Contrasting domains in the control of action the routine and the non-routine

It is argued that the systems for the Supervisory System’s top-down selection of schemas in contention scheduling has a different lateralisation of dorsolateral prefrontal cortex from th

Contrasting domains in the control of action: the routine and the non-routine

Tim Shallice

University College London and SISSA Trieste

Correspondence to: T Shallice, Institute of Cognitive

Neuroscience, Alexandra House, 17 Queen Square,

London WC13AR, UK

((Association Lecture of Attention & Performance XXI:Processes of Change in Brain andCognitive Development (ed M Johnson & Y Munakata), OUP, 2006.))

The Supervisory System model in which there are two cognitive levels in the control action is assessed It argued that there is a modulatory relation between the levels It is further argued that standard connectionist variables such as age of acquisition, familiarityand frequency are particularly useful for characterising behaviour produced by contentionscheduling, the lower-level system, when Supervisory System function is impaired By contrast, an analogy with symbolic AI models is used to theoretically motivate a

fractionation of Supervisory System processing as created by a set of functionally

selective and anatomically partially separable subsystems It is argued that the systems for the Supervisory System’s top-down selection of schemas in contention scheduling has

a different lateralisation of dorsolateral prefrontal cortex from the systems concerned with non-evident error detection and checking The former are held to be the more left lateralised by comparison with the latter

1 Introduction

The idea that there is a hierarchical organisation of the processes that control action with the higher levels modulating the operation of the lower ones is very old, going back at least to Hughlings Jackson Moreover in the more neurobiological versions it is

commonplace to view the prefrontal cortex as the summit of the hierarchy (e.g

Luria,1966; Fuster, 1989; Dehaene & Changeux, 1997; Miller & Cohen, 2001.)

A second very common idea in experimental psychology is that there are two domains of the control of action – automatic and non-automatic (controlled) (e.g Shiffrin &

Schneider, 1977) - and there are related models in developmental psychology (eg

Karmiloff-Smith, 1986) The model of Norman & myself (1980, 1986) (see also

Shallice, 1982) essentially combines these two ideas in proposing two domains of

processing – that of the Supervisory System and of contention scheduling - with the former only realising its effects through modulation of the latter and with the SupervisorySystem localised in prefrontal cortex

In these respects the Norman-Shallice model may be thought of as merely one variant of the combination of two now standard, although not universally accepted, perspectives in cognitive neuroscience It has, though, an additional rather different conceptual

dimension It was developed not only from a conflation of the experimental psychology and the neurobiology of levels of action control; it also represented in two respects an interface between two different modelling traditions - the connectionist, at that time represented by interactive activation modelling (e.g McClelland & Rumelhart, 1981) and

of symbolic AI

Thus as far as the selection of which schemas control the processing and effector systemsthey require, contention scheduling, the system the Supervisory System modulates, is viewed as operating in an interactive activation fashion with units corresponding to overlearned single motor or cognitive skills – action and thought schemas In routine mode schemas receive activating input from both higher-level (source) schemas and fromobject-trigger systems In nonroutine mode additional activation to schemas is provided

by the Supervisory System However, in addition it is also useful to conceive of the overall system within a more symbolic framework as operating in an analogous fashion

to production systems The symbolic aspect is used principally in the processes that follow selection and in particular how the “arguments” of schemas are set on selection – where, with what and on what the thought or action skill operates – which depend upon the simultaneous state of object representation systems Moreover, the initial verbal account has been realised more recently computationally in interactive activation

simulations of Cooper & Shallice (2000) based on the everyday task of coffee

preparation

The intellectual origins of the Supervisory System concept were different In earlier discussions of the automatic/controlled distinction (e.g Shiffrin & Schneider, 1977) what processes lead to controlled processing was only vaguely specified In the Norman-Shallice model it was held to be a materially and conceptually separable system

Conceptually that different systems could be involved in routine and non-routine

operations came from classical artificial intelligence There the idea was quite standard that in addition to the processes used for the effecting of routine selection of routine operations, there are special processes that come into play in situations where routine responding does not lead to the attaining of goals (e.g Sussman, 1975; Newell, 1990) The main thrust of the model developed by Norman and myself (1980, 1986) was to argue for a prefrontally located Supervisory System, coming into play in non-routine situations to modulate the operation of a system which effects routine action – contention scheduling Thus the distinction between the situations in which the two types of system come into play was derived from symbolic AI

One may view the contrast in the computational principles on which the two systems operate from a related but not identical perspective Perner (2003) has argued that the lower-level system representations involved in contention scheduling are implicit, as theyare procedural representations By contrast he argues that the higher (Supervisory) level

“is defined by the necessity to entertain predication and fact-explicit representation and toexercise content control over the lower level” (p225) He illustrates this with the

example of a child given the instructions ‘Put the green cards into the left box’, where thechild cannot represent their meaning in a ‘predication-implicit’ way, as no card is actually

being presented He continues “That also means that not only predication to instances but also that they are not real but only hypothetically considered instances needs to be made explicit … The same explicitness is, of course, also required for planning,

reasoning and entertaining hypotheses before one can come to a conclusion which action sequence is best to employ.” (p225)

It is possible to consider the computations carried out by the cognitive subsystems on twodimensions One concerns the number of input variables that need to be taken into account and the complexity of their interactions As these increase computational

procedures which optimise constraint satisfaction, ones using gradient descent principles are likely to be optimal ie systems operating on broadly connectionist principles, of which interactive activation models are a simple version The second dimension is the degree to which the values of intermediate products, of input variables themselves, and indeed which are the critical input variables may be subject to revision As this

characteristic increases so the value of having explicit how and why intermediate

products are arrived at becomes increasingly valuable One basic theme of the paper is that high values on the first dimension are more critical in the computations of contentionscheduling and high values on the second for those of many aspects of the Supervisory System

More recently the model has been developed by Stuss et al (1995) and Shallice &

Burgess (1996) to confront a major conceptual inadequacy in the original model How

completely unspecified Thus the concept is derided by Dennett (1998) as ‘an ominously wise overseer – homonculus who handles the hard cases in the workshop of

consciousness’ (p288) Since the logic of the original paper was partially derived from the idea that the postulating of homunculi of reduced power could be progressive, itself derived from Dennett (1978), the source of this criticism was rather odd However the sentiment was common (see e.g Baddeley, 1996) One major strand of development of the model has been to confront this objection However this has principally been done byanalogy with the deeply unfashionable conceptual framework which was critical in the initial development of the model, namely symbolic artificial intelligence The essence of the Mark II model of Shallice & Burgess (1996) was the assumption that in confronting non-routine situations a number of qualitatively very different types of computational operations are required which were held to be the province of anatomically separable higher-level subsystems It further used the assumption of functional specialisation, which can be broadly but unrigorously specified, as that if the phenotype of homo sapiensincludes cognitive tasks sufficiently different in their computational requirements from all others in its repertoire then they would be implemented at least in part in separable regions of cortex ((Footnote: For simplicity I will adopt the terminology of the Shallice-Burgess paper.))

Four interlinked issues are addressed through the paper One major issue is whether the distinction between higher and lower-level control of action is well captured by the idea

of their being the provinces of different systems with ‘modulation’ being an appropriate characterisation of their relation The second major issue whether “routine-ness” is an

appropriate concept to use to contrast the different properties of the two levels of

systems The third concerns the relevant saliency of the constraint satisfaction and explicitness dimensions respectively Thus I will argue that a broadly connectionist framework is particularly useful in characterising the lower-level contention scheduling system, as discussed above This is considered in the next section Then I will address whether symbolic AI ideas are of value when considering the Supervisory System; in particular I will consider the perspective derived from symbolic AI that the” ominously wise overseer-homonculus" can be tamed by its being fractionated into different

subsystems This is addressed in the following two sections The final issue, considered

in the last section, is whether the division relates to the phenomenological one between willed and automatic action

2 Basic Findings: Associationism and Contention Scheduling

In functional imaging, that the activation of prefrontal cortex declines as tasks become less novel has been shown in a variety of paradigms (e.g Raichle et al.,1994; Jueptner et

al, 1997) Moreover lesions to prefrontal cortex affect the ability to confront non-routine situations appropriately (e.g Shallice & Evans, 1978; Knight, 1984) This fits with a prefrontally localised system being critical for non-routine tasks However while the absence of such evidence would undermine the current approach, it could be explained in

a variety of ways; for instance non-routine situations may make greater demands on working memory

Stronger evidence is provided by the findings that lesions to prefrontal cortex lead to behaviour characterisable as that controlled by contention scheduling alone Through thelast 100 years of its history psychology has been concerned to model routine operations –initially through S-R psychology, then somewhat separately in more computational form through production systems and in a more neurobiological one through connectionism A key aspect of the model is the role that stimulus-response (action schema) associations play in the genesis of action For such frameworks, factors such as frequency,

familiarity, age-of-acquisition, priming, interference and so on become important

dependent variables Predictions made from accounts of contention scheduling inherit this tradition A key aspect of the model is the role that stimulus-response (or better trigger-stimulus representation to action-schema) associations play in the genesis of action So prefrontally impaired behaviour should be especially sensitive to: (i)

familiarity, (ii) age-of-acquisition and (iii) (implicitly) frequency-of-application of a rule

In a variety of prefrontal syndromes the patient’s actions are behaviours triggered by stimuli with which they are strongly associated even when they have been instructed not

to respond in this way and gain nothing by so doing These include the grasp reflex (De Renzi & Barbieri, 1992), utilisation behaviour (Lhermitte, 1983; Shallice et al, 1989), the anti-saccade type of tasks (Paus et al, 1991) and forms of the anarchic hand syndrome (Della Sala, Marchetti & Spinnler, 1991; Humphreys & Riddoch, 2003) In such

syndromes action selection can be at the level of effector selection (Riddoch et al, 1998)

or of so-called ‘source’ schema (Shallice et al, 1989)

The analogue of these behaviours in problem-solving is the especial vulnerability of frontal patients in situations in which the elicitation of ‘capture errors’ (Reason, 1979; Norman, 1981) is potentiated by stimulus displays (Della Malva et al, 1993) In other words prefrontal patients have difficulty in suppressing inappropriate responses triggered

by familiar S-R bonds in contention scheduling; they cannot overcome potentiated but incorrect responses A second relevant problem-solving phenomenon is the strikingly good performance of prefrontal patients in situations where rule abstraction is required if the rule is one which corresponds to the inbuilt tendencies of the contention scheduling system Thus Verin et al (1993) found that an alternation rule was attained more rapidly

by prefrontal patients than normal adults when the subject must make one choice once every 15 seconds; the prefrontal patients made virtually no errors Moreover 5- and 6-year old children but not those of 7+ also found the task very easy (Houde et al, 2001), supporting the assumption that the contention scheduling system obeys age-of-acquisitionprinciples

Where frequency is concerned perhaps the most direct support is provided by the use of the random generation task In the random generation task the subject must produce a sequence of numbers which approximate randomness as closely as possible; that is the responses must not satisfy any given rule inappropriately frequently Baddeley (1986) analysed the task He argued that as there were no external stimuli, any schema

controlling obeying a rule which is operating to elicit any one response must be inhibited prior to the next response and an alternative schema activated On the model both these

comparison with a sham TMS control a single TMS pulse to the left dorsolateral

prefrontal cortex led to a rough doubling in the rate of responses which involved counting

up or down in ones In other words a response which followed the most frequently applied rule occurred much more often following rTMS to left prefrontal cortex, even when it was clearly inappropriate Moreover using fMRI Jahanshahi et al (2000) found that as the speed of generation was increased so the degree of randomness plunged dramatically as rates increased from 1 per 3s to 1 per 0.5s The curve was mirrored in theactivation of one area of cortex, again the left dorsolateral prefrontal cortex (see fig 1) Thus with the proviso, to be discussed later, that the effects were essentially restricted to one part of the prefrontal cortex – the left dorsolateral region – the predictions were well corroborated using this paradigm

These neuropsychological effects of activating selection of action schema only from representations of trigger stimuli - a S-R type of operation – fit well with a phenomenon described from the normal literature on task switching Allport and Wylie (2000)

investigated how switching was affected by the specific stimuli used on a previous run of the other task Subjects alternated between short runs of two tasks – colour naming and (colour) word reading – so creating a Stroop-type situation However only some of the colours presented for naming had occurred in the preceding (colour) word reading task Allport & Wylie found that on ‘switch’ trials, but not on non-switch trials, subjects were slower to give the first colour naming response if the name had been previously primed than if it was in the ‘unprimed’ set It is on switch trials that the schema unit controlling the effecting of one task set has to be inhibited and a second unit activated which controls

the other task-set The pattern of results is well simulated on an interactive activation model which is essentially isomorphic with contention scheduling in the relevant respect (see Gilbert & Shallice, 2002).

The Allport and Wylie result is only one of a range of results from the classical

Attention-Performance areas reviewed by Hommel (2000) that show that S-R effects can occur independently of the conscious intentions of the subject and be in conflict with them Moreover in line with standard associationism such effects of irrelevant S-R associations increase with practice (McLeod & Dunbar, 1988) At the same time

Hommel also showed that “automatic” S-R translation effects do not generally occur independently of the conscious intentions of the subject, in that the degree to which they are manifested depends typically upon the conscious intention of the subjects concerning the task that they are attempting to carry out As will be discussed in the last section, on the current approach conscious volitional responses require the Supervisory System The first type of effect Hommel described then corresponds on the model to a situation where

a schema is selected which had activation derived only from triggering by external stimuli in contention scheduling and producing responses in conflict with those of other schemas elicited by top-down modulation from the Supervisory System However, typically the two types of environmentally driven and top-down influences would

combine to lead to schema selection, so giving rise to his second type of phenomenon Moreover the ‘determining tendency’ conceptual framework of Ach (1905), to which Hommel appeals, is as we will see entirely compatible with the current model

3 The Fractionation of the Supervisory System

How does the cognitive system enable subjects to cope with non-routine situations? In symbolic AI, architectures based on classical planning systems were found to be slow, cumbersome and fragile in their operation (see Russell, 1995, for review) This led to their replacement by decentralised reactive-planning systems such as those of Brooks,

1991 However more recently a number of defects of reactive-planners have become apparent and researchers have developed three-layer architectures (e.g Gat, 1991;

Elsaesser & Slack, 1994) (see Glasspool, 2004) In a three-level planner, such as that of Gat (1998), the bottom layer is a Brooks-style reactive controller The highest level is a

“deliberative” system capable of reasoning and goal-directed behaviour It can be viewed

as performing the overall function of the Supervisory System The middle layer

assembles sequences of single behaviours and enables higher-level goals to be achieved

by different lower-level behaviours depending on the overall situation Glasspool has argued that the middle-layer of the Gat-type architecture is analogous to contention scheduling on the current approach So current planning frameworks used in symbolic

AI have a set of systems on different levels which resemble those in the Norman-Shallicemodel

A system for coping with non-routine situations does not have to be internally unitary Indeed in a recent model of the deliberative level used in medical expert decision-making– the Domino model – Fox and Das (2000) argue that six different computational

domains are required in order to carry out the key processes necessary from the detection

of the existence of a potential symptom to the execution of the care procedure The

processes may however be called recursively Fox and Das argue that these six different domains each need to be underpinned by a different logic, that is that the logical

constraints which the implementation aims to satisfy involve six different sets of axioms

There are a number of approaches to conceptualising executive functions which have an origin in symbolic AI; one example is EPIC (Kieras et al, 2000) In this paper I will follow the Supervisory System mark II model (Shallice & Burgess, 1996) (see fig2) where this conceptual approach is combined with the use of dissociation logic to isolate subsystems neuropsychologically However, the model fits well with the stages

outlined by Fox and Das, namely:

(i) The first computational domain is that required in the articulation of a goal

This involves abstracting from the current situation aspects that are important

to improve

(ii) The second is that involved in the production of one or more procedures for

attaining the goal

(iii) The third is selection between them and the consequent decision to act (iv) The fourth is the articulation of the selected procedure into a sequence of

implementable steps

(v) The fifth is the realisation of the steps as actions

(vi) The final one is the checking that these actions are indeed ameliorating the

situation, that is they assist in realising the goal or goals If not, the cycle begins again with a new goal of patching the procedure or aborting it

Such a division of the whole process into a series of stages of this type is far from

original to the Domino model (see e.g Ben-Yishay & Diller, 1983) However what is original is that each of the stages is held to correspond to different domains which are each sufficiently computationally distinct to require a separate underlying logic

The current approach, which is based on the the Supervisory System mark II model presupposes that processes requiring the Supervisory System are involved in each of these stages However only the processes involved in the second, fifth and sixth of the above stages will be considered in this paper The approach aims to show that in carryingout these stages, processes involving the Supervisory System are used which are partially anatomically distinct from each other This partial anatomical separability corresponds to the different underlying logics on the Fox and Das model Whether other stages also involve separable processes in prefrontal cortex remains to be investigated

Consider the second stage the articulation of a possible procedure for attaining the goalspecified in the first stage, when this is not directly elicited through the use of learned contingencies by the combination of situation and goal On the model the procedure corresponds to a series of outputs from the Supervisory System which lead to the

activation of a set of schemas in contention scheduling For reasons to be given in the final section, outputs from a Supervisory System to contention scheduling correspond to conscious states of the organism Articulated procedures then are ‘strategies’ Strategies which normal subjects typically develop and apply in confronting a novel situation are much less used by frontal patients (see e.g Owen et al, 1990; Burgess & Shallice, 1996a)

For instance, in the Owen et al (1990) study, subjects were presented with a set of

randomly arranged outline boxes on a monitor The subject can select any box to open For any one search-cycle only one such target is available and the subject must search through the boxes until it is found Then a target will become available in another box and so on until there has been a target in each of the boxes Normal subjects generally learn the strategy of searching in a fixed spatial pattern, for each configuration of boxes Frontal patients use this strategy much less It should be noted that using a fixed spatial search procedure requires subjects to abstract that they will have a memory problem if they search randomly So production of an appropriate strategy has a critical prior stage

of abstracting over the situation, behaviour and thought processes That the prefrontal cortex should be involved in the determination of the procedure by which the SupervisorySystem actively modulates the activation level of schemas in contention scheduling fits well the now standard idea that dorsolateral prefrontal cortex produces top-down

modulatory control over lower level structures (e.g Dehaene & Changeux, 1997; Frith, 2000; Miller & Cohen, 2001)

One can differentiate at least three different ways, by which humans achieve a strategy in

a non-routine situation: by spontaneous strategy production, by explicit problem-solving and by having previously determined on a procedure to use if the situation arises – intention realisation I will return later to the second and third (which can be considered part of stage five on the Domino model) I will assume that the first of the three the evolutionarily primary and restrict consideration to it for the present

I will contrast strategy production with a second supervisory subprocess, which is a key part of stage six of the Domino model This is the process of detecting a discrepancy between the situation resulting from the implementation of the strategy and the

requirements of satisfying a goal If such a discrepancy is detected then this leads to a new goal being set up to ‘patch’ the ‘bug’ or abandon the goal Very frequently the environment forces us to detect an error that we have made However, spontaneous detection of non-evident errors also frequently occurs in everyday life situations (see e.g Rizzo, Ferrante & Bagnara, 1995) These authors argue for a 4-stage process of coping with errors: production of the evidence for detecting a mismatch, explicit detection of the mismatch as arising from the subject's own behaviour, error identification and error recovery As mismatches, in particular, can be complex, specific mechanisms for non-evident error detection would seem likely to be required Thus a mismatch can be based

on rather abstract properties such as the length of time an operation is taking, it can involve a comparison with the intuitive value a result should have and can even involve inferences (Rizzo et al, 1995; Burgess and Shallice 1996b)

As with the processes responsible for strategy production, this subprocess comes in an explicit and in a non-explicit form The evolutionarily primary form is spontaneous non-evident error detection In its explicit form it is checking Checking is sometimes

viewed as merely the repeating of cognitive operations that have already been carried out.However there is critical difference On the second pass the non-evident error detection system is potentiated Moreover in many cases even deliberate checking is more than

mere recapitulation It can, for instance, involve the attempt to obtain further informationcompatible or not with a putative solution (see e.g Burgess and Shallice, 1996b).

The essence of the computation in non-evident error detection is a matching operation very different from the active construction process required in strategy production The two processes collectively correspond to the ‘manipulation and monitoring’ of Petrides’s (1994) characterisation of dorsolateral prefrontal function However within the

Supervisory System model they are complimentary The former is concerned with positively modulating the contention scheduling system from above in order to achieve goals The latter is concerned with negatively interrupting the operation of the contentionscheduling system because it is failing to achieve task goals Therefore on the argument for functional specialisation given earlier, it would seem likely that the two types of process strategy production and non-evident error detection are partially

anatomically separable In the next section I consider evidence that while both involve dorsolateral prefrontal cortex, the former is relatively more lateralised to the left and the latter relatively more to the right The contrasting lateralisations are not assumed to be all-or-none Moreover in any particular case one would also expect an additional

influence orthogonal to this one of the well-known material specific lateral biases in the operation of prefrontal cortex (see Wagner)

4 The Relative Differential Lateralisation of Strategy Production and

Non-Evident Error Detection Processes

The neuropsychological investigations of prefrontal function referred to in section 2 weremainly conducted before the early 1990s when most patients did not have MRI scans Where comparisons were carried out of possible contrasting effects of prefrontal lesions, the contrast was typically whether the lesion affected the left or right prefrontal cortex This conflated the very differing effects of lateral, medial and orbital prefrontal lesions

It was not until the mid-1990s when Stuss and Alexander introduced the method of contrasting Left Lateral, Right Lateral, Superior Medial and Inferior Medial prefrontal lesions (see, e.g Stuss et al 1998), that contrasting effects of right and left lesions began

to be reliably obtained (with respect to lateral prefrontal lesions)

Thus of the studies discussed in section 2 only one showed differing effects of left and right prefrontal lesions This was the study of random number generation of Jahanshahi

et al (1998), where the effect of rTMS to induce a routine procedure (counting) was muchgreater if applied to the left lateral prefrontal cortex than to the right Moreover as far as the avoidance of counting in ones is concerned, the functional imaging study of

Jahanshahi et al (2000) gave a similar left lateralised result

However there is a simple alternative explanation for such effects Representations of number to which alternative rules or procedures can be applied are verbal and so may involve specifically left hemisphere regions for this reason To avoid simple stimulus-based explanations of this sort it is important to contrast two theoretically important processes when both use the same material

Our first contrasting set of findings with respect to lateralisation when the same material

is used occurred in a surprising area – memory for categorised word lists (Fletcher, Shallice & Dolan, 1998a; Fletcher et al, 1998b) In free recall of word lists where the words are selected from a small number of categories, it is well known that the optimal strategy is to organise the encoded words into categories (Mandler, 1967) It was also known from neuropsychological studies that the prefrontal cortex is critical for use of thisstrategy, if words are presented randomly so that the appropriate categories have to be abstracted (Incisa & Milner, 1993; Gershberg & Shimamura, 1995) In our study

(Fletcher et al, 1998a) when this process was disrupted by a demanding secondary task, there was a specific loss of activation in left dorsolateral prefrontal cortex

However, might one not just argue that this condition involved verbal material and so the lateralisation of the effect would be explained in terms of the material used? Such an explanation would not account for a contrasting lateralisation which was unexpectedly obtained in a complementary retrieval experiment (Fletcher et al, 1998b) In this

experiment encoding had taken place outside the scanner In the experimental condition subjects had to retrieve the whole word list by free recall at a 1 per 4 s rate with the next word to be recalled cued only by the instruction ‘Next’ This was compared with a condition in which subjects had to retrieve 16 words in paired associate fashion, each being elicited by its previously presented individual subcategory label, e.g poet->

Browning Here when the activations involved in carrying out the different tasks were compared the effects obtained were in the right frontal lobe In particular, retrieval of the

organised list activated right dorsolateral cortex significantly more than did retrieval of anequal number of words through individual paired associate recall

Why might this have occurred? In a neuropsychological investigation of a closely relatedexperimental procedure of free recall of 16-word categorised lists, Stuss et al (1994) had found that patients with right frontal lesions, but not left, produce over twice as many words in their output protocols which are repeats as did normal subjects As their

memory performance was in other respects quite good and considerably better than that

of patients with left frontal lesions, it appeared that it was not that the right frontal

patients had forgotten but that instead that they had an editing or checking impairment and did not remove from their output prior to producing them, words which they had already recalled Thus the two complimentary PET experiments on encoding and retrieval

of categorical lists fit with the idea that there are relative differences in lateralisation between strategy generation and non-evident error detection

I will assume from the results just considered and the findings of Jahanshahi et al that at least for verbal material spontaneous strategy production depends critically on left

dorsolateral prefrontal cortex I will therefore concentrate on the complementary evident error detection and checking processes which are less easy to investigate

non-experimentally As far as these processes are concerned it will be assumed that they will (i) occur under conditions of uncertainty, (ii) occur when a plausible alternative needs to

be rejected, (iii) occur in time after the solution has been initially achieved In addition damage to such processes will be assumed to (iv) lead to failures in criterion setting with

subjects being more prone to false positive errors and (v) make subjects particularly liable to capture errors Moreover it will be assumed that (vi) the control of non-evident error detection and checking cannot simply be reduced to the deployment of cognitive effort.

The studies of Fletcher et al (1998ab) involved encoding and retrieval in episodic

memory Memory experiments might not seem the most natural vehicle to use to analysecontrol processes However they have a great advantage when compared to, say,

problem-solving, in allowing repeated trials which involve complex control processes in

a domain where the other processes required vary little across trials Although episodic memory experiments can be used to analyse the nature of control processes, they also involve many other processes which are not well understood I will therefore consider memory experiments only from the perspective of the properties discussed earlier in this section that non-evident error detection and checking processes should have

I will consider four experiments which relate to the relative uncertainty of the subject in the memory retrieval situation – two of which used the remember/know paradigm

(Henson et al, 1999a; Eldridge et al, 2000), one which used confidence judgements (Henson et al, 2000) and a fourth which examined incorrect versus correct source

judgements given following prior correct item judgements (Cansino et al, 2002) If one contrasts the less confident (or accurate) compared with the more confident (or accurate) responses, i.e the condition which should give rise to the greater checking activity, there

cases specifically right (Henson et al, 1999a; Eldridge et al, 2000) By contrast the more confident condition produced greater activation in mid-dorsolateral or anterior prefrontal sites on the left This result is not predicted by the current model, but nor is it in conflict with it.

The second type of situation in which error detection and checking is to be expected is when a highly plausible alternative must be rejected; a capture error is to be avoided In the memory context, proactive interference is provided by the existence of a retrieval environment where the potential for capture errors is considerable In Henson et al (2002) the one cortical region which showed significantly greater activation in trials where proactive interference was present compared to control trials was the right

dorsolateral prefrontal cortex

A third criterion concerns the time at which an error detection or checking process shouldoccur fMRI lacks the temporal discriminatory power to give adequate timing of the activation of dorsolateral prefrontal cortex in memory retrieval studies However a variety of electrophysiological studies have reported a wave beginning more than 1 second after stimulus presentation occurring at electrodes over right dorsolateral

prefrontal cortex (e.g Wilding & Rugg, 1996) To my knowledge no study using source analysis with this paradigm has been run However, if the identification of where the source is by using the localisation of the surface electrodes proves correct then this provides support that the right dorsolateral prefrontal response fits with criterion (iii) for

checking In general across studies this wave occurs bilaterally but is greater over the right than left prefrontal cortex.

Criterion (iv) holds that an impairment in checking will lead to an inadequate setting of criteria Evidence for this comes from a striking study of the effects of the creation of temporary lesions by rTMS on encoding or retrieval Rossi et al (2001) presented

subjects with 96 magazine pictures for 2 seconds each at a roughly 1 per 20 sec rate An hour later they were presented with 48 of these together with 48 distractors at the same rate rTMS or sham TMS was given over a left or right prefrontal site lasting for 500 msec to coincide with the onset of the encoding of the recognition stimulus rTMS over right prefrontal cortex at encoding or over left prefrontal at retrieval led to performance which was no different from the effects of sham TMS However rTMS over left

prefrontal cortex at encoding or over right prefrontal cortex at retrieval led to impaired performance with d’ values between 0.5 and 0.8 by comparison with d’ of about 1.4 for the two intact conditions As critically, the criterion C was significantly reduced only onecondition from the value of 0.97 found in the baseline condition When right prefrontal rTMS was given at retrieval C was only 0.28 and there was a false alarm rate of 41%, roughly double that in the sham and baseline conditions

A second study of Sandrini et al (2003) used word pairs It too showed a strong effect of the laterality of rTMS at retrieval (with an effect for right but not for left prefrontal cortex), but unlike the earlier study rTMS had an effect at encoding, when administered