复制成功
  • 图案背景
  • 纯色背景
  •   |  注册
  • /
  • 批注本地保存成功,开通会员云端永久保存 去开通
  • task switching a pdp model

    下载积分:2000

    内容提示: Cognitive Psychologydoi:10.1006/cogp.2001.0770, available online at http://www.idealibrary.com onTask Switching: A PDP ModelSam J. Gilbert and Tim ShalliceInstitute of Cognitive Neuroscience, University College London, London, United KingdomWhen subjects switch between a pair of stimulus–response tasks, reaction timeis slower on trial N if a different task was performed on trial N  1. We present aparallel distributed processing (PDP) model that simulates this effect when subjectsswitch between word read...

    亚博足球app下载格式:PDF| 浏览次数:2| 上传日期:2014-08-28 01:19:50| 亚博足球app下载星级:
    Cognitive Psychologydoi:10.1006/cogp.2001.0770, available online at http://www.idealibrary.com onTask Switching: A PDP ModelSam J. Gilbert and Tim ShalliceInstitute of Cognitive Neuroscience, University College London, London, United KingdomWhen subjects switch between a pair of stimulus–response tasks, reaction timeis slower on trial N if a different task was performed on trial N  1. We present aparallel distributed processing (PDP) model that simulates this effect when subjectsswitch between word reading and color naming in response to Stroop stimuli. Reac-tion time on ‘‘switch trials’’ can be slowed by an extended response selection pro-cess which results from (a) persisting, inappropriate states of activation and inhibi-tion of task-controlling representations; and (b) associative learning, which allowsstimuli to evoke tasks sets with which they have recently been associated (as pro-posed by Allport & Wylie, 2000). The model provides a good fit to a large bodyof empirical data, including findings which have been seen as problematic for thisexplanation of switch costs, and shows similar behavior when the parameters areset to random values, supporting Allport and Wylie’s proposal.© 2001 Elsevier ScienceKey Words: task switching; task set; Stroop effect; parallel distributed processing;executive functions.Atkinson and Shiffrin (1968) proposed a distinction between relativelypermanent cognitive structures, such as short- and long-term memory, andcontrol processes which harness those fixed structures in order to attain spe-cific goals. This distinction was elaborated in the following years (e.g.,Posner & Snyder, 1975; Shiffrin & Schneider, 1977) and has been generallyaccepted (Shallice, 1994). Yet research with normal subjects in the decadesfollowing Atkinson and Shiffrin’s article paid relatively little attention to thecontrol processes required to select and organize fixed cognitive structures(with some notable exceptions, e.g., Gopher, Weil, & Siegel, 1989; Logan,1979, 1980, 1985; Spelke, Hirst, & Neisser, 1976).In recent years, studies have begun to investigate the control processesof normal subjects when they switch between different cognitive tasks onThe research reported in this article was supported by a research studentship awarded toSam Gilbert by the UK Medical Research Council. We thank Gordon Logan, NachshonMeiran, Glenn Wylie, Nick Yeung, and an anonymous reviewer for their helpful commentson an earlier version of this article.Address correspondence and reprint requests to Sam Gilbert, Institute of Cognitive Neuro-science, Alexandra House, 17 Queen Square, London WC1N 3AR, United Kingdom. E-mail:sam.gilbert@ucl.ac.uk. Fax: 44 (0)20 7916 8517.0010-0285/01 $35.00© 2001 by Elsevier ScienceAll rights reserved. 2GILBERT AND SHALLICEsuccessive trials (e.g., Allport, Styles, & Hsieh, 1994; Meiran, 1996; Rog-ers & Monsell, 1995; and contributions to Monsell & Driver, 2000). Thetask-switching paradigm therefore appears to offer a valuable tool for study-ing ‘‘executive control’’ (Logan, 1985; Monsell, 1996), both in normal sub-jects and in patients (e.g., Rogers, Sahakian, Hodges, Polkey, Kennard, &Robbins, 1998). However, there has been disagreement over the interpreta-tion ofexperiments involving task switching, making it difficult to draw firmconclusions from them about executive control. Here, we seek to clarify thisdebate by presenting a computational model of task switching.MAIN EMPIRICAL FINDINGS AND THEORETICALINTERPRETATIONSA number of methodologies have been developed for studies of taskswitching. One approach is simply to compare pure and mixed blocks, i.e.,to compare blocks where the subject performs the same task on every trialwith blocks in which s/he must alternate between two tasks on successivetrials (e.g., Allport et al., 1994; Jersild, 1927; Spector & Biederman, 1976).Other studies have employed the ‘‘alternating runs’’ paradigm (Rogers &Monsell, 1995), where subjects are required to switch tasks predictably everynth trial, where n is at least 2. This has the advantage that ‘‘switch’’ trials(where the task differs from the one performed on the previous trial) can becompared with ‘‘nonswitch’’ (or ‘‘repeat’’) trials within the same block. Ina third methodology for task switching experiments, each trial is precededby a cue which instructs the subjects which task to perform (e.g., Meiran,1996; Sudevan & Taylor, 1987). This paradigm makes the requirement toswitch tasks unpredictable.Experiments using each ofthese paradigms have demonstrated costs, bothin reaction time and error rate, for switch compared with nonswitch trials(or mixed compared with pure blocks). These ‘‘switch costs’’ have beenreported to vary from zero to several hundred milliseconds per item, de-pending on the experimental conditions (see below). The main test of themodel presented in this article will be whether it can predict the effects ofvarious experimental manipulations on switch costs.Explanations ofthe Switch CostAn important concept in studies oftask switching is the ‘‘task set,’’ whichmay be defined loosely as the set of cognitive operations required to effec-tively perform a task (‘‘To form an effective intention to perform a particulartask, regardless of which of the range of task-relevant stimuli will occur, isto adopt a task-set’’; Rogers & Monsell, 1995, p. 208, emphasis in original).This concept is often used in the same way as the concept of the actionor thought ‘‘schema’’ (Norman & Shallice, 1986). The distinction betweenindividual S-R mappings and task sets, composed ofall ofthe individual S-R A PDP MODEL OF TASK SWITCHING3mappings which are required to carry out an experimental task, is crucial. Aswe shall see, an important issue concerns the degree to which task switchingshould be understood in terms of processes occurring at the level of discreteS-R mappings rather than those which occur at the level of the task set (seealso Monsell, Taylor, & Murphy, 2001).Two main theoretical accounts of the switch cost have been put forward.Allport and colleagues (e.g., Allport et al., 1994; Allport & Wylie, 2000)have suggested that switch costs index an interference effect caused by acarryover of the previous task set into switch trials (this will be called the‘‘task carryover account’’). As well as a carryover of the previously activetask set into switch trials, inhibition of competing task sets can also persiston switch trials, when the previously competing task is now required. Ac-cording to this theory, there is no need to postulate differences between thehigherlevel cognitive processes that occurduring switch and nonswitch trialsin order to explain switch costs; switch trials are simply prolonged by greatercompetition caused by the carryover effect (i.e., switch costs reflect a formofpriming). This account does not deny the involvement ofcontrol processesin task switching. Without such processes, the subject could never switchtasks at all. What is denied by this account is that these control processesare measured in any direct way by the switch cost.An opposing explanation of switch costs has been advanced by Monselland colleagues (Monsell, Yeung, & Azuma, 2000; Rogers & Monsell, 1995).This hypothesis proposes that switch costs do reflect the duration ofa stage-like executive control process which reconfigures the cognitive system forthe upcoming task. However, it is held that this process cannot be completeduntil the arrival of the first stimulus of the new task (i.e., it is a stimulus-driven ‘‘exogenous control process’’). An additional component of theswitch cost is hypothesized to reflect the operation of an ‘‘endogenous con-trol process’’ which can be executed before the arrival of the first stimulusin the new task. In a similar account, Rubinstein et al. (2001) attribute switchcosts to the duration ofa‘‘goal shifting’’ stage, which can be executed beforethe arrival of a stimulus, and a ‘‘rule activation’’ stage which must awaitstimulus presentation.Evidence for the Task Carryover AccountEvidence for the task carryover account of switch costs has come fromfindings which demonstrate an effect of the performance of an earlier taskon subsequent switch costs (Allport et al., 1994; Allport & Wylie, 2000; seealso Mayr & Keele, 2000). Two ofthese findings are ofparticular theoreticalrelevance.Asymmetric switch costs. Allport et al. (1994, Experiment 5) carried outan experiment where subjects switched between word reading and colornaming in response to incongruent Stroop stimuli (e.g., ‘‘red’’ written inblue ink; Stroop, 1935). As expected, word reading yielded faster reaction 4GILBERT AND SHALLICEtimes than color naming (see MacLeod, 1991). However, this experimentalso produced a most unexpected result. Reaction times for word-readingtrials were slower when color naming had been performed on the previoustrial (i.e., there was a switch cost), but Allport et al. failed to detect a switchcost for color-naming trials that followed performance of the word-readingtask. In other words, the switch cost appeared to be confined to the switchfrom the nondominant into the dominant (i.e., better learned, easier) task.More recent experiments have replicated this asymmetry in switch costs(Allport & Wylie, 2000; Wylie & Allport, 2000). Allport and Wylie foundthat there is a cost for a switch into the color-naming task, but it is smallerthan the cost of a switch into the word-reading task. Meuter and Allport(1999) reported an analogous finding when subjects switched between digitnaming in their dominant and nondominant languages. This paradoxicalfinding—larger reaction time costs for a switch into a better learned, moredominant task—is difficult to explain if switch costs reflect the time takento reconfigure the cognitive system for the upcoming task. Why should aswitch into an easier, better learned tasktake longer to complete than a switchinto a less familiar task? Allport et al. argue that the result can be explainedif the primary determinant of the switch cost is the nature of the previoustask. When subjects name the color of an incongruent Stroop stimulus, All-port et al. hypothesize that inhibition of the word-reading task may be re-quired. According to the task carryover account, this inhibition will persiston a switch trial where word-reading is now appropriate, leading to a largeswitch cost. But in the absence of any requirement to suppress color namingin order to perform the word-reading task, there will be no carryover ofinhibition into color naming switch trials, hence the small or absent switchcosts.Item specific costs. Allport and Wylie (2000, Experiment 5) investigatedthe extent to which the switch cost is caused by the repetition of stimulibetween the two tasks. Subjects alternated between short runs of color nam-ing and word reading, but only a subset of the stimuli that appeared in theword-reading task were also presented for color naming. Thus, it was possi-ble to compare two types of stimuli in the word-reading task: ‘‘primed’’stimuli, which had also appeared recently in the color-naming task, and ‘‘un-primed’’ stimuli, which were only everseen in the word-reading task. Allportand Wylie found that the reaction time to primed stimuli on switch trialswas slower than the reaction time to unprimed stimuli, but there was noreliable difference in reaction time between primed and unprimed stimuli onrepeat trials. Thus, there was a greater switch cost when the stimulus on theswitch trial was primed. This item specific component of the switch costprovides further evidence for the hypothesis that switch costs involve acarryover effect caused by prior performance of a different task. The findingsuggests that stimuli might themselves evoke the task sets with which they A PDP MODEL OF TASK SWITCHING5were recently associated, even when this task set is not appropriate. As aresult, Allport and Wylie (2000) have updated the ‘‘task set inertia’’ (TSI)theory of Allport et al. (1994). They propose an ‘‘associative-TSI’’ accountof switch costs, according to which stimuli are able to evoke recently associ-ated task sets from memory. When a stimulus appears on one trial, associatedwith taskA, and then reappears on a subsequent switch trial, requiring perfor-mance oftaskB, Allport and Wylie hypothesize that taskset A may neverthe-less be evoked by the presentation of the stimulus (cf. perceptual ‘‘triggerconditions’’ in Norman & Shallice, 1986). The resulting competition be-tween task sets A and B may lead to an extended response selection process,explaining the enhanced switch cost obtained for word reading when thestimulus was recently presented for color naming.Evidence for the Exogenous Control Process AccountIn arguing for the exogenous control process account, Monsell and col-leagues (Monsell et al., 2000; Rogers & Monsell, 1995) have demonstratedthe robustness of the switch cost even when subjects have long intervalsbetween trials to prepare for the upcoming task. They provided evidence thatthere is no further reduction in switch costs after the first 600 ms or so ofthe preparation interval: an asymptotic ‘‘residual switch cost’’ remains (e.g.,Rogers & Monsell, 1995). This is hypothesized to correspond with the timetaken by the execution of the exogenous control process, which must awaitstimulus presentation and is therefore insensitive to the preparation interval.A second line of evidence for the exogenous control process accountcomes from studies which have investigated task switching using the alter-nating runs paradigm when there are more than two trials ofeach task beforea switch. Experiments with run lengths of four (Rogers & Monsell, 1995,Experiment 6) and eight (Monsell, Azuma, Eimer, Le Pelley, & Strafford,1998) trials before each switch have found that the cost of task switchingis confined to the first trial of a run (see Fig. 1). Rogers and Monsell (1995)argue that if switch costs result from a carryover effect from the previoustask, they should dissipate more gradually, rather than being limited to thefirst trial ofa run. Similarly, Monsell (1996) points out that ‘‘although havingample time to prepare before a predictable task switch does not eliminateswitch cost, performing the task just once appears to do so’’ (p. 138). Thispattern of results does not seem to be compatible with the idea that switchcosts simply reflect a form of priming. It can more easily be explained ifswitch costs reflect a one-offprocess oftask set reconfiguration at the begin-ning of switch trials.An empirical argument against the task carryover account has been madeby Monsell et al. (2000). They demonstrated a number of cases in whichthe cost of a switch from a stronger to a weaker task is greater than the costof a switch in the reverse direction, i.e., the opposite asymmetry of switch 6GILBERT AND SHALLICEFIG. 1.ment 6, as a function of the position in a run of four trials, averaged over both tasks, adaptedfrom Rogers and Monsell (1995, Fig. 5).Mean reaction time (RT) and error rate in Rogers and Monsell’s (1995) Experi-costs to that reported by Allport et al. (1994). Thus, it does not seem to bea general rule that switches from a less dominant to a more dominant taskyield greater RT costs than switches in the reverse direction.Ironically, recent evidence (e.g., Meiran et al., 2000; Salthouse, Fristoe,McGurthy, & Hambrick, 1998) has also shown that it is not a general rulethat switch costs are confined to the first trial in a run. Nevertheless, it stillappears to be a challenge to the task carryover account that switch costs canbe confined to the first trial in a run, at least sometimes. Similarly, the findingof larger costs of a switch into a dominant task, even if it is not universal,appears to challenge the exogenous control process account.It is of course possible that the switch cost measures a combination ofboth a carryover of task set (leading to an extended response selection pro-cess) and an exogenous control process (see Meiran, 2000a, 2000b, for suchan account). Indeed, Monsell et al. (2000, p. 254) have pointed out that ‘‘itis perfectly possible that an extra ‘control’ process is required precisely inorder to overcome . . . interference, once it arises as the result ofthe stimulusretrieving a recently activated task-set or the inhibition associated with arecently suppressed task-set.’’ Thus, the task carryover and exogenous con-trol process accounts need not be mutually exclusive. However, before con-sidering a combination of these two accounts, we should first investigate the A PDP MODEL OF TASK SWITCHING7abilities of each one to account for the data on its own. To do this, eachaccount needs to be specified more precisely in orderto assess its consistencywith the available empirical data (of course, a sufficiently vague theory isconsistent with any empirical data). Second, we need to establish where thecrucial differences between the theories lie and which differences are betterclassified as differences in emphasis. Computational modeling can play arole in both ofthese steps. First, by implementing a theory in a computationalmodel, the theory can be specified precisely, allowing a clear assessment ofits ability to explain the empirical data. Second, ifit were possible to producea model onto which two competing theories may be mapped, this wouldsuggest that the difference between the theories is not a fundamental one.A crucial area of disagreement is the extent to which involuntary persis-tence of task sets can explain the dramatic improvement in reaction timefrom the first to the second trial in a run, following a switch oftask. AlthoughMonsell et al. (2000) ‘‘certainly accept that there are relatively long-termcarry-over effects of the kind that Allport and colleagues have demon-strated,’’ they are inclined to doubt whether the dramatic improvement inRT from switch trials to immediately successive nonswitch trials can be ex-plained ‘‘even in part’’ by a carryover effect. Thus, the exogenous controlprocess account does not deny that reaction time in some circumstances maybe affected by a carryover of task set, but it claims that this is insufficientas an explanation of the drop in reaction time from switch to subsequentnonswitch trials.In order to test this claim, we have implemented a version of the taskcarryover account in a computational model. In the remainder of this article,we take an existing computational model of task performance in pure-blockconditions and extend it in accordance with the task carryover account ofswitch costs. We then test its performance in mixed-blocks against the empir-ical data. Our reasoning is as follows. If an existing computational modelof pure-block performance, augmented with the mechanisms required by thetask carryover account of task switching, can produce similar behavior tosubjects in comparable experimental conditions, then the task carryover ac-count would be strengthened. Furthermore, if the model is successful, thiswould indicate that, contrary to the arguments ofRogers and Monsell (1995)and Monsell et al. (2000), a task carryover account can offer a sufficientexplanation ofthe relevant set oftask-switching phenomena. Thus, the exog-enous control process invoked by alternative accounts may be unnecessaryto explain the cost of task switching.The experimental tasks that will be simulated are word reading and colornaming in response to Stroop stimuli. This domain was chosen for threereasons. First, a relatively large corpus of data has been accumulated con-cerning the effects of switching between dominant and nondominant taskson reaction times in domains such as Stroop word reading/color naming(Allport et al., 1994; Allport & Wylie, 1999, 2000; Monsell et al., 2000), 8GILBERT AND SHALLICEbilingual language switching (Meuter & Allport, 1999), and pairs of taskswith different S-R compatibility (Monsell et al., 2000). Thus, there is enoughdata to allow the evaluation of a computational model against the perfor-mance of human subjects. A second reason for this choice is the theoreticalweight attached to the finding of ‘‘paradoxical’’ asymmetric switch costs(Allport et al., 1994; Monsell et al., 2000), which requires a pair of tasks ofdifferent ‘‘strengths.’’ The final reason is that an influential model of theStroop effect has been developed by Cohen and colleagues (Cohen,Braver, & O’Reilly, 1996; Cohen, Dunbar, & McClelland, 1990; Cohen &Huston, 1994; Cohen & Servan-Schreiber, 1992; see also Phaf, Van derHeijder, & Hudson, 1990; Zhang, Zhang, & Kornblum, 1999). Thus, an ex-isting computational model of pure-block performance is already available.This strategy is similar to the approach taken by Logan and Gordon (2001),who added control processes to an existing model ofvisual attention (Bunde-sen, 1990) in order to simulate situations that require switching betweentasks. However, the type of model employed by Logan and Gordon (2001)and the phenomena simulated are rather different from those tackled here,making direct comparison between the two models difficult.THE MODELAlthough the present model is based on the earlier models of the Strooptask by Cohen, Dunbar, and McClelland (1990) and Cohen and Huston(1994), it also has many modifications. Below, we provide a full descriptionof our model, followed by a brief comparison with the earlier models onwhich it is based.We have implemented an interactive activation model (McClelland &Rumelhart, 1981), composed of two separate pathways, for word readingand color naming (see Fig. 2). In each pathway, there are three input units(representing, in the word pathway, the words ‘‘red,’’ ‘‘green,’’ and ‘‘blue’’and, in the color pathway, the colors red, green, and blue). In addition, eachpathway has three output units, representing the responses ‘‘red,’’ ‘‘green,’’and ‘‘blue.’’ In other words, each possible response is represented twice,once in the word-reading pathway and once in the color-naming pathway.Thus, the model has a total of six input units and six output units. Eachinput unit has a positive connection with its corresponding output unit. Forexample, in order to simulate a stimulus of the word ‘‘red’’ written in greenink color, the ‘‘red’’ word input unit and the ‘‘green’’ color input unit wouldboth be activated. This would send activation to the ‘‘red’’ output unit inthe word-reading pathway and the ‘‘green’’ output unit in the color-namingpathway.Processing in the model, i.e., the passing ofactivation between units alongtheir connections, is iterated for a number of cycles. This allows the simula- A PDP MODEL OF TASK SWITCHING9FIG. 2.Architecture of the present model.tion ofreaction time: on each cycle, ‘‘evidence’’ is collected from the activa-tion values of the six output units, two of which represent each possibleresponse, ‘‘red,’’ ‘‘green,’’ or ‘‘blue.’’ When the evidence for one of thesethree responses passes a fixed threshold, the trial is terminated. In this way,it is possible to compare the number of cycles required for the model toreach its response threshold with the mean reaction time of human subjects.The connection strengths from the input to the output units are strongerin the word-reading pathway than in the color-naming pathway. This simu-lates the greater experience that people have of naming written words thancolors. As a result, the word-reading output units become more strongly ac-tive than the color-naming output units when the model is presented witha Stroop stimulus. The evidence for the response represented in the word-reading pathway is therefore greater than the evidence for the response repre-sented in the color-naming pathway and as a result the model will tend torespond by ‘‘reading out’’ the word that it is presented with. However, sincepeople are able to name the ink color of a color-word, even when the colorand word are incongruent, some mechanism is required to prevent the modelfrom always executing the word-reading task. This is provided by the color-naming and word-reading ‘‘task demand’’ units, which send activation totheir corresponding pathways. For example, when the color-naming task de- 10GILBERT AND SHALLICEmand unit is activated, it sends activation to the output units in the color-naming pathway, allowing them to win competition with the output units inthe word-reading pathway. As well as sending a positive input to the outputunits oftheir corresponding pathway, the task demand units also send a nega-tive (i.e., inhibitory) input to the output units of the other pathway.The word-reading and color-naming output units also send activation backto the task demand units. This introduces feedback as well as feedforwardconnectivity into the model, allowing activity in the color and word pathwaysto modulate activity in the taskdemand units. Similar connections were intro-duced into the model of Cohen and Huston (1994) in order to simulate phe-nomena such as attentional capture, where stimuli are able to ‘‘draw atten-tion’’ to themselves (e.g., Posner, 1980). The task demand units receive anadditional ‘‘top-down control input,’’ which specifies which task the modelshould perform. The word and color output units are interconnected, so thatcongruent word and color response units (e.g., the two ‘‘red’’ units) havereciprocal positive connections and incongruent pairs of units (e.g., word‘‘red’’/color green) have reciprocal negative connections. Finally, there arelateral inhibitory connections between all units within the word output‘‘module’’ (i.e., set of word output units), color output module and task de-mand module. This encourages the network to settle into stable states withno more than one unit active in each module.Comparison with Earlier ModelsAlthough the model is based on the earlier models of Cohen, Dunbar, andMcClelland (1990) and Cohen and Huston (1994), its architecture differsfrom them in two main respects. First, there are three possible words andcolors (red, green, and blue) in the present model as opposed to just two inthe earlier models. This was chosen because the Cohen et al. (1990) modelhas been criticized for failing to capture differences between word readingand color naming when the set size is increased beyond two (Kanne, Balota,Spieler, & Faust, 1998; but see Cohen, Usher, & McClelland, 1998, for areply).A second difference is that, unlike our model, the earlier models includeda ‘‘winner-take-all’’ response layer. In the Cohen et al. (1990) model, theunits corresponding to the word and color output units in the present modelsent inputs into a pair ofresponse units. Thus the word-reading ‘‘red’’ outputunit and the color-naming ‘‘red’’ output unit sent activation to a single ‘‘red’’response unit. Evidence was collected from these response units in order todetermine when each trial should end. These additional units are unnecessaryin the present model: since we have interconnected the word and color outputunits there is a beneficial effect of activating congruent output units and adetrimental effect of activating incongruent output units. This plays a role A PDP MODEL OF TASK SWITCHING11in our model similar to the convergent inputs into the response units of theCohen et al. (1990) and Cohen and Huston (1994) models.Implementation ofthe Task Carryover AccountThe model was extended in two ways in order to implement the task carry-over account oftask switch costs. First, rather than reinitializing the networkat the beginning of every trial, the model was modified so that the state ofthe task demand units can persist into successive trials. As a result, the mostrecently implemented task set remains active at the beginning of the nexttrial and the most recently inhibited task set remains inhibited. This can beseen as an implementation of the hypothesis put forward by Allport et al.(1994) that control states persist, involuntarily, from one trial to the next.The most simple way to implement this would be to start each trial with thetask demand units in the state they were in at the end of the previous trial.However, this could well lead to ‘‘perseverative’’ behavior, with the modelunable to switch fromone taskinto the other. Thus, a ‘‘squashing’’ parameterwas introduced so that the activation levels of the task demand units follow-ing each response are squashed, i.e., set to some proportion oftheir activationlevels at the end of the previous trial. Such reductions in activation levelsbetween trials are common in models of sequential processes (e.g., Burgess,1995; Dayan, 1998; O’Reilly & Farah, 1999) and seem to be biologicallyplausible. The task demand units in our model are thought to reflect activityin prefrontal cortex (see Miller & Cohen, 2001) and single-cell recordingstudies of monkeys performing cognitive tasks have demonstrated a sharpreduction in the firing rate ofcells in prefrontal cortex following the produc-tion of a response (see Fuster, 1997, pp. 121–134; for an example in someways akin to the task switching paradigm see Asaad, Rainer, & Miller,2000).An additional requirement for an implementation of the associative-TSIaccount, as opposed to the earlier theory of Allport et al. (1994), is somemechanism for individual stimuli to evoke task sets with which they haverecently been associated. In order to achieve this, we added a connectionfrom each ofthe stimulus input units to both task demand units. Thus, stimuliare able to evoke task sets by providing an extra input into the task demandunits. The weights of these connections between stimulus input and taskdemand units are determined by Hebbian learning at the end of each trial,so that the weights between coactive units are adjusted in proportion to theproduct of their activation values. One potential danger with this learningalgorithm is that the weights between a pair ofrepeatedly coactive units willgrow without bound, which may lead to such a strong input into the taskdemand units that the model is unable to switch task. In order to avoid thisproblem, the weights between stimulus input and task demand units are resetto zero at the end of each trial, before the new weights are calculated, so 12GILBERT AND SHALLICEthat the effects of learning on trial N persist only for trial N  1. This is asimplifying assumption rather than a theoretical position which we adopt.There is strong evidence for such item-specific priming effects lasting forlonger than one trial (e.g., Allport & Wylie, 2000, Experiment 5; Waszak,Hommel, & Allport, submitted). However, long-term priming effects, thoughclearly of great interest, are not addressed in this work.Multiple Inputs to Task Demand UnitsAs well as the control input that indicates which task is to be performed,each task demand unit also receives an input from the stimulus input units,from the color and word output units and from the other task demand unit.When we discuss the model’s performance we refer especially to two ofthese inputs: the control input, which indicates which task is currently appro-priate (we refer to this as the ‘‘top-down input’’), and the input that the taskdemand units receive from the stimulus input units (we refer to this as the‘‘bottom-up input’’). One crucial feature ofthe top-down input is that it is notequal for the two task demand units. We assume that the control mechanismprovided by the task demand units is required more for the color than theword task, since the color-naming pathway is weaker (see, e.g., LaBerge &Samuels, 1974; Posner & Snyder, 1975 for similar ideas). Thus, the top-down input received by the color task demand unit, when color naming isthe required task, is greater than the top-down input received by the wordtask demand unit on word-reading trials. As we shall see, this difference intop-down input plays an important role in the model’s behavior.Operation ofthe ModelThe steps taken to simulate a trial are as follows: (1) For trials other thanthe first, the task demand units are set to a proportion of their activationvalues at the end ofthe previous trial. This proportion is set by the squashingparameter discussed above. The activations of the stimulus input and outputunits are set to zero.1(2) The appropriate top-down input is added to the netinput of the color or the word task demand unit, depending on which taskis required. This input is added to the task demand unit’s net input on everycycle. (3) The preparation interval begins. With all of the stimulus inputunits set to zero, the top-down input is applied...

    关注我们

  • 新浪微博
  • 关注微信公众号

  • 打印亚博足球app下载
  • 复制文本
  • 下载task switching a pdp model.XDF
  • 您选择了以下内容