CONFLICT MONITORING of made with ch hand. Finally,Grasby et al.(1993)had participants listen to an requently re ersed er and m 2 1o fou ACC ame stud Corbetta et al (1991 study,one option is to attribu this findin in the g ion of the ERN there gain oth grew.gre A number of other the bility that ACC activation ma been rel d ERN eversas than do Is with smaller ERNs(Gehri et a provide d by similarity effect,the that pa ed y for the list if it is c sed of similar ading entrie this oh 993)3 Third in a stu where part nts were asked t king then a Z s a ulus in Given such ay,any transient cor veen pro sing in the tim e et al study i ikely to mpanied by the sort of crosstalk to cally similar words studies The close ciation bet onflict Accounting for ACC Activation:Simulation Study l of stu wed above ties that do not fall into any of t egories we laid out nficts in information processing nle D'Es al (1995)used fMRIto c the inwe des ACC activity c either arti culat On the hasis of the efold A firs on of the role of alk ir CC activation in this stud was to make the ac ed so far 0 er divided atte n study.Corbetta.Miezin.Dob 19g yer. second the hile for the nges alc ong th hird goa s to lay the groundwork for furth ling d only one of these dimen In a divided ttention d tor in any of the ations make sly an ition de mor ne of the t ains in w hich activation ssibility parallel evaluation differen 199 under Although t and err nine th th o 1990).the task most of thes models. vith ns like sals are likely to involve e the ve have onal role in trie Iving beh Lain e the t al. 1987),apoi in late
CONFLICT MONITORING 629 the strength of the response made with each hand. Participants very frequently reversed errors, and the EMG results indicated clearly that when this occurred, there was typically temporal overlap between the error and error-correcting responses. This same study provides evidence consistent with the idea that this transient response conflict is n critical factor in the generation of the ERN; EEG data indicated that the ERN coincided with the period of response overlap on error trials. A number of other findings corroborate the connection between response conflict and the ERN. First, error trials associated with the largest ERN amplitudes more frequently involve response reversals than do trials with smaller ERNs (Gehring et al., 1993). Thus, the largest ERNs are associated with error trials where there is the strongest evidence for belated activation of the correct response. Second, an ERN appears, even in association with correct responses, if these are subsequently reversed (Gehring et al., 1993).3 Third, in a study where participants were asked to withhold their responses until 2 s after stimulus presentation, no ERN was observed in association with errors (Dahaene et al., 1994). Given such a delay, any transient competition between processing pathways is likely to have resolved by the time of response delivery. Thus, incorrect responding in the Dahaene et al. study is unlikely to have been accompanied by the sort of crosstalk to which we attribute the ERN. Residual studies. The close association between conflict and ACC activation in the studies we have reviewed is reinforced by the fact that conflict also appears to play a role in ACC activation studies that do not fall into any of the three categories we laid out above. For example, D'Esposito et al. (1995) used fMRI to compare ACC activity during two simple tasks performed either singly or concurrently, observing greater activation in the latter condition. On the basis of the earlier discussion of the role of crosstalk in dual-task performance, it is clear how ACC activation in this study can be explained as a response to conflict. In another divided attention study, Corbetta, Miezin, Dobmeyer, Shulman, and Petersen (1991) measured brain activity with PET while participants monitored forms in a visual display for subtle changes along the dimensions of color, shape, and direction of movement. In a focused attention condition, participants monitored only one of these dimensions. In a divided attention condition, participants searched for changes in any of the three dimensions. Greater ACC activation was associated with the divided attention condition. Participants made more errors in this condition, and so it may be possible to attribute ACC activation in this study to errors. However, another interesting (and closely related) possibility is that the parallel evaluation of different stimulus dimensions led on some trials to crosstalk between pathways supporting "same" and "different" responses. Although the published data do not allow a conclusive evaluation of this possibility, it is consistent with the reported higher frequency of misses (incorrect "same" judgments) in the divided attention condition. In another study, Baker et al. (1996) found ACC activation in association with performance of the Tower of London task. Because the solution to this task is rarely immediately apparent to the unpracticed participant, competition or conflict among alternative actions seems likely to be involved. As we have already noted, certain computational models accord such competition a pivotal functional role in triggering problem-solving behavior (e.g., Laird et al., 1987), a point to which we will return in later discussion. Finally, Grasby et al. (1993) had participants listen to and immediately repeat word lists from 2 to 13 items long. Using PET, they found that ACC activation increased with list length. As in the Corbetta et al. (1991) study, one option is to attribute this finding to errors, for the frequency of errors rose along with list length. However, there are again other potential explanations that involve conflict. One is that, as list length grew, greater response competition occurred during the retrieval process. Even more intriguing is the possibility that ACC activation may have been related to interference among lexical representations being maintained in working memory. One way of examining this latter possibility is provided by the phonological similarity effect, the fact that participants asked to repeat a list of words shows relatively poor memory for the list if it is composed of similar-sounding entries (Baddeley, 1966). If, as appears reasonable, this phenomenon can be assumed to derive from interference among representations being held in working memory, then a potentially informative experiment might be to measure ACC activity during retention of short word lists, comparing activity levels during maintenance of phonologically similar and dissimilar items. If the ACC is responsive to conflict among representations in working memory, then greater activation should be seen in the condition using phonologically similar words. Accounting for ACC Activation: Simulation Study 1 In connection with each group of studies reviewed above, we briefly proposed how ACC activation can be understood as a response to the occurrence of conflicts in information processing. In the present section, we describe a set of three computer simulations in which this interpretation of the ACC literature is more fully articulated. The objectives of these modeling studies were threefold. A first goal was to make the account we have presented so far more explicit, providing a precise indication of what we intend by such terms as crosstalk, conflict, top-down control, and conflict monitoring. A second goal was to confirm the sufficiency of these constructs, as we have used them in the conflict monitoring hypothesis, to account for the results of ACC activation experiments. The third goal was to lay the groundwork for further modeling work, reported in Part 2, that examines the entire feedback loop running from conflict monitoring to cognitive control. Each of the present simulations makes use of a previously and independently implemented computational model of a single task from one of the three primary domains in which ACC activation has been reported. To examine the role of conflict in response override, we consider a model of the Stroop task (Cohen & Huston, 1994); for underdetermined responding, a model of stem completion (McClelland & Rumelhart, 1981); and to evaluate the relation between conflict and error commission, we examine a model of the Eriksen task (Servan-Schreiber, 1990), the task most frequently used in studies of the ERN. To each of these models, we apply a quantitative measure of conflict, allowing the models to be used in simulations of the 3 Although trials involving overt reversals are likely to involve the strongest coactivation of correct and incorrect responses, the account we are proposing does not require that any actual reversal occur, only that activation of the correct pathway occur while activation of the incorrect pathway is still present
630 BOTVINICK.BRAVER,BARCH,CARTER,AND COHEN YationsofcoOmpcingrtprcsntaio5accaualAthouti5a )Berlyne's een shown to engage the General Methods ropy (this involves multiplying en by the average activation the set of competing repr ons).In the has been dra ure of them quasi-empiric em to test the consis y of th whic satisfies Berlyne's criteria while being act and me simpl work as the mumber of free para oursimulation of -Σa,aw ( nts should not be taken to imply that the Smolensky.McClclland.Smo of basic mpatible) ns are cap in the this level Like other co Energy rises models,the c nes w e will consider he ends on the activatior s of the two units.be to a subs hat his imp Whe al input in a final outpu on pro ssing relie As in the cognitives n conflicts hetweer tations ir e net .th were set in previous studies and election.In the simulation studies sented bere itoring In each simulation,the un th the of res ext led by of each model.We not vated in thischo commonality of resp onse selectio This unit takes input m the asic ork and nputes the mined r of the size that of how confict might be afirst step toward defir thesis s addressed in the defined as the simultaneous ls we will co here,ine d informati inhibiting units quantified.Berlyne (197.1960 who discusses this s,itisworhnotiagtatS edcionctvationg with mpetin re of c Hop where units shar
630 BOTVINICK, BRAVER, BARCH, CARTER, AND COHEN behavior of the ACC as it has been observed in brain activation experiments. Each study provides an explicit account of the mechanisms that give rise to conflict, comparing their role across task conditions that have been shown to engage the ACC to different degrees. General Methods Selection of models. Although the models we consider are examined in a novel context, they are not themselves new. Each has been drawn from the published literature and is considered here in its original form. The fact that these models were formulated independently of present hypotheses allows us to approach them quasi-empirically, using them to test the consistency of the conflict monitoring hypothesis with current theories of information processing in specific tasks known to engage the ACC. Leaving the models' original parameters intact and using the same simple computation to determine conflict across all three studies reduces the number of free parameters associated with our simulation of ACC activation to zero. Of course, these points should not be taken to imply that the models used were selected in a disinterested or theory-neutral fashion. On the contrary, the three models implement a shared set of basic assumptions about information processing that also form part of the background for the conflict monitoring hypothesis. Specifically, they assume that information processing is parallel, distributed, and interactive. These assumptions are captured in the connectionist framework, within which all three models were conceived (McClelland, 1992; Rumelhart & McClelland, 1986). Like other connectionist models, the ones we will consider here are composed of identical processing units, each carrying a realvalued activity level, which excite and inhibit one another through weighted connections. When external input is applied to a subset of the units, information propagates through the network, resulting in a final output activation pattern. Information processing relies upon the strength of the network's connections, which can either be set by hand or by a number of training algorithms. Again, these values were set in previous studies and used unmodified in our current simulations. Implementing conflict monitoring. In each simulation, the underlying model adopted from the literature is extended by the addition of a single conflict-monitoring unit (see Figures 1-3). This unit takes input from the basic network and computes the current amount of conflict prevailing there. This component of the simulations raises the important question of how conflict might be measured. As a first step toward defining a method for accomplishing this, conflict may be operationally defined as the simultaneous activation of incompatible representations. In the models we will consider here, incompatible representations (e.g., representations of alternative responses) correspond to units interconnected by inhibitory weights. Thus, conflict can here be defined as the simultaneous activation of mutually inhibiting units. Although this makes it clear what conflict involves at a qualitative level, it is a more difficult question how conflict should be quantified. Berlyne (1957, 1960), who discusses this issue at length, offers a list of desiderata for a measure of conflict: (a) It should increase with the absolute activation of the competing representations; (b) it should increase with the number of competing representations; and (c) it should be maximal when the activations of competing representations are equal. Although it is an empirical question how conflict might be measured by the brain (a point we consider further in the General Discussion), Berlyne's criteria provide a reasonable starting point for considering alternative possibilities. Berlyne noted that there are many potential measures of conflict that would meet his specifications. He himself adopted one based on the information-theoretic expression for entropy (specifically, this involves multiplying entropy by the average activation in the set of competing representations). In the present context, this approach carries the technical disadvantage that it requires activation levels to be translated into probability values, a step that in turn requires peripheral assumptions. In the present studies, we chose a different measure of conflict— Hopfield energy—which satisfies Berlyne's criteria while being based on values specified directly by the models we examine. Hopfield (1982) defined the energy in a recurrent neural network as (1) where a indicates unit activity and both subscripts are indexed over all units in the set of interest (related measures are discussed by Rumelhart, Smolensky, McClelland, & Hinton, 1986, and Smolensky, 1986). To see how energy reflects conflict, consider a single pair of mutually inhibiting (incompatible) units. When both are inactive, energy is zero, consistent with the absence of conflict. Energy remains at zero if only one of the units becomes active, once again mirroring the level of conflict. Energy rises above this level only if both units are active. The particular value for energy then depends on the activation values of the two units, becoming largest when both units are maximally active and thus most strongly in conflict.4 Note that his implementation of conflict does not involve any additional parameters, and this preserves the zero-parameter nature of our simulations. As in the cognitive system, conflicts between representations in connectionist networks can occur at a variety of levels of processing, including stimulus evaluation, memory and set representation, and response selection. In the simulation studies presented here, we focus exclusively on the role of response conflict, measuring energy over units in the output layer of each model. We were motivated in this choice by the commonality of response selection processes among tasks that involve response override, underdetermined responding, and error commission, which led us to hypothesize that ACC activation in these domains might be accounted for in terms of conflict at this level of processing. Although this is the hypothesis addressed in the simulation studies, there are reasons for leaving open the possibility that conflict at other levels of processing might also be relevant to ACC activation, a point to which we return in later discussion. Simulation procedure. In each study, the underlying model is used to simulate information processing in conditions that have been reported to engage the ACC to different degrees. In the 4 For completeness, it is worth noting that concurrent activation of units interconnected by excitatory links causes a reduction in energy. In this regard, energy is more than a measure of conflict; it measures compatibility or consistency. This interesting aspect of Hopfield's (1982) formula does not come into play in the simulations to be reported here, where units share only inhibitory connections
CONFLICT MONITORING 63 em co letionm nd in rval. mpared ept wh re exp itly not t s n in Figure 1.the current simulation added a ed fo step of pr mon se l ring unit as vation level tudy.the of this ask.the acti the strongest ACC activation. Simulation IA:The Stroop Task the conflic ongruent cond s this is be both output un tow rd thei o)r esting monitoring-in a ntation is rem oss condit mde ta osed b appear ciate are d H he m puts eac its in ea nomeicsaast conflic layer are by symmetrical ved to engage on provides an 12 Q.6 t panel:I o of the Su H 46 T.A. dge,M MIT Press.C RihtpeEneny sured in th En cyele of proce ing.and the data hown are mean
CONFLICT MONITORING 631 Stroop model, congruent, neutral, and incongruent trial conditions are compared; in the stem completion model, the stem completion task is compared with word reading; and in the Eriksen model, correct responses are compared with errors. Except where explicitly noted, simulations are run according to the procedure originally used for each of the basic models as reported in the literature. With each step of processing, the conflict monitoring unit assumes an activity equal to the current level of energy in the output layer of the underlying model. In each study, the activity of this unit is compared across conditions, with the prediction that the greatest activity will be observed in the condition associated with the strongest ACC activation. Simulation 1A: The Stroop Task In this first simulation study, we introduce the basic elements of the proposed framework by considering the origins of ACC activation—and, by implication, the role of conflict monitoring—in a response override task. Method. Stroop performance was simulated using a model proposed by Cohen and Huston (1994), shown in Figure 1 (left). This model is based on an earlier feed-forward model (Cohen et al., 1990), revised to include recurrent connections and interactive processing (both of which are amenable to our measurement of conflict). The model includes input units for display color and word identity. The appropriate units in each group connect reciprocally via excitatory weights to an output layer with units representing potential responses. In addition, the model includes a task demand layer with units standing for word reading and color naming, respectively. The task demand units serve to bias activation in the model so that either word or color inputs may dominate response activation. As shown in Figure 1, units within every layer are interconnected by symmetrical negative weights. The procedure used in simulating a trial is detailed in Cohen and Huston (1994). Briefly, one of the task units is activated during an initial priming interval, during which the output units are inhibited to prevent premature responses. The input pattern is then applied and the response-layer inhibition removed. As illustrated in Figure 1, the current simulation added a conflict monitoring unit that takes inputs from the response layer of the underlying model, taking on an activation level equal to the energy in that layer on the current cycle of processing. In order to account for the findings regarding ACC activation in neuroimaging studies of the Stroop task, the activation of the conflict monitoring unit was evaluated during simulation of incongruent, congruent, and neutral conditions in the color-naming task. Results. Results are shown in Figure 1 (right). As with ACC activation in neuroimaging studies of the Stroop task, activation of the conflict monitoring unit was higher in the incongruent condition than in the congruent or neutral conditions. As shown in the figure, activation rose rapidly in all three conditions; this is because both output units move toward their (nonzero) resting activity levels once the inhibition they receive prior to stimulus presentation is removed. Differences in energy across conditions soon appeared, however, with incongruent trials associated with the highest degree of energy. The increased activity of the conflict monitoring unit on incongruent trials reflects the occurrence of crosstalk within the Cohen and Huston (1994) model. On incongruent trials, word and color inputs each activate a different set of units in their corresponding pathways. The intersection of these two pathways in the output layer (in addition to the other sectors of the model) causes conflict between the response units, and this in turn raises the activity of the conflict monitoring unit. Discussion. Response override tasks have been repeatedly observed to engage the ACC. This first simulation provides an Conflict ffitofiftorinQ RMDORM Ink Color Task Dwnand Word 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 Cycte Figure 1. Left panel: Illustration of the Stroop model discussed in Simulation Study 1. From "Progress in the Use of Interactive Models for Understanding Attention and Performance," by J. D. Cohen and T. A. Huston, 1994, in C. Umilta and M. Moscovitch, Attention and Performance XV (Figure 18.8, p. 462), by J. D. Cohen and T. A. Huston, 1994, Cambridge, MA: MIT Press. Copyright 1994 by MIT Press. Adapted with permission. Eq. = equation; R = red; G = green; C = color-naming; W = word-reading; N = neutral. Right panel: Energy as measured in the response layer of the Cohen and Huston (1994) model during simulation of congruent, neutral, and incongruent trials in the Stroop task. Arrows indicate average response times. Energy was recorded for each cycle of processing, and the data shown are means based on 100 trials in each condition
632 BOTVINICK.BRAVER.BARCH.CARTER,AND COHEN a third set o ndition of the Stoop task isa g.and in fact the byinhibtC through the co enting letter s blan the 3- ontrolsenalcoitagr m the olor-naming task. weaken or-naming uni e tn ng inte nd thus to high .Here.he assumes an activati on equal to the curer also trar stable I ACC on (partial inp activation during incong ntedintewod-unitrlayrerof In the w d-rea n uers of e nalneuroinmaging A oral studies (. troop interf ence effect if in one are ran esuls.canbctndc tood as de the final outcome in the word for sedtotheword-readingtask)The Carter ()s full wordsand paths by which reache ently.Be oral set of other words (e.g the FSin Fmight tivate FIST),the ls when were rare.Eve vord un ciated with wor y ng the input.For word stems,pro trials were rare. two units ese wor Simulation IB:Stem Completion ter units associal function.Using the approach taken in Simu tion task odelsimdlation2Bmhepeeataniclkealocontain
632 BOTVINICK, BRAVER, BARCH, CARTER, AND COHEN illustration of how the idea of conflict monitoring can be used to explain this finding. When an element is added to an existing model of a typical response override task, acting to transform the occurrence of conflict into an activation-based signal, a pattern is observed across conditions that parallels that observed in ACC activation studies. The color-naming condition of the Stroop task is a classic example of controlled information processing, and in fact, the Cohen and Huston (1994) model was originally proposed as a basic model of control function. Control is implemented here through the color-naming and word-reading units, insofar as these units bias information flow through the rest of the system in accordance with task demands. It is interesting that varying the control signal coming from these units impacts the degree to which conflict occurs during stimulus processing. In simulations of the color-naming task, specifically, weakening the input from the color-naming unit on incompatible trials leads to increasing interference between color and word inputs and, thus, to higher peak energy.5 This aspect of the model's behavior fits well with the idea that conflict might serve as an indicator of insufficient control, as it means that conflict is most likely to occur when control is weak. It also translates into a testable prediction: If ACC activation reflects conflict detection, then, on the basis of the model, ACC activation during incongruent trials in the Stroop task should vary inversely with the strength of control, defined as the effort to attend exclusively to color. We recently tested this prediction in a functional neuroimaging study (Carter et al., 2000). Here, the strength of top-down control was influenced indirectly by manipulating trial-type frequency. As shown by a number of behavioral studies (e.g., Lindsay & Jacoby, 1994; Logan & Zbrodoff, 1979), participants display a smaller Stroop interference effect if incongruent trials are frequent than if they are rare. In our terms, frequent incongruent trials lead to a high-control state (a tight focus on the color-naming task as opposed to the word-reading task). The Carter et al. (2000) study exploited this phenomenon to test for the predicted relationship between control state and ACC activation. Participants performed the Stroop task while undergoing fMRI. Trial-type frequency was varied across blocks; in one half of the blocks, incongruent trials occurred frequently, in the other half, relatively infrequently. Behavioral results confirmed the expected effect of trial-type frequency on control state. Participants were faster on incompatible trials when these were frequent than when they were rare. Eventrelated scan acquisition allowed evaluation of the time course of ACC activation on individual trials. As predicted, peak activation on incongruent trials differed as a function of trial-type frequency, with greater activity occurring during blocks where incompatible trials were rare. Simulation IB: Stem Completion As in response override tasks, we have attributed ACC engagement in underdetermined responding tasks to the engagement of a conflict monitoring function. Using the approach taken in Simulation 1A, this proposal was tested against a relevant model of information processing, in this case a model of the stem completion task. Method. Stem completion can be simulated using the interactive activation (IA) model of word reading introduced by McClelland and Rumelhart (1981; Rumelhart & McClelland, 1982), illustrated in Figure 2 (left). The model consists of three interconnected sets of processing units. External input is applied to a layer encoding featural elements—vertical, horizontal, and diagonal line segments—from which individual letters are constructed. Activation feeds forward from this feature layer to a layer of units representing individual letters. This layer connects to a third set of units, each standing for an individual four-letter word. Between layers, compatible units (e.g., the unit for the letter A in the first slot and the word unit for ALSO) are connected by excitatory weights, and incompatible ones by inhibitory weights. There are also symmetrical inhibitory connections between each pair of units in the word layer. Stem completion can be simulated in the IA model by presenting letters in the first two positions, leaving the third and fourth slots blank. Given such input, the model completes it, settling into a final state dominated by a word unit (and corresponding letter units) representing a word beginning with the two letters presented, similar to what would have resulted had all four letters of the word been present. As in Simulation 1A, a conflict monitoring element was added to the underlying model. Here, the conflict monitoring unit takes its input from the units in the word layer and assumes an activation equal to the current level of energy in that layer. In order to account for the finding of ACC activation in association with stem completion, the activity of the conflict monitoring unit was evaluated during simulations of both stem completion (partial input) and word reading (full input). A total of 20 words were chosen at random from the corpus represented in the word-unit layer of the model. In the word-reading condition, each word was presented in full to the feature layer. In the stem-completion condition, only the first two letters of each word were presented, with the last two slots receiving no input. As in Simulation 1A, energy was measured at regular intervals throughout each trial. Results. Results are shown in Figure 2 (right). Whereas presenting a full word led to only a fleeting rise in energy, stem presentation led to much greater and sustained levels of energy. As in Simulation 1A, these results can be understood as deriving from the different degrees of crosstalk involved in the two task conditions. Although the final outcome in the word layer is similar for full words and word stems, the paths by which the network reaches its final representation entail quite different amounts of crosstalk. For full words, the process is fairly straightforward. The input for each letter activates its corresponding feature units and letter unit. The selected letter units together strongly activate one word unit. Although subgroups of letters might also weakly activate a small set of other words (e.g., the FIS in FISH might activate FIST), the support for the fully specified word is stronger, and this word unit quickly dominates the word layer. The small increase in energy associated with word reading corresponds to the minor conflict among words partially matching the input. For word stems, processing unfolds differently. Initially, the input activates one letter unit in each of the first two letter positions. These two units together activate a wide range of word units (FI_ will activate FISH, FIND, FINE, FIRE, etc.). These word units compete through inhibitory interconnections, also sending activation to the letter units associated with them in the third and fourth positions of the letter layer. Although this conflict is ultimately resolved in 5 The effects of varying task-unit input were first explored in an earlier version of the model by Cohen et al. (1990). Usher and Cohen (2000) have replicated and extended these findings in the context of the Cohen and Huston (1994) model. Simulation 2B in the present article also contains relevant findings
CONFLICT MONTTORING 633 0.07 E4.1 0.06 0.5 ⊙⊙⊙D :@O⑥o@OO 0.02 0.01 @⊙@可@⊙@⊙ Right pan 38 d layer of the IA m del during simu greater than for word reading.giving rise to the greater amount of io.Thompson-Schilt(19) Underdetermined responding tasks make upan se by that of the wn to activate the h response strength io as an index of the degree to which eact fMRL the mant respons the models we have adopted.it in both types of task isth aitongbyri sls,greater activation was superficialdifcreneces This finding was recently replicated by Barch et al.(2000)ina tudy that also tested f rther predictions based directly on the system state. sof stem letiot del's the way tie fact that,in the IA mode the specific stem tested.Energy varies with the degreeto which ten the d equal in stren th)." s had the esocaed st conflict.b se the sideredinthecontexofthcconnictmoaitoringypothcsi inginewcwdrta ing and ste mplitude than th ould )dg the di o the ongly pre respon inpu
CONFLICT MONITORING 633 Conflict monitoring 0.07 I I'l l I 1 I 1 * Cycto Figure 2. Left panel: Illustration of the IA model. From "An Interactive Activation Model of Context Effects in Letter Perception: Part I. An Account of Basic Findings," by J. L. McClelland and D. E. Rumelhart, 1981, Psychological Review, 88, Figure 3, p. 380. Copyright 1981 by the American Psychological Association. Eq. = equation. Right panel: Energy as measured in the word layer of the IA model during simulation of word reading and stem completion. *Energy during simulation of stem completion with the weight of inhibitory connections in the letter layer set to match those in the word layer, in order to prevent two-way ties between words. Energy was recorded every five cycles of processing. favor of one word that completes the input pattern,6 the degree of crosstalk sustained in the response selection process is much greater than for word reading, giving rise to the greater amount of energy observed. Discussion. Underdetermined responding tasks make up an important subset of the tasks that have been shown to activate the ACC. The results of the present simulation demonstrate how the engagement of the ACC in this setting can be understood in terms of a conflict monitoring function. This result stems from the same factor that produced the results of Simulation IA. On the basis of the models we have adopted, conflict in both types of task is the result of crosstalk between processing pathways. Thus, despite the superficial differences between response override and underdetermined responding tasks, the ACC activation associated with both can be understood as a response to precisely the same type of system state. As in Simulation 1 A, consideration of the factors that affect the degree of conflict in the underlying model leads to testable predictions. One example involves the fact that, in the IA model, the degree of crosstalk associated with stem completion depends on the specific stem tested. Energy varies with the degree to which words other than the eventual winner are excited by the stem. Stems that activate one completion much more strongly than any other will be associated with the least conflict, because the preferentially activated word unit quickly suppresses its competitors. Considered in the context of the conflict monitoring hypothesis, this leads to the prediction that stem completion should engage the ACC more strongly when the stem presented is associated with several completions than when the stem is associated with one strongly preferred response. A finding related to this prediction has been reported in the context of another underdetermined responding task, verb generation. Thompson-Schill et al. (1997) recorded the frequency with which specific responses were elicited in this task by a set of nouns. For each noun, they divided the frequency of the most frequent response by that of the second most frequent, using this response strength ratio as an index of the degree to which each noun was associated with a single predominant response. Using fMRI, the group compared brain activation during completion of stems with high and low response strength ratios. Consistent with the conflict monitoring hypothesis, greater ACC activation was observed for low-response-ratio nouns. This finding was recently replicated by Barch et al. (2000) in a study that also tested further predictions based directly on the 6 In simulations of stem completion using the model's original parameters, the settling process sometimes resulted in a two-way tie between word units. This is reflected in a plateau in the average energy trajectory, as shown in Figure 2. We tested whether the occurrence of these ties might be responsible for the higher levels of energy during stem completion by introducing reciprocal inhibitory weights between each pair letter units, similar to those in the word layer (and equal in strength). This had the effect of eliminating deadlocks, but without otherwise affecting the differences between processing in the word-reading and stem-completion conditions. The resulting energy trajectory, shown as a dashed line in Figure 2 (right), remained significantly greater in amplitude than the baseline condition, confirming that the differences in energy between the two conditions were not due to the incidental occurrence of two-way ties, but instead to the transient competition among processing pathways triggered by presentation of the word-stem input