior. Such bias can be controlled by many stimulus attributes, including selection by spatial location, by simple object features, and by complex conjunctions of features. 3. Within the ventral stream, which underlies object recognition, top-down biasing inputs resolve competition mainly between objects located within 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