Large Scale Neurocognitive NetworksUnderlying Episodic Memory

Lars Nyberg and Jonas PerssonUmeadeg University

Reza Habib Endel Tulving and Anthony R McIntoshRotman Research Institute

Roberto CabezaUniversity of Alberta

Sylvain HoulePet Centre Clarke Institute of Psychology

Abstract

amp Large-scale networks of brain regions are believed tomediate cognitive processes including episodic memoryAnalyses of regional differences in brain activity measured byfunctional neuroimaging have begun to identify putativecomponents of these networks To more fully characterizeneurocognitive networks however it is necessary to useanalytical methods that quantify neural network interactionsHere we used positron emission tomography (PET) tomeasure brain activity during initial encoding and subsequentrecognition of sentences and pictures For each type ofmaterial three recognition conditions were included whichvaried with respect to target density (0 50 100) Analysisof large-scale activity patterns identified a collection of foci

whose activity distinguished the processing of sentences vspictures A second pattern which showed strong prefrontalcortex involvement distinguished the type of cognitive process(encoding or retrieval) For both pictures and sentences themanipulation of target density was associated with minoractivation changes Instead it was found to relate to systematicchanges of functional connections between material-specificregions and several other brain regions including medialtemporal right prefrontal and parietal regions These findingsprovide evidence for large-scale neural interactions betweenmaterial-specific and process-specific neural substrates ofepisodic encoding and retrieval amp

INTRODUCTION

Large-scale neural networks are believed to mediatecognitive processes including memory (Fuster 1997Mesulam 1990) The nature of these networks can bestudied with functional neuroimaging techniques thatallow the examination of the whole brain simulta-neously For episodic memory (Tulving 1993) it hasbeen hypothesized that left prefrontal regions work inconcert with more posterior regions during encoding toform enduring memory representations (Tulving Mar-kowitsch Craik Habib amp Houle 1996) Similarly it hasbeen proposed that right prefrontal brain regions inter-act with posterior regions when stored information isrecovered (Markowitsch 1995)

There has been little direct evidence to support theseproposed neural interactions between prefrontal andposterior regions (cf Fuster 1997) A role of prefrontalbrain regions in episodic encoding and retrieval is

suggested by several studies (for reviews see NybergCabeza amp Tulving 1996 Tulving Kapur Craik Mos-covitch amp Houle 1994a) but the neural responses thatdefine actual recovery of event information remain tobe defined Evidence for involvement of posteriorregions comes from studies showing regional changesin blood flow correlated with success in retrieval A fewstudies have found increased activation of medial tem-poral lobe (MTL) regions including the hippocampusunder conditions of higher levels of retrieval (egNyberg McIntosh Houle Nilsson amp Tulving 1996aRugg Fletcher Frith Frackowiak amp Dolan 1997Schacter Alpert Savage Rauch amp Albert 1996) MTLactivation may however reflect a conscious recollectiveprocess associated with recovery (Schacter et al 1996)Furthermore increased activity associated with higherlevels of memory performance has been observed insome neocortical brain areas including superior tem-

copy 2000 Massachusetts Institute of Technology Journal of Cognitive Neuroscience 121 pp 163ndash173

poral (Heckers Rauch Goff et al 1998 Nyberg Tul-ving Habib et al 1995 Rugg et al 1997) and medialparietal (Kapur et al 1995) regions In other studiescomparisons between conditions of higher and lowerlevels of recovery have generated no evidence ofdifferential activation in posterior brain regions (egBuckner et al 1998 Schacter Buckner Koutstaal Daleamp Rosen 1997) Thus activation analyses provide aninconclusive picture regarding recovery-related effects

Analysis of functional connectivity represents a moredirect assessment of the hypothesized neurocognitiveinteractions between prefrontal and posterior brainregions during episodic memory Functional connec-tivity refers to the correlation of activity among brainregions (Friston 1994) and can be used to exploretask-specific changes in the interactions between re-gions independently of relative changes in activity Astudy of functional connections between prefrontal andmedial temporal cortices during episodic retrieval foundthat a right posterior inferior prefrontal region and a lefthippocampal region were bound in a similar pattern offunctional connectivity only during conditions of higherbut not lower levels of retrieval (McIntosh NybergBookstein amp Tulving 1997) The connectivity patternincluded bilateral inferior temporal and retrosplenialregions These results which were not evident fromactivation analyses indicated that functional interac-tions between frontal and posterior regions is relatedto level of recovery However as level of recovery atretrieval was manipulated by having subjects encodeitems in a shallow or deep manner the results couldreflect changes related to the way the items wereinitially processed rather than recovery-related changesper se

The present study was explicitly designed to examineactivation changes and changes in functional connectivityassociated with level of recovery Regional brain activitywas monitored by positron emission tomography (PET)during encoding and retrieval (yesno recognition) ofverbal (sentences) and nonverbal (pictures) episodicinformation Three different retrieval conditions wereincluded for each kind of material with the conditionsvarying in terms of target density (0ndash50ndash100) Weexpected that increasing target density would lead toincreases in recovery of information with the under-standing that this manipulation also relates to increasingfamiliarity (Mandler 1980) It is widely assumed thatrecognition can have different bases such as familiarityor recollection and these may relate to different types ofinformation In certain paradigms (for example Yoneli-nas amp Jacoby 1995) recollection is defined as remember-ing some criterial feature whereas familiarity may rely onperceptual characteristics The present design did notallow separation between responses based on familiarityor recollection and hence any changes in neural re-sponses related to the manipulation of target density canreflect either or a mixture of these

Our strategy for data analysis was as follows First weused a partial-least-squares (PLS) analysis (McIntoshBookstein Haxby amp Grady 1996) to identify task-re-lated activation changes PLS identifies spatial patterns ofbrain activity that represent the optimal associationbetween brain images and a block of contrast vectorscoding for the experimental design Of particular inter-est here was whether significant activation changesrelating to type of cognitive process (encoding vsretrieval) and to level of recovery (lowndashmediumndashhigh)would be observed Second we used PLS to explorechanges in functional connectivity related to level ofrecovery This usage of PLS can be seen as an extensionof analyses of pairwise regional interrelations in which aregion is selected and it is analyzed how activity in thisregion correlates across subjects with activity in the restof the brain (for example Horwitz 1989) PLS can sortthe correlations into what is similar and different acrosstasks hence facilitating comparisons of correlationmaps across experimental conditions Here we wereinterested in whether correlations between activity inregions involved in encoding and retrieval of sentencesand pictures (as determined by the initial task PLSanalysis) and activity in the rest of the brain wouldchange systematically as a function of level of recovery(lowndashmediumndashhigh)

RESULTS

Behavioral Results

Increasingly more items were recovered across retrievalconditions for both sentences and pictures (Table 1)The effect of target density was significant but not theeffect of materials or the target density by materialsinteraction The similar behavioral pattern across sti-muli suggests that any differences in activity betweensentences and pictures do not result from performancedifferences

Activation Changes

The PLS analysis of task-dependent activations identifiedtwo highly significant activity patterns (permutation testplt001) PLS is a multivariate analysis of activationchanges that serves to identify distributed systems thatas a whole relate to some aspect of the experimentaldesign These systems can nevertheless be characterizedby their most salient components (peak voxels) asassessed by reliability assessment (see Grady McIntoshRajah amp Craik 1998) Table 2 lists coordinates for thepeak voxels of the two patterns A more completeappreciation of the spatial distribution of the differentneural systems can be obtained from Figure 1

The first pattern related to processing of pictures vssentences independent of the type of cognitive process(Figure 1A) Compared with sentence processing pic-ture processing resulted in increased activity in posterior

164 Journal of Cognitive Neuroscience Volume 12 Number 1

visual and medial temporal regions especially in theright hemisphere and sentence processing stronglyactivated left temporal and frontal regions relative topicture processing The second pattern distinguishedthe type of cognitive process (encoding vs retrieval)across materials and level of recovery (Figure 1B)Compared to retrieval encoding was most stronglyassociated with increased activity in bilateral temporalregions and also in a left dorsolateral prefrontal regionRetrieval compared to encoding was most stronglyassociated with increased activity in left lateral parietalcortex and in right anterior prefrontal cortex

The PLS analysis of activation changes showed nosignificant effects relating to the manipulation of targetdensity Similarly in univariate statistical analyses (Fris-ton et al 1995) which explicitly tested for activationchanges related to the manipulation of target density noeffect was significant after correction for multiple com-parisons The response in some regions was howeversignificant at an uncorrected plt001 These regions arelisted in Table 3 For both pictures and sentences(lsquolsquooverallrsquorsquo) the response in several prefrontal regionstended to be related to target density Importantly thelocation of these prefrontal regions did not overlap withthe location of the right anterior prefrontal region whichwas associated with retrieval independently of level ofrecovery (see Table 2 and Figure 1B) For pictures atendency to a recovery-related effect in posterior regionswas seen in medial occipito-parietal cortex whereas the

only corresponding effect for sentences was observed inthe right insular cortex

Changes in Functional Connectivity

We next used PLS to identify systematic changes infunctional connections as a direct test of whether neuralnetwork interactions change in relation to the ma-nipulation of target densitylevel of recovery (McIntoshet al 1997) The material-specific regions identified bythe first pattern from the task PLS analysis were used asstarting points (seed voxels) As determined by the taskPLS analysis these regions were activated during encod-ing as well as retrieval If recovery of information en-gages brain regions challenged during acquisition of thesame information (eg Squire Knowlton amp Musen1993) we reasoned that these material-specific regionswould be good candidates to show recovery-relatedeffects The specific goal of the analyses was to seewhether the image-wide functional connections of theseed voxels changed in a systematic way as a function oftarget density That is we examined whether the corre-lations between activity in the seed regions and activityin any other brain regions changed such that acrossconditions they became increasingly more positive orincreasingly more negative For example we expected tofind a data pattern showing that the correlation betweenactivity in a seed voxel and activity in a set of distributedbrain regions would be negative in the 0-target condi-

Table 1 Mean Number of Yes-Responses (SDrsquos) as a Function of Condition and Material

Condition Sentences Pictures

Low (0 targets) 127 (168) 136 (136)

Medium (10 targets) 864 (136) 945 (288)

High (20 targets) 1564 (277) 1600 (257)

The scores in the low condition represent false alarms (average false alarm rate=7) The false alarm rate in the medium condition averaged 9 AnANOVA on number of yes-responses was used to analyze the effect of target density The results showed a significant effect of condition[F(220)=2088 plt001] The effect of material and the conditionpoundmaterial interaction were non-significant (pgt20)

Table 2 Material-Specific and Process-Specific Brain Regions

Effect Region (x y z)

Picture processinggtSentence processing L occipital (ndash 4 ndash 98 ndash 8)R occipital (6 ndash 86 ndash 12)R medial temporal (28 ndash 36 ndash 20)

Sentence processinggtPicture processing L temporal (ndash 56 ndash 38 0 ndash 52 ndash 6 ndash 8 ndash 44 0 ndash 12)Left frontal (ndash 52 12 16 ndash 42 26 4 ndash 44 ndash 8 40)

EncodinggtRetrieval L temporal (ndash 56 ndash 26 ndash 28)R temporal (50 ndash 60 4 36 ndash 20 ndash 24)L frontal (ndash 20 26 48)

RetrievalgtEncoding R frontal (36 46 12 20 46 0)L parietal (ndash 40 ndash 38 36)

Nyberg et al 165

tion weakly to moderately positive in the 50-targetcondition and strongly positive in the 100-targetcondition We also expected to observe the reversedpatternmdashmore negative correlations as a function ofcondition These two sets of outcomes will be consid-ered equivalentmdashtaken as evidence for recovery-relatedchanges in functional connectionsmdashno attempt will bemade to interpret the sign of correlations

The seed voxel PLS analysis identified a significanteffect ( p=0014) relating to target density duringpicture recovery when activity in a left occipital region(xyz=ndash 4 ndash 98 ndash 8) was used as seed (Figure 2A) Asrevealed by the task PLS analysis this occipital regionwas significantly more activated during picture than

sentence processing The peak salient voxels from theimages in Figure 2A are listed in Table 4 and as anaid for interpretation the univariate correlations ofthese peak voxels are given as well (For all listedpeaks there was an orderly change in correlationsfrom lowndashmediumndashhigh) Increasingly more positivecorrelations across scans were observed between theoccipital seed voxel and regions in bilateral frontalcortex and in left cuneusprecuneus (displayed inwhite color) Increasingly more negative correlationswere observed between the seed voxel and regions inbilateral medial temporal cortex bilateral occipitalcortex and in left middle temporal cortex (displayedin black)

Figure 1 (A) Mean brain scores and associated singular image from PLS for the first pattern that distinguished between sentence and pictureprocessing Regions in blue were relatively more active during picture processing regions in yellow were relatively more active during sentenceprocessing (B) Mean brain scores and associated singular image from PLS for the second pattern that distinguished between encoding and retrievalRegions in blue were more active during encoding and regions in yellow were more active during retrievalNote Areas are displayed on a standard magnetic resonance image from ndash 28 mm to +48 mm relative to the anteriorndashposterior commissure (ACndashPC) line (in 4-mm increments) The right side of the image corresponds to the right side of the brain A brain score is the sum of the cross-productof the raw voxel value and the salience for that voxel Salience is a weight associated with each voxel which express its relation to a specific contrast(for example encoding vs retrieval) The brain scores are analogous to factor scores derived from factor analysis

166 Journal of Cognitive Neuroscience Volume 12 Number 1

For sentences a corresponding effect ( p=014) wasobserved when a voxel in the left-frontal cortex(xyz=ndash 52 12 16) was used as seed (Figure 2B) Thisregion showed increased activity during sentence pro-cessing compared to picture processing (eg Figure1A) The peak voxels from the images in Figure 2B arelisted in Table 4 along with the univariate correlations ofeach peak (All listed peaks showed an orderly change incorrelations from lowndashmediumndashhigh) Increasingly morepositive correlations across scans were found betweenthe seed voxel and regions in bilateral cuneus bilateralmiddle temporal gyrus and right frontal cortex (dis-played in white color) Increasingly more negative cor-relations were observed between the seed voxel andregions in left medial-temporal cortex (displayed inblack)

Despite the fact that the location of the seed voxelsdiffered markedly (left occipital cortex for pictures leftfrontal cortex for sentences) inspection of the images inFigure 2A and b suggested considerable overlap in thepatterns of functional connections To more formallyassess overlap the reliable peak voxels from each imagewere cross-multiplied (see Methods) The crossproductis indicated in red in Figure 2A and B and the peaksalong with corresponding univariate correlations arepresented in Table 4 Salient regions of both spatialpatterns were located in cuneusprecuneus thalamusbilateral temporal cortex anterior right prefrontal cor-tex and in the anterior cingulate As detailed abovesalient regions were also identified near left hippocam-pus in both analyses but the locations did not overlap

Similar but less strong trends to recovery-relatedchanges in functional connectivity were observed whenother regions from the first pattern of the task PLSanalysis were used as seeds such as the right medial-temporal region (xyz=28 ndash 36 ndash 20) for pictures andthe left lateral-temporal region (xyz=ndash 56 ndash 38 0) forsentences By contrast when peak voxels from thesecond pattern of the task PLS were used as seeds noneof these showed a reliable change in functional connec-tivity that related to target density Collectively theseobservations indicate that recovery-related changes in

functional connectivity involve material-specific regionsactivated during both encoding and retrieval rather thanmaterial-general regions activated during encoding orretrieval

DISCUSSION

The present results provide clear evidence for large-scalenetworks related to episodic recognition memory Webegin by discussing the results which can be character-ized as lsquolsquorecovery-independentrsquorsquo and then turn to dis-cuss brain responses which seem sensitive to level ofrecovery

Recovery-Independent Effects

Systems-level patterns of co-activation were observedthat related to the two dominant dimensions of thisexperiment One represented activity that discrimi-nated between the type of material processed (picturesvs sentences) and was comprised of occipito-temporalcortices (more active for pictures) and left prefrontaland temporal cortices (more active for sentences)(eg Grady et al 1998) The second pattern wasassociated with the cognitive process (encoding vsretrieval) and consisted of bilateral temporal and leftdorsolateral prefrontal cortices during encoding andright prefrontal and left parietal cortices that weremore active during retrieval There were no large-scaleactivity patterns to suggest that there was an interac-tion of material-type and cognitive process whichindicates that encoding and retrieval were superim-posed on material-specific activity This is in line withprevious findings that cognitive operations can operatewithin the same material-specific networks (KohlerMoscovitch Winocur Houle amp McIntosh 1998a)Thus although we do not refute the possibility oftop-down modulation of material-specific neural re-sponses our data show that processing of pictures(or sentences) tends to activate the same neuralsystem regardless of whether subjects perform encod-ing or retrieval operations

Table 3 Brain Regions Showing Increased Activity as a Function of Target Density

Material Region

Overall (sentences amp pictures) L frontal (ndash 42 46 20 ndash 20 64 8)R frontal (30 28 ndash 8)R insula (32 6 12)

Pictures L precuneus (ndash 16 ndash 74 36)L posterior cingulate (ndash 24 ndash 62 8)

Sentences L frontal (ndash 22 62 0)R insula (32 ndash 8 4)

Regional effects as determined by univariate analyses (Friston et al 1995) The analyses included all retrieval scans with a contrast of [ndash 1 0 1 ndash 1 01] testing for an overall effect and [ndash 1 0 1 0 0 0] vs [0 0 0 ndash1 0 1] testing for materialndash specific effects All listed peaks were significant at anuncorrected plt001 (Zgt309) and had a spatial extent ofgt10 voxels when plotted at a threshold of Z=258

Nyberg et al 167

tion weakly to moderately positive in the 50-targetcondition and strongly positive in the 100-targetcondition We also expected to observe the reversedpatternmdashmore negative correlations as a function ofcondition These two sets of outcomes will be consid-ered equivalentmdashtaken as evidence for recovery-relatedchanges in functional connectionsmdashno attempt will bemade to interpret the sign of correlations

The seed voxel PLS analysis identified a significanteffect ( p=0014) relating to target density duringpicture recovery when activity in a left occipital region(xyz=ndash 4 ndash 98 ndash 8) was used as seed (Figure 2A) Asrevealed by the task PLS analysis this occipital regionwas significantly more activated during picture than

sentence processing The peak salient voxels from theimages in Figure 2A are listed in Table 4 and as anaid for interpretation the univariate correlations ofthese peak voxels are given as well (For all listedpeaks there was an orderly change in correlationsfrom lowndashmediumndashhigh) Increasingly more positivecorrelations across scans were observed between theoccipital seed voxel and regions in bilateral frontalcortex and in left cuneusprecuneus (displayed inwhite color) Increasingly more negative correlationswere observed between the seed voxel and regions inbilateral medial temporal cortex bilateral occipitalcortex and in left middle temporal cortex (displayedin black)

Figure 1 (A) Mean brain scores and associated singular image from PLS for the first pattern that distinguished between sentence and pictureprocessing Regions in blue were relatively more active during picture processing regions in yellow were relatively more active during sentenceprocessing (B) Mean brain scores and associated singular image from PLS for the second pattern that distinguished between encoding and retrievalRegions in blue were more active during encoding and regions in yellow were more active during retrievalNote Areas are displayed on a standard magnetic resonance image from ndash 28 mm to +48 mm relative to the anteriorndashposterior commissure (ACndashPC) line (in 4-mm increments) The right side of the image corresponds to the right side of the brain A brain score is the sum of the cross-productof the raw voxel value and the salience for that voxel Salience is a weight associated with each voxel which express its relation to a specific contrast(for example encoding vs retrieval) The brain scores are analogous to factor scores derived from factor analysis

166 Journal of Cognitive Neuroscience Volume 12 Number 1

For sentences a corresponding effect ( p=014) wasobserved when a voxel in the left-frontal cortex(xyz=ndash 52 12 16) was used as seed (Figure 2B) Thisregion showed increased activity during sentence pro-cessing compared to picture processing (eg Figure1A) The peak voxels from the images in Figure 2B arelisted in Table 4 along with the univariate correlations ofeach peak (All listed peaks showed an orderly change incorrelations from lowndashmediumndashhigh) Increasingly morepositive correlations across scans were found betweenthe seed voxel and regions in bilateral cuneus bilateralmiddle temporal gyrus and right frontal cortex (dis-played in white color) Increasingly more negative cor-relations were observed between the seed voxel andregions in left medial-temporal cortex (displayed inblack)

Despite the fact that the location of the seed voxelsdiffered markedly (left occipital cortex for pictures leftfrontal cortex for sentences) inspection of the images inFigure 2A and b suggested considerable overlap in thepatterns of functional connections To more formallyassess overlap the reliable peak voxels from each imagewere cross-multiplied (see Methods) The crossproductis indicated in red in Figure 2A and B and the peaksalong with corresponding univariate correlations arepresented in Table 4 Salient regions of both spatialpatterns were located in cuneusprecuneus thalamusbilateral temporal cortex anterior right prefrontal cor-tex and in the anterior cingulate As detailed abovesalient regions were also identified near left hippocam-pus in both analyses but the locations did not overlap

Similar but less strong trends to recovery-relatedchanges in functional connectivity were observed whenother regions from the first pattern of the task PLSanalysis were used as seeds such as the right medial-temporal region (xyz=28 ndash 36 ndash 20) for pictures andthe left lateral-temporal region (xyz=ndash 56 ndash 38 0) forsentences By contrast when peak voxels from thesecond pattern of the task PLS were used as seeds noneof these showed a reliable change in functional connec-tivity that related to target density Collectively theseobservations indicate that recovery-related changes in

functional connectivity involve material-specific regionsactivated during both encoding and retrieval rather thanmaterial-general regions activated during encoding orretrieval

DISCUSSION

The present results provide clear evidence for large-scalenetworks related to episodic recognition memory Webegin by discussing the results which can be character-ized as lsquolsquorecovery-independentrsquorsquo and then turn to dis-cuss brain responses which seem sensitive to level ofrecovery

Recovery-Independent Effects

Systems-level patterns of co-activation were observedthat related to the two dominant dimensions of thisexperiment One represented activity that discrimi-nated between the type of material processed (picturesvs sentences) and was comprised of occipito-temporalcortices (more active for pictures) and left prefrontaland temporal cortices (more active for sentences)(eg Grady et al 1998) The second pattern wasassociated with the cognitive process (encoding vsretrieval) and consisted of bilateral temporal and leftdorsolateral prefrontal cortices during encoding andright prefrontal and left parietal cortices that weremore active during retrieval There were no large-scaleactivity patterns to suggest that there was an interac-tion of material-type and cognitive process whichindicates that encoding and retrieval were superim-posed on material-specific activity This is in line withprevious findings that cognitive operations can operatewithin the same material-specific networks (KohlerMoscovitch Winocur Houle amp McIntosh 1998a)Thus although we do not refute the possibility oftop-down modulation of material-specific neural re-sponses our data show that processing of pictures(or sentences) tends to activate the same neuralsystem regardless of whether subjects perform encod-ing or retrieval operations

Table 3 Brain Regions Showing Increased Activity as a Function of Target Density

Material Region

Overall (sentences amp pictures) L frontal (ndash 42 46 20 ndash 20 64 8)R frontal (30 28 ndash 8)R insula (32 6 12)

Pictures L precuneus (ndash 16 ndash 74 36)L posterior cingulate (ndash 24 ndash 62 8)

Sentences L frontal (ndash 22 62 0)R insula (32 ndash 8 4)

Regional effects as determined by univariate analyses (Friston et al 1995) The analyses included all retrieval scans with a contrast of [ndash 1 0 1 ndash 1 01] testing for an overall effect and [ndash 1 0 1 0 0 0] vs [0 0 0 ndash1 0 1] testing for materialndash specific effects All listed peaks were significant at anuncorrected plt001 (Zgt309) and had a spatial extent ofgt10 voxels when plotted at a threshold of Z=258

Nyberg et al 167

For sentences a corresponding effect ( p=014) wasobserved when a voxel in the left-frontal cortex(xyz=ndash 52 12 16) was used as seed (Figure 2B) Thisregion showed increased activity during sentence pro-cessing compared to picture processing (eg Figure1A) The peak voxels from the images in Figure 2B arelisted in Table 4 along with the univariate correlations ofeach peak (All listed peaks showed an orderly change incorrelations from lowndashmediumndashhigh) Increasingly morepositive correlations across scans were found betweenthe seed voxel and regions in bilateral cuneus bilateralmiddle temporal gyrus and right frontal cortex (dis-played in white color) Increasingly more negative cor-relations were observed between the seed voxel andregions in left medial-temporal cortex (displayed inblack)

Despite the fact that the location of the seed voxelsdiffered markedly (left occipital cortex for pictures leftfrontal cortex for sentences) inspection of the images inFigure 2A and b suggested considerable overlap in thepatterns of functional connections To more formallyassess overlap the reliable peak voxels from each imagewere cross-multiplied (see Methods) The crossproductis indicated in red in Figure 2A and B and the peaksalong with corresponding univariate correlations arepresented in Table 4 Salient regions of both spatialpatterns were located in cuneusprecuneus thalamusbilateral temporal cortex anterior right prefrontal cor-tex and in the anterior cingulate As detailed abovesalient regions were also identified near left hippocam-pus in both analyses but the locations did not overlap

Similar but less strong trends to recovery-relatedchanges in functional connectivity were observed whenother regions from the first pattern of the task PLSanalysis were used as seeds such as the right medial-temporal region (xyz=28 ndash 36 ndash 20) for pictures andthe left lateral-temporal region (xyz=ndash 56 ndash 38 0) forsentences By contrast when peak voxels from thesecond pattern of the task PLS were used as seeds noneof these showed a reliable change in functional connec-tivity that related to target density Collectively theseobservations indicate that recovery-related changes in

functional connectivity involve material-specific regionsactivated during both encoding and retrieval rather thanmaterial-general regions activated during encoding orretrieval

DISCUSSION

The present results provide clear evidence for large-scalenetworks related to episodic recognition memory Webegin by discussing the results which can be character-ized as lsquolsquorecovery-independentrsquorsquo and then turn to dis-cuss brain responses which seem sensitive to level ofrecovery

Recovery-Independent Effects

Systems-level patterns of co-activation were observedthat related to the two dominant dimensions of thisexperiment One represented activity that discrimi-nated between the type of material processed (picturesvs sentences) and was comprised of occipito-temporalcortices (more active for pictures) and left prefrontaland temporal cortices (more active for sentences)(eg Grady et al 1998) The second pattern wasassociated with the cognitive process (encoding vsretrieval) and consisted of bilateral temporal and leftdorsolateral prefrontal cortices during encoding andright prefrontal and left parietal cortices that weremore active during retrieval There were no large-scaleactivity patterns to suggest that there was an interac-tion of material-type and cognitive process whichindicates that encoding and retrieval were superim-posed on material-specific activity This is in line withprevious findings that cognitive operations can operatewithin the same material-specific networks (KohlerMoscovitch Winocur Houle amp McIntosh 1998a)Thus although we do not refute the possibility oftop-down modulation of material-specific neural re-sponses our data show that processing of pictures(or sentences) tends to activate the same neuralsystem regardless of whether subjects perform encod-ing or retrieval operations

Table 3 Brain Regions Showing Increased Activity as a Function of Target Density

Material Region

Overall (sentences amp pictures) L frontal (ndash 42 46 20 ndash 20 64 8)R frontal (30 28 ndash 8)R insula (32 6 12)

Pictures L precuneus (ndash 16 ndash 74 36)L posterior cingulate (ndash 24 ndash 62 8)

Sentences L frontal (ndash 22 62 0)R insula (32 ndash 8 4)

Regional effects as determined by univariate analyses (Friston et al 1995) The analyses included all retrieval scans with a contrast of [ndash 1 0 1 ndash 1 01] testing for an overall effect and [ndash 1 0 1 0 0 0] vs [0 0 0 ndash1 0 1] testing for materialndash specific effects All listed peaks were significant at anuncorrected plt001 (Zgt309) and had a spatial extent ofgt10 voxels when plotted at a threshold of Z=258

Nyberg et al 167

The finding that brain regions generally activatedduring encoding included left inferior temporal anddorsolateral prefrontal areas and brain regions generallyactivated during retrieval involved right anterior pre-frontal and left parietal areas is consistent with muchprior research (see Cabeza and Nyberg 1997) Several ofthese regions have been suggested to form generalencoding and retrieval networks regardless of type ofevent information (Nyberg et al 1996b) The asym-metric involvement of left and right prefrontal areasduring encoding and retrieval provides support for theHERA model (Nyberg et al 1996 Tulving et al 1994a)This model holds that the left prefrontal cortex isdifferentially more involved in episodic encoding than

is the right prefrontal cortex whereas the right prefron-tal cortex is differentially more involved in episodicretrieval than is the left prefrontal cortex The activationof the right anterior prefrontal cortex during retrieval ishardly controversial as activation of this region duringepisodic retrieval has been demonstrated for severaldifferent tasks and kinds of material It is more note-worthy that the left prefrontal cortex was differentiallyactivated during encoding for both verbal and nonverbalmaterial This is because although left prefrontal activa-tion has been observed in several other studies ofnonverbal episodic encoding (see Nyberg et al 1998for a recent summary) it has been argued that the partof the HERA model that predicts left-lateralized prefron-

Figure 2 (A) Results from the analysis of seed voxel correlations for left occipital area 18 The plot of correlation of brain scores with the occipitalvoxel by condition shows a roughly linear change in the covariance of this voxel across picture retrieval conditions with the singular image displayedon the left Peak salient voxels with a positive loading (correlations becoming more positive across scans) are displayed in white Peak salient voxelswith a negative loading (correlations becoming more negative across scans) are shown in black Regions shown in red are those for which a similarpattern of functional connections were observed for both seed voxels (that is left area 18 for pictures and left area 44 for sentences) (B) Resultsfrom the analysis of seed voxel correlations for left frontal area 44 The plot of correlation of brain scores with the frontal voxel by condition shows aroughly linear change in the covariance of this voxel across sentence retrieval conditions with the singular image displayed on the left Peak salientvoxels with positive loading are shown in white and peak salient voxels with negative loading are shown in black Regions shown in red are thosefor which a similar pattern of functional connections were observed for both seed voxelsNote Areas are displayed on a standard magnetic resonance image from ndash 28 mm to +48 mm relative to the anteriorndashposterior commissure (ACndashPC) line (in 4-mm increments) The right side of the image corresponds to the right side of the brain

168 Journal of Cognitive Neuroscience Volume 12 Number 1

tal activation during encoding has low predictabilitywhen applied to encoding studies involving nonverbalmaterials (Kelley et al 1998a 1998b) The peak in thepresent study (located in area 8) fell outside the area ofleft prefrontal cortex (areas 45 and 47) most stronglyassociated with episodic encoding (but see for exampleHaxby et al 1996) but our results nevertheless suggestthat the predictability of the model may not be as low ashas been argued

A final point to consider in this section is the con-sistent activation of the right prefrontal cortex acrossthe different retrieval conditions As noted above acti-vation of the right anterior prefrontal cortex duringepisodic retrieval has been observed in many functionalneuroimaging studies but the interpretation of thisresponse has been the subject of some debate This isso even if the discussion is limited to episodic recogni-tion According to one position the right anteriorprefrontal activation is reflecting retrieval attemptretrie-val mode (Kapur et al 1995 Nyberg et al 1995) Thebasis for this suggestion was findings that the level ofregional activation relative to a common baseline didnot vary as a function of target density or level ofrecognition success (manipulated by varying the levelof processing at study) A conflicting position holds thatthe prefrontal cortex mediates post-retrieval processing(Rugg Fletcher Frith Frackowiak amp Dolan 1996) This

position is based on findings of greater prefrontalactivation in retrieval conditions involving many targetscompared to retrieval conditions involving fewer tar-gets (see also Tulving et al 1994b) A recent study byWagner Desmond Glover amp Gabrieli (1998) providedsupport that right prefrontal regions mediate processesassociated with retrieval attempt rather than retrievalsuccess Wagner et al showed that under standardrecognition instructions level of right prefrontal activa-tion is not affected by target density They also showedthat if subjects are informed about the relative propor-tion of targets and lures conditions of high targetdensity tend to be associated with increased rightprefrontal activation Possibly this is because beinginformed that most items will be new reduces theinvolvement of retrieval attempt processes hence low-ering the level of right prefrontal activation Thisobservation helps to explain some findings of greaterright prefrontal activation when recognition of old andnew items were directly contrasted (Tulving et al1994bmdashsubjects were informed) and to the extentthat subjects on-line can discoverguess that mostitems are nonstudied it may be possible to accountfor similar findings when subjects were not informedabout the proportion targets and lures (Rugg et al1996) By showing that the right anterior prefrontalresponse did not vary as a function of level of recovery

Table 4 Peak Regions of Functional Connectivity Patterns with Associated Univariate Seed-Correlations

Material Region (x y z)Correlation

( lowndashmediumndashhigh)

Pictures (seed=ndash 4 ndash 98 ndash 8) L frontal (ndash 40 42 0) ndash 048 013 046L frontal (ndash 44 16 ndash 4) ndash 047 0 071L precuneus (ndash 10 ndash 66 20) ndash 048 006 085L cuneus (ndash 24 ndash 80 32) ndash 047 003 068R frontal (24 50 32) ndash 074 ndash 019 050R frontal (40 28 ndash 8) ndash 079 ndash 010 045R occipital (32 ndash 96 0) 029 ndash 075 ndash 076L occipital (ndash 36 ndash 82 ndash 8) 070 024 ndash 052L temporal (ndash 44 ndash 36 4) 028 ndash 027 ndash 086R hippocampus (22 ndash 14 ndash 16) 004 ndash 002 ndash 085L hippocampus (ndash 24 ndash 14 ndash 16) 009 ndash 001 ndash 069

Sentences (seed=ndash 52 12 16) L temporal (ndash 56 ndash 50 12) ndash 011 029 073R temporal (52 ndash 60 24) ndash 042 004 052R frontal (46 16 28) ndash 035 025 064R cuneus (2 ndash 66 8) ndash 053 ndash 015 013L cuneus (ndash 2 ndash 92 20) ndash 073 ndash 035 ndash 020L hippocampus (ndash 32 ndash 22 ndash 12) 029 ndash 010 ndash 062

Overlap Pictures SentencesAnterior cingulate (2 38 28) ndash 040 ndash 039 058 013 055 080R frontal (26 58 24) ndash 053 ndash 022 055 ndash 013 013 059L cuneusprecuneus (ndash 22 ndash 88 28) ndash 035 0 061 ndash 032 015 028R thalamus (12 ndash 22 4) 065 ndash 012 ndash 048 050 ndash 024 ndash 048L temporal (ndash 42 ndash 26 0) 034 ndash 051 ndash 061 039 ndash 009 ndash 014R temporal (40 ndash 10 ndash 20) 035 0 ndash 060 022 ndash 049 ndash 060

Nyberg et al 169

(target density) the present study is in agreement withthe conclusion by Wagner et al and more generallywith the retrieval attemptmode account Importantlythough the analyses of activation changes and con-nectivity changes pointed to an association betweenprefrontal involvement and level of recovery Takentogether these observations hint that the retrievalattemptmode account and the post-retrieval (success)account are not mutually exclusive but distinct pre-frontal regions may mediate each of these processesBelow we discuss recovery-sensitive responses in moredetail

Recovery-Dependent Effects

The PLS analysis of task-dependent activations suggestedthat recovery-related changes in regional activity wereweak (no significant activity pattern relating to level ofrecovery was identified) This suggestion was supportedby the outcome of univariate analyses which explicitlytested for recovery-related effects (material-general aswell as material-specific effects) None of the identifiedeffects was significant following correction for multiplecomparisons which appears consistent with some pre-vious findings (Buckner et al 1998 Schacter et al 1997Wagner et al 1998) However as some of the regionaleffects which were significant at an uncorrected plt001were located in brain areas previously associated withlevel of recoverytarget density (prefrontal cortex andprecuneus) we will provide some discussion of theseeffects

Starting with prefrontal cortex activity in two leftprefrontal areas showed a tendency to recovery-relatedeffects The peaks fell in the middle frontal gyrus in ornear area 10 Similarly one of the areas identified byRugg et al (1996) to be sensitive to target density waslocated in left area 10 We also found that activity in aright inferior prefrontal voxel showed a recovery-relatedtrend In a previous study (Nyberg et al 1996a) it wasobserved that activity in this area tended to be greaterfor subjects who correctly recognized many items thanfor subjects who recognized few items Collectivelythese studies suggest that activity in specific prefrontalregions is modulated by level of recovery

Turning now to the precuneus region as was noted inthe Introduction activity in this area has been related tolevel of target density in previous studies (Kapur et al1995) Kapur et al argued that their observation ofincreased activity in posterior cortical regions includingthe precuneus during recognition of studied wordscould reflect reactivation of stored engrams SimilarlyRoland and Gulyas (1995) presented evidence which ledthem to suggest that the precuneus area is a storage sitefor visual patterns and data obtained by Krause et al(1999) led these authors to conclude that successfulretrieval is dependent on reactivation of engrams inposterior multimodal association cortices especially

the precuneus These previous proposals are consistentwith the present demonstration of a recovery-relatedeffect in precuneus It is unclear however why theeffect was only observed for picturesmdashwords were usedin both the Kapur et al and Krause et al studies It maybe that episodic recognition of the kind of verbalmaterial used in this study differs in significant waysfrom episodic word recognition (and from recognitionof nonverbal information) Nonetheless our findingsfrom the picture conditions provide further evidencethat precuneus is a posterior region that shows recov-ery-related effects

In addition to testing for activation changes weexplored whether neural network interactions changedin relation to the manipulation of target densitylevel ofrecovery We found that the success of the recovery ofinformation in terms of the number of items correctlyrecognized can be directly related to the interactionsamong brain regions That is using regions showingdifferential activity related to the type of material weidentified patterns of systems-level covariation acrossretrieval conditions that changed as the number of itemsrecognized increased For sentences activity in a leftfrontal region showed a change in correlation patternthat mapped on to the increase in old items acrossrecognition scans For pictures a corresponding effectwas found for a region in left occipital cortex Thusthrough the examination of functional connectivity weprovide empirical evidence that the level of recovery canbe directly related to the interactions among brainregions Such an observation is anticipated by theoriessuggesting that cognition is supported by the operationsof large-scale neural systems (Mesulam 1990)

Interestingly even though different regions were usedas lsquolsquoseedsrsquorsquo for sentences and for pictures there wassignificant overlap in the areas identified in the systems-level patterns (indicated in red in Figure 2) Overlappingregions included the anterior cingulate gyrus thalamusright prefrontal cortex and cuneusprecuneus More-over for both seeds hippocampal regions were impli-cated although the specific location of these differed forsentences and pictures The overlapping areas may bepart of a neurocognitive system that interacts withmaterial-specific regions in binding together differentcomponents of the episodic experience (John Easton ampIsenhart 1997 McIntosh et al 1997) In light of theactivation data discussed above it is striking that theoverlapping areas included regions in right prefrontalcortex and in precuneus This is so even if the peaksidentified in the two sets of analyses did not closelyoverlap and taken together the findings strongly im-plicate areas in prefrontal cortex and precuneus inaspects of recovery It is also noteworthy that theconnectivity analyses pointed to a role of hippocampalregions in recovery despite the fact that regional activitydid not differ as function of level of recovery Differentialactivation of hippocampal regions during episodic re-

170 Journal of Cognitive Neuroscience Volume 12 Number 1

trieval has been related to conscious recollection andorconfidence (see Nyberg et al 1996a Schacter et al1996) Possibly the observation here of lack of differ-ential activation of hippocampal regions across retrievalconditions signals equal levels of confidence in theresponsesmdashregardless of whether these involve correctrecognition or correct rejection The finding in theconnectivity analyses that involvement of hippocampalregions nevertheless relates to level of recovery is inagreement with a recent demonstration that structuresin the medial temporal lobe interact with specific poster-ior neocortical brain regions depending on the type ofinformation retrieved (Kohler McIntosh Moscovitch ampWinocur 1998b) Finally in addition to the overlappingregions the patterns for sentences and pictures in-cluded unique regions Notably for sentences bilateralposterior temporal regions were involved whereas forpictures bilateral lateral occipital regions were impli-cated The interactions among material-specific regionsmay be related to recovery of modality-specific aspectsof an experience

The discussion of the functional connectivity resultshas focused on the findings when a left prefrontal seedwas used for sentences and a left occipital seed was usedfor pictures As noted in Results similar trends wereseen when other material-specific regions from the firstpattern of the task PLS analysis were used as seeds (suchas a right medial-temporal region for pictures and a leftmiddle temporal region for sentences) Together withthe observation that recovery-related changes in func-tional connections were not seen when seeds from thesecond pattern of the task PLS analysis (general encod-ing and retrieval regions) were used this indicates thatrecovery-related effects were restricted to material-spe-cific regions commonly involved in encoding and retrie-val While this appears consistent with theoreticalproposals (see for example Squire Knowlton amp Mu-sen 1993) it must be noted that recovery-related effectswere not seen for all material-specific regions Moreoverit is quite possible that recovery-related effects wouldhave been found for other regions This relates to ageneral problem with connectivity analysesmdashselectionof regions Therefore the present results may best beseen as an empirical demonstration that level of episodicrecovery maps on to changes in functional connectivityrather than as a complete characterization of the in-volved networks As such our findings extend recentobservations that other kinds of learning and memoryare based on changing interactions between brain re-gions (Buchel Coull amp Friston 1999 McIntosh Cabezaamp Lobaugh 1998)

CONCLUSION

In conclusion the present results show that distinctencoding and retrieval networks operate within com-mon material-specific networks In line with previous

findings actual recovery of information manipulated byvarying levels of target density tended to be associatedwith activation changes in distinct regions In additionevidence was provided that level of recovery can bedirectly related to the interactions among brain regionsNotably our demonstration that material-specific re-gions interacted with common regions during recoveryshow that they were operating within overlapping neur-al systems (McIntosh in press) Thus although thespecific events differed (verbal vs nonverbal) the inter-actions defining higher levels of recovery involved simi-lar regions The activation of material-specific regionsand their interactions with common areas can be con-sidered a large-scale neural network related to successfulretrieval of episodic memories

METHODS

Cognitive Task

Each encoding condition included 45 items Subjectswere instructed to try to learn as many items as possiblefor a later test of memory and to press with their righthand any of two mouse buttons after having viewed eachitem (the items were presented on a computer screenthat was placed above the subjectsrsquo heads) The verbalmaterials consisted of a sentence frame and a semanti-cally related word (for example hairy on the outside butdelicious on the insidemdashcoconut) The nonverbal ma-terial were scenic pictures of coastlines and wateranimals vegetation and mountains Following eachencoding condition subjects were given three memorytests in a counterbalanced order They were instructedto press with their right hand the left or right mousebutton depending on whether or not they recognized agiven item from the study list Each memory test in-cluded four studied and four non-studied items beforethe scan interval and one studied and one non-studieditem after the scan interval During the scan interval onetest included 20 non-studied items (0-target condi-tion) another test included 10 studied and 10 non-studied items (50-target condition) and the third testincluded 20 studied items (100-target condition) Nostudied items were repeated in the same retrieval scanor in different retrieval scans Half of the subjects werepresented the verbal conditions prior to the pictorialconditions the other half were given the pictorial con-ditions first During encoding and retrieval the presen-tation rate was 25 sec per item (ISI=05 sec)

Image Acquisition

Subjects were given eight PET scans (one scanencoding-and retrieval condition) The scans were obtained with aGEMS-Scanditronix PC 2048-15B head scanner usingbolus injections of 40 mCi H2

15O and 60-sec dataacquisition scans The study was approved by the Hu-man Subjects Use Committee of Baycrest Center All

Nyberg et al 171

subjects gave written informed consent (four femaleseven male mean age=284 years range=20ndash39 years)One additional subject was tested but had to be ex-cluded due to software problems during scanning

Image Analysis

All images were realigned to the subjectsrsquo first scan byusing AIR (Woods Mazziotta amp Cherry 1993) trans-formed into Talairach space (Talairach amp Tournoux1988) and smoothed to 10 mm by using SPM-95 (Fristonet al 1996) Partial least squares (PLS) was used toidentify patterns of brain activity related to the differenttasks (McIntosh et al 1996) PLS was also used toanalyze task-related differences in functional connectiv-ity (McIntosh et al 1997) Peak voxels used to charac-terize the singular images from PLS analyses wereselected based on their reliability through bootstrapestimation of standard errors (Grady et al 1998) Twosets of weights are derived for the PLS analysis offunctional connectivity one for the seed voxel withineach scan and one for each voxel in the remainder of theimage (the singular image) The variation in the weightfor the seed voxel across scans indicates whether thepattern of covariances in the singular image represents atask-related difference in functional connectivity Forexample if the seed voxel weights are similar acrossscans this would represent a common pattern of func-tional connectivity If seed voxel weights follow a linearchange from 0 to 50 to 100 targets then that wouldrepresent a change in functional connectivity that mapson to the change in target density Using a set of linearcontrast coding for target density we statistically eval-uated the PLS results for such a pattern by covarying thecontrast with the seed voxel weights for each latentvariable The significance of the target-density effect wasassigned using permutation tests (McIntosh amp Gonzalez-Lima 1998) The overlapping regions in the two singularimages were identified by thresholding (at a ratio ofvoxel weight to bootstrap standard error greater than 2)the singular images for each seed analysis to includeonly the reliable peak voxels and computing the cross-product of the two images Therefore the overlappingregions are the strongest and most reliable voxels thatcontributed to the singular image in both seed PLSanalyses

Acknowledgments

This work was supported by a grant from HSFR (Sweden) to LN E T and A R M and from NSERC (Canada) to ET Wethank the staff at the PET center for help with PET scanningThe helpful comments by two anonymous reviewers aregratefully acknowledged We also thank colleagues who haveprovided useful input on this work especially Paul FletcherStefan Kohler and Karl-Magnus Pettersson

Reprint requests should be sent to Anthony R McIntoshRotman Research Institute 3560 Bathurst Street Toronto

Ont M6A 2E1 Canada Electronic mail mcintoshpsychutorontoca

REFERENCES

Buckner R L Koutstaal W Schacter D L Dale A M RotteM amp Rosen B R (1998) Functional-anatomic study ofepisodic retrieval II Selective averaging of event-relatedfMRI trials to test the retrieval success hypothesis Neuro-image 7 163ndash175

Buchel C Coull J T amp Friston K J (1999) The predictivevalue of changes in effective connectivity for human learn-ing Science 283 1538ndash1541

Cabeza R amp Nyberg L (1997) Imaging cognition An em-pirical review of PET studies with normal subjects Journalof Cognitive Neuroscience 9 1ndash26

Friston K J (1994) Functional and effective connectivity inneuroimaging A synthesis Human Brain Mapping 2 56ndash78

Friston K J Holmes A P Worsley K J Poline J-P Frith CD amp Frackowiak R S J (1995) Statistical parametric mapsin functional imaging A general linear approach HumanBrain Mapping 2 189ndash210

Friston K J Ashburner J Frith C D Pline J-B Heather JD amp Frackowiak R S J (1996) Spatial registration andnormalization of images Human Brain Mapping 2 165ndash189

Fuster J M (1997) Network memory Trends in Neuros-ciences 20 451ndash459

Grady C L McIntosh A R Rajah M N amp Craik F I M(1998) Neural correlates of the episodic encoding of pic-tures and words Proceedings of the National Academy ofSciences USA 95 2703ndash2708

Haxby J V Ungerleider L G Horwitz B Maisog J MRapoport S I amp Grady C L (1996) Face encoding andrecognition in the human brain Proceedings of the NationalAcademy of Sciences USA 93 922ndash927

Heckers S Rauch S L Goff D et al (1998) Impaired re-cruitment of the hippocampus during conscious recollectionin schizophrenia Nature Neuroscience 1 318ndash323

Horwitz B (1989) Functional neural systems analyzed by useof interregional correlations of glucose metabolism In J-PEwert amp MA Arbib (Eds) Visuomotor Coordination (873ndash892) New York Plenum Press

John E R Easton P amp Isenhart R (1997) Consciousnessand cognition may be mediated by multiple independentcoherent ensembles Consciousness and Cognition 6 3ndash39

Kapur S Craik F I M Jones C Brown G M Houle S ampTulving E (1995) Functional role of the prefrontal cortex inretrieval of memories A PET study NeuroReport 6 1880ndash1884

Kelley W M Miezin F M McDermott K B Buchias R LRaichle M E Cohen W J Olliraer J M Akbudak ELonturu T E Snyder A Z amp Petersen S E (1998a)Hemispheric specialization in human dorsal frontal cortexand medial temporal lobe for verbal and nonverbal memoryencoding Neuron 20 927ndash936

Kelley W M Buckner R L amp Petersen S E (1998b) Re-sponse from Kelley Buckner and Petersen Trends in Cog-nitive Science 2 921

Kohler S Moscovitch M Winocur G Houle S amp McIntoshA R (1998a) Networks of domain-specific and general re-gions involved in episodic memory for spatial location andobject identity Neuropsychologia 36 129ndash142

Kohler S McIntosh A R Moscovitch M amp Winocur G(1998b) Functional interactions between the medial tem-

172 Journal of Cognitive Neuroscience Volume 12 Number 1

subjects gave written informed consent (four femaleseven male mean age=284 years range=20ndash39 years)One additional subject was tested but had to be ex-cluded due to software problems during scanning

Image Analysis

All images were realigned to the subjectsrsquo first scan byusing AIR (Woods Mazziotta amp Cherry 1993) trans-formed into Talairach space (Talairach amp Tournoux1988) and smoothed to 10 mm by using SPM-95 (Fristonet al 1996) Partial least squares (PLS) was used toidentify patterns of brain activity related to the differenttasks (McIntosh et al 1996) PLS was also used toanalyze task-related differences in functional connectiv-ity (McIntosh et al 1997) Peak voxels used to charac-terize the singular images from PLS analyses wereselected based on their reliability through bootstrapestimation of standard errors (Grady et al 1998) Twosets of weights are derived for the PLS analysis offunctional connectivity one for the seed voxel withineach scan and one for each voxel in the remainder of theimage (the singular image) The variation in the weightfor the seed voxel across scans indicates whether thepattern of covariances in the singular image represents atask-related difference in functional connectivity Forexample if the seed voxel weights are similar acrossscans this would represent a common pattern of func-tional connectivity If seed voxel weights follow a linearchange from 0 to 50 to 100 targets then that wouldrepresent a change in functional connectivity that mapson to the change in target density Using a set of linearcontrast coding for target density we statistically eval-uated the PLS results for such a pattern by covarying thecontrast with the seed voxel weights for each latentvariable The significance of the target-density effect wasassigned using permutation tests (McIntosh amp Gonzalez-Lima 1998) The overlapping regions in the two singularimages were identified by thresholding (at a ratio ofvoxel weight to bootstrap standard error greater than 2)the singular images for each seed analysis to includeonly the reliable peak voxels and computing the cross-product of the two images Therefore the overlappingregions are the strongest and most reliable voxels thatcontributed to the singular image in both seed PLSanalyses

Acknowledgments

This work was supported by a grant from HSFR (Sweden) to LN E T and A R M and from NSERC (Canada) to ET Wethank the staff at the PET center for help with PET scanningThe helpful comments by two anonymous reviewers aregratefully acknowledged We also thank colleagues who haveprovided useful input on this work especially Paul FletcherStefan Kohler and Karl-Magnus Pettersson

Reprint requests should be sent to Anthony R McIntoshRotman Research Institute 3560 Bathurst Street Toronto

Ont M6A 2E1 Canada Electronic mail mcintoshpsychutorontoca

REFERENCES

Buckner R L Koutstaal W Schacter D L Dale A M RotteM amp Rosen B R (1998) Functional-anatomic study ofepisodic retrieval II Selective averaging of event-relatedfMRI trials to test the retrieval success hypothesis Neuro-image 7 163ndash175

Buchel C Coull J T amp Friston K J (1999) The predictivevalue of changes in effective connectivity for human learn-ing Science 283 1538ndash1541

Cabeza R amp Nyberg L (1997) Imaging cognition An em-pirical review of PET studies with normal subjects Journalof Cognitive Neuroscience 9 1ndash26

Friston K J (1994) Functional and effective connectivity inneuroimaging A synthesis Human Brain Mapping 2 56ndash78

Friston K J Holmes A P Worsley K J Poline J-P Frith CD amp Frackowiak R S J (1995) Statistical parametric mapsin functional imaging A general linear approach HumanBrain Mapping 2 189ndash210

Friston K J Ashburner J Frith C D Pline J-B Heather JD amp Frackowiak R S J (1996) Spatial registration andnormalization of images Human Brain Mapping 2 165ndash189

Fuster J M (1997) Network memory Trends in Neuros-ciences 20 451ndash459

Grady C L McIntosh A R Rajah M N amp Craik F I M(1998) Neural correlates of the episodic encoding of pic-tures and words Proceedings of the National Academy ofSciences USA 95 2703ndash2708

Haxby J V Ungerleider L G Horwitz B Maisog J MRapoport S I amp Grady C L (1996) Face encoding andrecognition in the human brain Proceedings of the NationalAcademy of Sciences USA 93 922ndash927

Heckers S Rauch S L Goff D et al (1998) Impaired re-cruitment of the hippocampus during conscious recollectionin schizophrenia Nature Neuroscience 1 318ndash323

Horwitz B (1989) Functional neural systems analyzed by useof interregional correlations of glucose metabolism In J-PEwert amp MA Arbib (Eds) Visuomotor Coordination (873ndash892) New York Plenum Press

John E R Easton P amp Isenhart R (1997) Consciousnessand cognition may be mediated by multiple independentcoherent ensembles Consciousness and Cognition 6 3ndash39

Kapur S Craik F I M Jones C Brown G M Houle S ampTulving E (1995) Functional role of the prefrontal cortex inretrieval of memories A PET study NeuroReport 6 1880ndash1884

Kelley W M Miezin F M McDermott K B Buchias R LRaichle M E Cohen W J Olliraer J M Akbudak ELonturu T E Snyder A Z amp Petersen S E (1998a)Hemispheric specialization in human dorsal frontal cortexand medial temporal lobe for verbal and nonverbal memoryencoding Neuron 20 927ndash936

Kelley W M Buckner R L amp Petersen S E (1998b) Re-sponse from Kelley Buckner and Petersen Trends in Cog-nitive Science 2 921

Kohler S Moscovitch M Winocur G Houle S amp McIntoshA R (1998a) Networks of domain-specific and general re-gions involved in episodic memory for spatial location andobject identity Neuropsychologia 36 129ndash142

Kohler S McIntosh A R Moscovitch M amp Winocur G(1998b) Functional interactions between the medial tem-

172 Journal of Cognitive Neuroscience Volume 12 Number 1

poral lobes and posterior neocortex related to episodicmemory retrieval Cerebral Cortex 8 451ndash461

Krause B J Schmidt D Mottaghy F M Taylor J HalsbandU Herzog H Tellmann L amp Muller-Gartner H-W (1999)Episodic retrieval activates the precuneus irrespective of theimagery content of word-pair associates A PET study Brain122 255ndash263

Mandler G (1980) Recognizing The judgement of previousoccurrence Psychological Review 87 252ndash271

Markowitsch H J (1995) Which brain regions are criticallyinvolved in the retrieval of old episodic memory Brain Re-search Reviews 21 117ndash127

McIntosh A R (in press) Mapping cognition to the brainthrough neural interactions Memory

McIntosh A R Bookstein F L Haxby J V amp Grady C L(1996) Spatial pattern analysis of functional brain imagesusing partial least squares Neuroimage 3 143ndash157

McIntosh A R Cabeza R E amp Lobaugh N J (1998) Analysisof neural interactions explains the activation of occipitalcortex by an auditory stimulus Journal of Neurophysiology80 2790

McIntosh A R amp Gonzalez-Lima F (1998) Large-scale func-tional connectivity in associative learning Interrelations ofthe rat auditory visual and limbic systems Journal of Neu-rophysiology 80 3148ndash3162

McIntosh A R Nyberg L Bookstein F L amp Tulving E(1997) Differential functional connectivity of prefrontal andmedial temporal cortices during episodic memory retrievalHuman Brain Mapping 5 323ndash327

Mesulam M M (1990) Large-scale neurocognitive networksand distributed processing for attention language andmemory Annals of Neurology 28 597ndash613

Nyberg L Tulving E Habib R et al (1995) Functional brainmaps of retrieval mode and recovery of episodic informa-tion NeuroReport 7 249ndash252

Nyberg L Cabeza R amp Tulving E (1996) PET studies ofencoding and retrieval The HERA model PsychonomicBulletin and Review 3 135ndash148

Nyberg L McIntosh A R Houle S Nilsson L-G amp TulvingE (1996a) Activation of medial-temporal structures duringepisodic memory retrieval Nature 380 715ndash717

Nyberg L McIntosh A R Cabeza R Habib R Houle S ampTulving E (1996b) General and specific brain regions in-volved in encoding and retrieval of events What where andwhen Proceedings of the National Academy of SciencesUSA 93 11280ndash11285

Nyberg L Cabeza R amp Tulving E (1998) Asymmetric frontalactivation during episodic memory What kind of specificityTrends in Cognitive Sciences 2 419ndash420

Roland P E amp Gulyas B (1995) Visual memory visual ima-gery and visual recognition of large field patterns by the

human brain Functional anatomy by positron emission to-mography Cerebral Cortex 5 79ndash93

Rugg M D Fletcher P C Frith C D Frackowiak R S J ampDolan R J (1996) Differential activation of the prefrontalcortex in successful and unsuccessful memory retrievalBrain 119 2073ndash2083

Rugg M D Fletcher P C Frith C D Frackowiak R S Jamp Dolan R J (1997) Brain regions supporting intentionaland incidental memory a PET study NeuroReport 81283ndash1287

Schacter D L Alpert N M Savage C R Rauch S L ampAlbert M S (1996) Conscious recollection and the humanhippocampal formation Evidence from positron emissiontomography Proceedings of the National Academy ofSciences USA 93 321ndash325

Schacter D L Buckner R L Koutstaal W Dale A M ampRosen B R (1997) Late onset of anterior prefrontal activityduring true and false recognition An event-related fMRIstudy Neuroimage 6 259ndash269

Squire L R Knowlton B amp Musen G (1993) The structureand organization of memory Annual Review of Psychology44 453ndash495

Talairach J amp Tournoux P (1988) Co-planar stereotaxicatlas of the human brain New York Thieme

Tulving E (1993) Elements of Episodic Memory New YorkOxford University Press

Tulving E Kapur S Craik F I M Moscovitch M amp HouleS (1994a) Hemispheric encodingretrieval asymmetry inepisodic memory positron emission tomography findingsProceedings of the National Academy of Sciences USA 912016ndash2020

Tulving E Kapur S Markowitsch H J Craik F I M HabibR amp Houle S (1994b) Neuroanatomical correlates of re-trieval in episodic memory Auditory sentence recognitionProceedings of the National Academy of Sciences USA 912012ndash2015

Tulving E Markowitsch H J Craik F I M Habib R ampHoule S (1996) Novelty and familiarity activations in PETstudies of memory encoding and retrieval Cerebral Cortex6 71ndash79

Wagner A D Desmond J E Glover G H amp Gabrieli J D E(1998) Prefrontal cortex and recognition memory Func-tional-MRI evidence for context-dependent retrievalprocesses Brain 121 1985- 2002

Woods R P Mazziotta J C amp Cherry S R (1993) MRI-PETregistration with automated algorithm Journal of Compu-ter Assisted Tomography 17 536ndash546

Yonelinas A P amp Jacoby L L (1995) The relation betweenremembering and knowing as bases for recognition Effectsof size congruency Journal of Memory and Language 34622ndash643

Nyberg et al 173