Intention, Action Planning, and Decision Making in - Andersen Lab

16 Pages·2012·1.04 MB·English

Neuron
Review
Intention, Action Planning, and Decision Making
in Parietal-Frontal Circuits
RichardA.Andersen1,*andHeCui1,2
1DivisionofBiology,CaliforniaInstituteofTechnology,Pasadena,CA91125,USA
2Presentaddress:BrainandBehaviorDiscoveryInstitute,MedicalCollegeofGeorgia,Augusta,GA30912,USA
*Correspondence:[email protected]
DOI10.1016/j.neuron.2009.08.028
Theposteriorparietalcortexandfrontalcorticalareastowhichitconnectsareresponsibleforsensorimotor
transformations.Thisreviewcoversnewresearchonfourcomponentsofthistransformationprocess:plan-
ning, decision making, forward state estimation, and relative-coordinate representations. These sensori-
motor functions can be harnessed for neural prosthetic operations by decoding intended goals (planning)
andtrajectories(forwardstateestimation)ofmovementsaswellashighercorticalfunctionsrelatedtodeci-
sionmakingandpotentiallythecoordinationofmultiplebodyparts(relative-coordinaterepresentations).
Introduction to action, including movement planning, decision making, the
Frontalandparietalareasarestronglyinterconnectedandfunc- formation of internal models, and coordinate transformations.
tiontogether formanyaspectsof actionplanning.Historically, ThereviewwillfocusontwoareasinthePPC:thelateralintrapar-
aroleoffrontallobeinactionhasbeenclear(FritschandHitzig, ietal area (LIP) and the parietal reach region (PRR). It will also
1870;Ferrier,1876).Theprimarymotorcortex(M1)isasourceof includeareasinthefrontallobeconnectedtoPPC,particularly
motorcommands(PenfieldandBoldrey1937;EvartsandThach, the dorsal premotor cortex (PMd), and other areas within the
1969),andmoreanteriorregionsofthefrontallobeareinvolved PPC,suchasarea5.Inthefinalsection,wewillshowexamples
inmanyhigher-levelaspectsofmovementplanninganddecision of‘‘proofofconcept’’inwhichtheaction-relatedactivityinthe
making (Miller and Cohen, 2001; Wise, 1985). The anterior PPC and PMd cortex can be decoded and used to provide
aspectoftheparietallobeiswellestablishedintheprocessing controlsignalsforneuralprostheticapplications.
of somatosensoryinformation (Mountcastle, 1957).Theposte-
riorparietalcortex(PPC)haspreviouslybeenconsideredimpor- MovementPlanning
tant for spatial attention, spatial awareness, and polysensory An important property of the cerebral cortex is its anatomical
integration (Critchley, 1953; Ungerleider and Mishkin, 1982; organization according to function. This fact is fortunate for
ColbyandGoldberg,1999).Inrecentyears,however,anumber neuroscientists, since this anatomical parcellation provides
of studies suggest that, although the PPC is involved in these a tractable approach to understanding cortical networks by
sensoryfunctions,ithasalsobeenshownindifferentcontexts studyingtheircomponentparts.
tobeimportantforaspectsofaction,includingmovementinten- ThePPChadpreviouslybeenconsideredasatypicalassoci-
tion and decision making (Mountcastle et al., 1975; Andersen, ation cortex containing largely two areas, Brodmann’s areas
1987; Andersen and Buneo, 2002; Gold and Shadlen, 2007; 5and7orVonEconomo’sareasPFandPG,basedoncytoarch-
Rizzolattietal.,1997;Kalaskaetal.,1997;Johnsonet.al.1996; itecture (Brodmann, 1909; Von Economo, 1929). As a typical
Burnodetal.,1999;Lacquanitietal.,1995;GrazianoandGross, associationcortex,PPC’sfunctionhasbeenthoughttoreceive
1998; Desmurget et al., 2009; Rushworth et al., 2001). In this convergent multisensory inputs, form a unitary map of space,
review,werefertointentionasmovementplanningatacognitive andthenrelayspatialinformationtothefrontalmotorareasto
levelratherthanatthelevelofmovementexecution(Andersen guide behavior. However, relatively recent neuron recording,
andBuneo,2002).Oneexampleindicativeofthismorecognitive neuroanatomical tracer, and BOLD imaging studies have
levelisthatthegoalsforvisuallyguidedreachmovementsare revealedaccumulatingevidenceofavarietyoffunctionalareas
encodedpredominantlyinvisualcoordinatesratherthanmuscle in the PPC (Andersen and Buneo, 2002; Mountcastle, 1998;
coordinates in the parietal reach region (PRR). Intention is not Rizzolatti et al., 1997; Graziano and Gross, 1998), particularly
meanttorefertopurposeorattitude(Schall,2004).Thestrong areas within the intraparietal sulcus (IPS; Blatt et al., 1990).
reciprocal connections between the PPC and broad areas of Furthermore, PPC is actively involved in movement planning
thefrontallobeanteriortoM1likelycomprisecircuitsforthese (Mountcastleetal.,1975;GnadtandAndersen,1988).Parietal
action-related processes (Andersen et al., 1990a; Goldman- and frontal areas share similar properties and work together
Rakic,1988).ThereviewwillpresentPPCwithintheframework through their association pathways in a collective manner
ofitsinvolvementwithanumberoffunctionsthatcanbebroadly (Johnson et al., 1996; Burnod et al., 1999). Individual areas in
classified as sensorimotor transformations (Andersen, 1987; the PPC have been found to encode different kinds of move-
Andersenetal.,1997;AndersenandBuneo,2002).Wewillfocus ments associated with different body parts. Area 5 represents
primarilyonnewresearchregardingfourrolesofPPCandasso- spatialinformationforlimbmovementandisinvolvedinreaching
ciatedfrontallobeareasinsensorimotortransformationsrelated arm movements (Lacquaniti et al., 1995; Kalaska et al., 1997).
568 Neuron63,September10,2009ª2009ElsevierInc.
Neuron
Review
Theanteriorintraparietalarea(AIP)appearsselectiveforgrasps
(Sakataetal.,1997;Baumannetal.,2009)andisinterconnected
with the ventral premotor cortex (PMv) (Tanne-Gariepy et al.,
2002), which also has activity related to grasp movements
(Rizzolatti et al., 1988). Inferior parietal lobule (IPL) neurons
have been demonstrated not only to encode specific acts but
alsotodischargeduringtheobservationofactsdonebyothers
(Fogassietal.,2005).ElectricalstimulationoftheIPLinhuman
patients triggered strong intention and desire to move their
body parts (Desmurget et al., 2009). This latter finding is very
important because recording data from monkeys show neural
correlates of intention (Gnadt and Andersen, 1988; Snyder
et al., 1997; Andersen and Buneo, 2002), but these human
experimentalinterventionsshowaroleofPPCintheawareness
of intention. Interestingly, stimulation of the premotor cortex
producedmovements,butthepatientsdeniedtheyhadmoved,
indicatingthatincreasedactivityinthepremotorcortexdidnot
leadtotheconsciousawarenessofintent.
Twoareasofparticularinteresttothisreview,LIPandPRR,are
respectively more selective for eye movements and reaching
(Andersen et al., 1987; Snyder et al., 1997; Quian Quiroga
Figure1. PopulationActivityofPRRNeuronsduringAutonomous
etal.,2006;CuiandAndersen,2007).LIPislocatedinapproxi-
TargetSelection
matelythemiddlethirdofthelateralbankoftheIPS.PRRwas Thetoprowshowsthemeanfiringrate(and95%confidencelimit)inthetarget
originallydefinedasanareamedialandposteriortoLIP(Snyder selectionforreachmovementstothepreferred(insidetheresponsefield,dark
etal.,1997)andmayhaveincludedmorethanonecorticalarea. curve)andnonpreferred(outsidetheresponsefield,lightcurve)target,aligned
tothetargetonset(time0).Bottomrowshowsthepopulationactivityinthe
ManysubsequentPRRstudieshavelargelytargetedthemedial
saccadetask.ReproducedwithpermissionfromScherbergerandAndersen
bankoftheIPS(Bhattacharyyaetal.,2009;Baldaufetal.,2008; (2007).
Pesaranetal.,2008;CuiandAndersen,2007;Scherbergerand
Andersen, 2007;Pesaran etal.,2006).Theseareas in turnare
largely connected to frontal lobe areas with similar functional The demonstration of effector specificity in no way excludes
selectivities—LIPtothefrontaleyefieldsandPRRtothePMd attention-drivenmodulationinadditiontothisspecificity(Snyder
(Andersen et al., 1985a, 1990a; Johnson et al., 1996; Tanne- etal.,1997;Andersenetal.,1997;QuianQuirogaetal.,2006).
Gariepyetal.,2002). However, the presence of planning activity in an area does
EffectorSpecificity caution against assuming that any increase in activity in PPC
Effectorspecificityingeneralreferstoactivitythatisspecificto during behavior must be attention related (for instance, see
planning to move or to moving a particular body part. In this ‘‘Potential Plans’’ section below) and emphasizes the impor-
review,wewillrefertothehandandeyepreferenceformove- tance of introducing controls to distinguish between attention
ment planning as effector specificity, and this term is meant andplanningcontributionstotheactivations.
to indicate relative, not absolute, specificity. For example, an A similar dissociation for reach and eye movements in PPC
areamaybeactiveforplanningareachorasaccade,butifit hasbeenseeninhumanimagingstudies.Thedegreeofdissoci-
issignificantlymoreactiveforoneplanovertheother,withall ationcomparedtooverlaphasvariedinthesestudies(Connolly
othervariablesbeingthesame,wewilllabeliteffectorspecific. etal.,2003;Astafievetal.,2003;Levyetal.,2007;Hagleretal.,
Since areas specific for reaching and looking are strongly 2007)andlikelyreflectsdifferencesintheexperimentaldesign.
interconnected within parietal and frontal cortex, no doubt for Experimentsthatfocusedonthe delay/planningperiod orthat
integrative purposes such as eye-hand coordination, it is not provided both the effector and target together rather than just
surprisingtofindsomedegreeofcommonactivation. theeffector,orthatusedreachingmovementsinsteadoffinger
Early studies examining neural activity during reaching and pointingmovementsprovidedthegreatestdegreeofseparation.
looking found a double dissociation, with LIP more active for PPCEncodesAutonomouslyChosenMotorPlans
saccades and PRR more active for reaches (Snyder et al., PPCneuronscanbeselectiveforexperience-dependentcate-
1997). A subsequent study showed that movement plans can gorical representations (Freedman and Assad, 2006; Toth and
be decoded better from populations of LIP and PRR neurons Assad,2002)andcognitivesetregardingtaskrules(Stoetand
thanthespatiallocationofthefocusofattention(QuianQuiroga Snyder,2004).Cuesareoftenusedinthesetasksandarestimuli
etal.,2006).Also,thelocalfieldpotentialsinPRRshowdistinctly thatinstructtheanimalsinwhattodo.Sinceearlierstudiesof
differentpatternsforreachingcomparedtolooking(Scherberger effector specificity in LIP and PRR used red stimuli to instruct
etal.,2005).Inautonomousreachtargetselectiontasks,PRR saccades and green stimuli to instruct reaches (Snyder et al.,
shows differential activity consistent with the spatial location 1997;QuianQuirogaetal.,2006),itispossiblethatthespecificity
of the chosen reach but little differential activity for saccade may be related to the meaning of the cue (i.e., red means
target selection (Scherberger and Andersen, 2007) (Figure 1). saccade,greenmeansreach)andstillbesensoryrelatedrather
Neuron63,September10,2009ª2009ElsevierInc. 569
Neuron
Review
thanreflectingactualplanning.Toaddressthispossibility,atask
wasdesignedinwhichthestimuliwerealwaysthesameandthe
monkeys autonomously chose whether to make a reach or
a saccade (Cui and Andersen, 2007). Thus, any difference in
activitycannotbeattributedtosensoryattributesofthestimuli
butratherreflectsthedecisionsandplansoftheanimals.
Figure 2 shows population activity from two animals for the
autonomouschoicetrials,withredindicatingwhenthemonkeys
choseasaccadeandgreenareach(CuiandAndersen,2007).
Themonkeysinitiallyfixtheirhandandeyesonafixationtarget
locatedstraightahead(Figure2A).Activityduringthisbaselineis
plottedfrom(cid:2)0.5to(cid:2)0.2s(Figures2Band2C).Next,atarget
appears in the response field of the neurons for 600 ms. The
response field is the restricted area in space that will activate
aneuronasaresultofastimulusappearingatthatlocationor
an action planned or executed to that location. This target is
composed of adjacent red and green parts (Figure 2A). When
the target is extinguished, the monkey chooses either a reach
or a saccade. The choice bias is balanced by having the
monkeyplayacompetitivegameagainstacomputeralgorithm
(Barraclough et al., 2004). The effector choice trials were
randomlyinterleavedwithinstructedtrialsinwhichthemonkey
is instructed to make a reach or a saccade by extinguishing
onlyoneofthecoloredcomponents(e.g.,iftheredpartstays
on, the monkey is instructed to make a saccade) (Figure 2A).
Theinstructedtrials(datanotshown)areintroducedforbehav-
ioralpurposesonly,sothemonkeydoesnotknowwhetherhe
is to make a decision or follow instructions until the target, or
partofthetarget,goesoff.Afterthetargetgoesoffinthedeci-
sion trials, there is a 600 ms delay during which the animals
decidewhichmovementtomake.Attheendofthisdelayperiod,
thefixationpointisextinguished,providingaGOsignalforthe
animalstomakethechosenmovement(Figure2A).
Duringtheperiodwhenthetargetisintheresponsefieldand
themonkeysdonotknowwhetheritisadecisionorinstruction
trial,theactivityishighinbothLIPandPRR(Figures2Band2C;
0–600ms).Whenthetargetextinguishes,theanimalsknowthey
arefreetochoose,buttheymustwithholdtheactionuntiltheGO
signal.Duringthisdelay(600–1200ms),theactivityseparatesin
LIPandPRR,withLIPcellsdifferentiallymoreactivewhenthe
animals choose a saccade and PRR cells differentially more
activewhentheychooseareach(Figures2Band2C).
Thiseffectorspecificityreflectingtheanimals’choiceforiden-
ticalvisualstimulimustberelatedtothedecisionandplanningof Figure2. PPCActivityduringEffectorChoiceTask
theanimalsandnottothesensorymeaningofthestimuli.Nor (A)Behavioralparadigminterleavedeffectordelay-instructedsaccade(top)
andreach(bottom)andeffectorchoice(saccadeorreach)trials(middle).
canthisdifferentialresponsebeduetospatialattention,since
(B)and(C)PopulationhistogramsaveragedacrossallisolatedLIP(B,n=100)
the targets are always in the same location in space, and it is andPRR(C,n=91)neuronsduringsaccade(red)andreach(green)chosen
only the effector choice that varies. Interestingly, at the time trials.Theverticalthinlinesindicatecueon,cueoff,andcentralfixationoff
(GOsignal),respectively.Thehorizontalthinlineindicatesbaselineactivity,
immediately after the GO signal and before the reaching arm
whichwasdefinedbymeanfiringrateduringthe300msintervalbeginning
movementismade,theactivityinLIPisstatisticallynodifferent from 500 ms before cue onset for both saccade and reach chosen trials.
frombaseline,eventhoughattentionisknowntobeattractedto Post-GOactivity(0–100msintervalafterGO)ofLIPpopulationwassignifi-
cantlyhigherthanthebaseline(p<0.005)ifthemonkeysdecidedtosaccade,
reach targets(Baldaufetal.,2006;Deubel etal.,1998).Bisley
butdroppedtobaseline(p>0.5)ifthemonkeysdecidedtoreach.Ontheother
and Goldberg (2003) have proposed that LIP forms a priority hand,post-GOactivityofthePRRpopulationwassignificantlyhigherthanthe
mapforattentionandthatitistherelativeamountofactivation baseline(p<0.0001)intrialsinwhichreacheswerechosen,butdroppedto
inLIPandnottheabsolutemagnitudethatindicatesthelocation baseline(p>0.8)intrialsinwhichsaccadeswerechosen.Reproducedwith
permissionfromCuiandAndersen(2007).
ofhighestpriority.Inthiscase,thereisnopremovementactivity
in the LIP population when the monkey is planning the reach,
arguingagainstthispriorityformulation.Likewise,inthisperiod
570 Neuron63,September10,2009ª2009ElsevierInc.
Neuron
Review
right after the GO signal, PRR is inactive when the monkey beenused toseparate sensory goalsfrommovement plansin
is choosing a saccade, even though saccades also attract antisaccade and antireach tasks. In some studies, monkeys
attention to the target (Deubel and Schneider, 1996; Kowler have been trained to move in the opposite direction to the
etal.,1996). appearance of a stimulus. It is reasoned that if the cells only
PotentialPlans encode the location of the stimulus, they are sensory related,
Both LIP and PRR showed vigorous activation during the cue andiftheyonlyencodethelocationofthegoal,theyaremove-
periodinthestudyofCuiandAndersen(2007)(0–600msepoch; mentrelated.
Figures2Band2C).Itcouldbearguedthatthisactivityreflects AntisaccadetaskswithrecordingsinLIPhaveyieldedmixed
sensory activity and/or top-down attention. Since the monkey results with respect to this sensory-motor dissociation, with
isnotsurewhetherhewillbeinstructedtomakeaneyeoran one report indicating largely sensory encoding (Gottlieb and
armmovementorwhetherhemaybefreetochooseaneyeor Goldberg, 1999) and another indicating largely movement
arm movement, the neural activity may also reflect potential encoding (Zhang and Barash, 2000, 2004). Differences in the
or default planning in which the monkeys form potential plans details of the behavioral tasks may account for these different
for both movements and then later select between the two. findings. In reach tasks, antireaches produce brief activation
Duringthe cueperiod, when astimulus ispresent, the activity forthecueinarea5of posteriorparietal cortexandPMd,fol-
ishigherthanduringthedelayperiodinwhichplanninganddeci- lowed by activity coding the intended direction of the reach
sionmakingtakeplace.Thisadditionalactivitymayreflectthe movement(Kalaska,1996).
fact that there is a visual stimulus present in the cue period Antireach experiments in PRR (Gail and Andersen, 2006)
andnot thedelay period, indicating that atleast acomponent produced results similar to those of Zhang and Barash (2000,
oftheactivityissensoryinnature. 2004) and Kalaska (1996). The task for the PRR experiments
Default plans for spatial locations have been reported in used four different directions for pro- and antimovements so
a number of sensorimotor and motor structures (e.g., Cisek thatthespatialtuningofthecellsforbothrulescouldbedeter-
andKalaska,2005;BassoandWurtz,1998;PlattandGlimcher, mined. Briefly flashed targets were used, and variable delays
1997).Ithasbeenproposedthatmakingdecisionsbetweentwo wereinterposedbeforetheGOsignaltotargetplanningactivity.
spatial locations may involve competition between potential Finally,thetaskruletobeappliedwasprovidedeachtimeatthe
plans(Cisek,2006),notdissimilartothecompetitivebiasmodel beginningofthetrial,priortothepresentationofthetargetcue.
forspatialattention(DesimoneandDuncan,1995).Infact,most Most cells showed tuning only to the planned reach direction
decision-makingmodelscontainacompetitionbetweenpoten- (45%).Asmallerproportion,termed‘‘visuomotorcells,’’showed
tial plans (Wang, 2008; Beck et al., 2008). Forming potential brieftuningtothetargetlocationfollowedbytuningtothemove-
plansmaynotonlybenefitdecisionmakingbutmayalsoreduce mentdirection(7%).Thenumberofcellstunedtoonlythetarget
reactiontimesandthushaveevolutionary/survivaladvantage. locationwasstatisticallyinsignificant.Thefactthatmostofthe
Potential movement plans toward multiple spatial targets cells encoded only the movement direction rules out attention
coexist in virtually the entire parietal-frontal network, even as a major contributing factor for these neurons. An attention
including primary motor cortex (Bastian et al., 2003). It would explanation also does not appear to apply to the visuomotor
be interesting to determine whether there are regions of the cells.ThetimeoftheGOsignalwasnotpredictable,anditwould
posteriorparietalandfrontalcortexthatdonotreflectpotential be expected that the monkeys shift spatial attention at least
plans and only code the decision outcome in effector choice partiallytowardthefixationpointwheretheGOsignaloccurred.
tasks. If such areas exist, it would suggest that PRR and LIP Instead, the movement goal tuning in the visuomotor tuned
carrysignalsrelatedtoboththepotentialplansandoutcomes neuronspersistedthroughoutthedelayperiodandatthepopu-
oftheeffectordecisionprocess,whileotherareaswithinparietal lationlevelbecamestrongesttowardtheendofthedelayperiod.
andfrontalcortexcarryonlytheoutcomeofthedecision.Such Subsequent modeling studies (Brozovic et al., 2007) have
a distinction would be consistent with the decision process examined how this context information might be integrated in
occurring within a parietal-frontal circuit that includes PRR PRRwiththetarget-stimuluslocationprovidedlaterinthetrial,
and LIP, with competition between potential plans and the similartothetaskparametersinthestudybyGailandAndersen
outcome being computed within that circuit. Areas that only (2006).Thenetworksintegratedthetargetlocationandcontext
represent the outcome could be downstream of this effector through a classic gain field mechanism (Zipser and Andersen,
decision process and receive information conveying results of 1988;Brozovicetal.,2007,2008).Themodelingstudiesshowed
deliberation from the decision network. However, it is also thatthecontextcouldoriginatefromfeedbackfromtheoutput
possible that the decision process occurs entirely outside of (motor) layer of the network, consistent with feedback from
PRRandLIPandtheseregionsreflectonlythepotentialplans frontallobestructures,orfrominputtothemiddlelayer,which
butare kept updated; that is,these areas arecoding decision represents PRR. Thus, the context information could originate
outcomes because there is only one plan after the decision is eitherfromtop-down(e.g.,fromfrontalcortexorotherparietal
made. areas)orbottom-up(e.g.,fromextrastriatevisualareas)sources,
Antimovements although the authors suggested that the former route is more
Not all movements are oriented toward a visible goal. For likely. A more recent study has shown that individual neurons
instance, one might wish to reach to a soda can but reach in PRR and PMd are gain modulated by context information
awayfromabee.Inthelattercase,thevisualstimulusandthe (Gail et al., 2009), consistent with the neural network models
goalofthemovementarenotcongruent.Thisdiscordancehas (Brozovicetal.,2007).
Neuron63,September10,2009ª2009ElsevierInc. 571
Neuron
Review
SequentialReachMovements takerouteAratherthanrouteBtothelabtoday).Anexample
Manynaturalreachbehaviorsentailsequencesofmovements. of the external variety is a perceptual decision, in which the
Frontal cortical areas have been shown toencode information subjectviewsanoisyorambiguousdisplayandmakesachoice
includingsubsequentmovementparts,directions,andtemporal basedonthesubject’spercept(Newsomeetal.,1989;Bradley
organization (Tanji and Shima, 1994; Shima and Tanji, 1998; et al., 1998; Dodd et al., 2001). Studies of area MT show that
Ninokuraetal.,2003;FujiiandGraybiel,2003;Ohbayashietal., trial-to-trial variation in neural firing can affect the perceptual
2003;LuandAshe,2005;Mushiakeetal.,2006;HistedandMiller, choice of animals in deciding which direction they perceive
2006;Averbecketal.,2006;Shimaetal.,2007).However,thefirst motion at low thresholds (Britten et al., 1996). In this type of
sequentialmovementstudyofPRRshowedactivityonlyforthe experiment, the direction of the perceived motion is indicated
movementthatwasnextinthesequence(BatistaandAndersen, byeyemovements,andareaLIPintegratesperceptualevidence
2001).Thisstudyusedacomplicatedtaskthatlikelyresultedin formakingthedecision(ShadlenandNewsome,1996).
the animals planning only one movement at a time. Recently, Internal(autonomous)decisions(Coeetal.,2002;Barraclough
theseexperimentshavebeenrepeatedwithasimplerparadigm etal.,2004;CuiandAndersen,2007)aresometimesreferredto
thatpromotesplanningtwomovementssimultaneously.Inthis as ‘‘free choice’’ in which selections are made concerning
newerparadigm,bothmovementplansaresimultaneouslyrep- ‘‘where,when,orhow’’(Haggard,2008).TheexampleinFigure1
resentedintheactivityofPRRneurons(Baldaufetal.,2008). showsselectionof‘‘where’’byPRRneurons(Scherbergerand
In the Baldauf et al. (2008) study, visual stimuli instructing Andersen, 2007). The effector choice task in Figure 2 uses
areach locationproduced ahugeresponseinPRR,butwhen a ‘‘how’’ decision (Cui and Andersen, 2007). An advantage of
the same stimulus was used as a timing cue and not a reach using a ‘‘how’’ decision task for neurophysiological study is
targetitproducedalmostnoresponse.Thiscontext-dependent that the locus of attention covaries in space with the ‘‘where’’
gatingforPRRissimilartothatseenforcuetargetsinantireach decisionsbutdoesnotwiththe‘‘how’’decisions,andthustwo
tasks mentioned above (Gail and Andersen, 2006). These sources of potential activation, spatial attention and planning,
findings contrast with much larger responses seen in LIP for canbemoreeasilyseparatedforeffectordecisions.
flashedirrelevantcues(BisleyandGoldberg,2003;Powelland Additional evidence for a role of LIP and PRR in decision
Goldberg,2000)ortimingcues(Camposetal.,2009).Thediffer- making is the finding that they encode the expected value of
ence in degree of visual response between LIP and PRR may the reward for a movement (Platt and Glimcher, 1999; Sugrue
indicatethat(especiallyflashed)stimulitendtoformautomatic et al., 2004; Musallam et al., 2004; Yang and Shadlen, 2007)
potentialeyemovementplansbutnotreachplans.Thisisplau- (Figure 7). Neurons in the putamen and caudate nucleus have
sible behaviorally, considering the much greater frequency been shown to encode action value (Samejima et al., 2005).
of saccades compared to reaches. Interestingly pop-out dis- Actionvalueisthevaluethatapotentialactionwouldproduce,
tracters (made salient by their physical properties) produce regardless of which action is chosen. It can be used to bias
less activity than non-pop-out targets when monkeys are per- selectionofaparticularaction.Intheoculomotorregionofthe
forming visual search witheye movements (Ipata etal., 2006). caudate,cellsarefoundthatcodeactionvalue,chosenvalue,
The authors proposed that this modulation was due to top- andthechoiceofthesaccade(LauandGlimcher,2008).Chosen
downmodulationofsalienceinLIP.However,itisalsopossible value refers tothe value that a chosen action really produces,
that top-down influences may regulate the level of activity of anditcanbeusedforreinforcementlearning.Itisnotcurrently
potential eye movement plans represented in LIP. Another clear whether LIP and PRR neurons encode action value or
possibilityisthatthereisastrongercouplingofsensorysignals chosenvalue(Rangel,2009).
withmovementplanninginLIPcomparedtoPRR. SelectionofanActioninParietal-FrontalCircuits:
Inrapidhand-movementsequences,attentioninhumanshas IntegratedandSerialModels
been shownto be distributed among target locations (Baldauf Decision making traditionally has been considered a separate
etal.,2006),similartotheactivityinPRR.Asdiscussedabove, process from action planning (Tversky and Kahneman, 1981),
the findings of effector specificity, the coding of mostly reach asillustratedinFigure3A.However,recentneurophysiological
goalsinantireachtasks,andthelackofevokedactivitytoflashed studiessuggestthatpotentialplansformovementstomultiple
timingcuesstronglysuggestthatPRRcodesreachplans.Itis target locations are simultaneously represented in a collection
possiblethatPRRhasatop-downinfluenceonextrastriateareas ofmotor-relatedareas(e.g.,ShadlenandNewsome,2001;Cisek
andbiasestheprocessingofsensorystimuli.Thiseffectwouldbe andKalaska,2005;BassoandWurtz,1998;PlattandGlimcher,
similartotheeffectoffrontaleyefield(FEF)activityonthemodu- 1997). Thus, target selection and movement preparation may
lationofattentioninV4(MooreandArmstrong,2002)andcould involve the same brain circuits and are performed in an inte-
beaccomplishedbydirectingattentiontoreachgoalsthrough grated manner as diagrammed in Figure 3B (Shadlen and
its feedback projections to visual areas. Such a mechanism Newsome, 2001; Fagg and Arbib, 1998; Cisek, 2006, 2007;
would indicate at least partially separate top-down control of Wang, 2008;), as opposed to a serial model in which decision
attentionforstimulithataretargetsforreachesandsaccades. makingoccursbeforeactionplanning(Schall,2002).Neverthe-
less,thisideahasonlybeentestedforspatialtargetselection,
DecisionMaking:ActionSelectioninParietal-Frontal whichinvolvesspatialattention,whichinturnengagesnumerous
Circuits brain areas. It remains unknown whether plan selection and
Decisionprocessesrangefromthosethatarelargelyexternally movement preparation are represented in segregated brain
driven(e.g.,stopwhenthelightisred)tointernallydriven(e.g., areas for other kinds of decision making. In the nonspatial
572 Neuron63,September10,2009ª2009ElsevierInc.
Neuron
Review
2006; Sejnowski and Paulsen, 2006; Salinas and Sejnowski,
2000).Inotherwords,ifcorticalareaAprojectstocorticalarea
B,andthephaseofthespikesfromcorticalareaAareinphase
withthemembraneoscillationsofcorticalareaB,andthisphase
issuchthatspikesarriveatthelowthresholdphaseoftheoscil-
lations,thenspikesfromareaAaremorelikelytoproducespikes
inareaB.Duringtheseperiods,therewouldbegreatercommu-
nicationorinfluenceofareaAonB.Anexperimentalprediction
suggests that during periods of greater communication there
willbegreatercoherencebetweenthephaseofspikinginone
areaandtheLFPinanother(ifinfactthephaseoftheincoming
spikesandthemembraneoscillationsareatthelowthreshold
periods of the oscillations). Changes in spike-field coherence
maybeausefulsignaturefortracingthedynamicsofcommuni-
cation between cortical areas (Pesaran et al., 2008; Gregoriou
Figure3. IllustrationsofTwoTheoreticalFrameworksofDecision
etal.,2009).
MakingandActionPlanning
(A)Traditionalserialmodelinwhichdecisionmakingisconsideredaprocess Increases in spike-field coherence have been recorded
separatedfromactionplanning. between PRR and PMd while monkeys select between reach
(B) Recent integrated model. Neurophysiological evidence suggests that targets(Pesaranetal.,2008).Insometrials,theselectionwas
potentialplansformovementstomultipletargetlocationsaresimultaneously
instructed, and in others the monkeys chose the targets. In
representedinacollectionofparietalandfrontalareas,asopposedtoaserial
modelinwhichdecisionmakingoccursbeforeactionplanning. the autonomous decision tasks, the PMd-PRR and PRR-PMd
spike-fieldcoherencesweregreaterforchoicethanforinstructed
trials (Figure 4). As mentioned above, it is possible that the
effectorchoice,PRRandLIParefoundtorepresentbothpoten- increasedcoherencemayrepresentdifferencesincommunica-
tialplansandtheoutcomeofthedecision,andthuspotentially tionbetweenthetwocorticalareasduringautonomouschoice
participateinthedecisioncircuitandcontributetothedelibera- andinstruction.Interestingly,onlyaboutaquarterofthepaired
tion.Itwillbeinterestingtodetermineifeffectordecisionsalso recordingsshowedsignificantspike-fieldcoherence,andthese
followthisintegratedmodeloriftheyhaveanadditionalhierar- pairsindicatedthedecisionoftheanimalearlierthanpairsthat
chicalcomponentthatcodesonlytheselectedplandownstream didnothavesignificantcoherence.Theseresultssuggestthat
ofdecisioncircuits. theremaybeasubsetofcellsconnectingthesetwoareasthat
In effector choice, if the deliberation is carried out by a coordinatesthedecisionprocess.
competition between potential saccade and reach plans, this TimingintheCircuit
competition may be at least partially carried out between LIP Timing within the parietal-frontal decision circuits provides
and PRR within the parietal lobe. From a theoretical point of insightintowhichareasmayencodethedecisionearlier.These
view,LIPandPRRsharethesame(predominantlyeye-centered) experiments require simultaneous recordings from the same
reference frame, which would benefit such a competitive animalsinordertokeepconstantsuchfactorsasleveloftraining,
computation. performance,andotherexperimentalvariablesthatmightinflu-
CorticocorticalCommunication encetimingcomparisons(MillerandWilson,2008).
Localfieldpotentials(LFPs)areoscillationsintheelectricalfield One would normally expect that decision-related activity
withinafewhundredstothousandsofmicronsoftherecording beginsearlierinfrontallobeareasandpassesbacktoparietal
microelectrode tip driven largely by synchronous activity of areas (Monosov et al., 2008). This seems to be the case with
synaptic potentials, but action potentials also contribute when instructedandautonomousdecisionsforselectingspatialloca-
they are sufficiently synchronous (Mitzdorf, 1987; Logothetis tions. Spike activity appears first in PMd and later in PRR for
and Wandell, 2004). Using combined optical and electrical theonsetofthetargetstimuli(Pesaranetal.,2008).Thisresult
recordingin V1cortex,ithas recentlybeenestimatedthatthe isabitsurprisingsinceitsuggeststhatarouteotherthanthrough
spread of the LFP is very local, with greater than 95% of the PRR,perhapssubcortical,producesthisveryearlyactivationof
signaloriginatingfromwithin250mmoftherecordingelectrode PMd. As shown in Figure 4, spike-field coherence timing also
(Katzneretal.,2009).Spikesofindividualneuronsoccurlargely suggeststhatthePMdtoPRRlinkofthecircuitisactivatedfirst
onthenegativepeaksoftheoscillations,suggestingthatduring followedwithinafewmillisecondsbyahand-shakebackfrom
this phase of the LFP oscillation the membrane potentials are PRRtoPMd.Forthismeasure,absolutetimingcannotbedeter-
closesttothresholdforspiking.Themagnitudesoftheoscilla- mined because the coherence is estimated using an analysis
tions in certain frequency bands are modulated with attention window of ±150 ms stepped in 10 ms increments; however,
andmotorpreparationintheparietal,occipital,andfrontallobe relative timing between the PMd-to-PRR and PRR-to-PMd
areas (Fries et al., 2001; Pesaran et al., 2002; Scherberger spike-field coherences can be determined because they are
et al., 2005; Murthy and Fetz, 1996; Sanes and Donoghue, obtained with the same analysis methods. Similar results of
1993; Mehring et al., 2003). It has been proposed that these frontalareasleadingparietalareashavebeenfoundineyemove-
oscillationsmaysynchronizewithinputstoanarea,increasing menttasksinwhichthesupplementaryeyefields(SEF)leadLIP
communication between regions of the brain (Mishra et al., (Coeetal.,2002).
Neuron63,September10,2009ª2009ElsevierInc. 573
Neuron
Review
Figure4. PMd-PRRSpike-FieldCoherenceIs
StrongerintheTrialsinwhichtheTargetWas
AutonomouslySelected
Populationaverage15HzPMd-PRRspike-fieldcoherenceis
plottedevery10ms:PMdspike-PRRfieldcoherence(solid);
PRRspike-PMdfieldcoherence(dashed).Freesearch(black);
instructed search (red). Coherence is z transformed before
averaging; 95% confidence intervals, Bonferroni corrected
(shaded). Reproduced with permission from Pesaran et al.
(2008).
efferencecopysignalstopredictthecurrentstate
of the limb and integrates this information with
delayed sensory observations to subsequently
learntoimprovethisestimate(Figure5A).
In addition to being used for online correction,
aforwardmodelcanalsobeusedtoestimatethe
sensoryconsequencesofamovement(Haarmeier
etal.,1997).Suchestimatescanbeusedtodistin-
guishmovementofaneffectorfrommovementin
the world. For instance, it is believed that the
apparent stability of the world during eye move-
ments, which sweep the visual scene across the
However,thetimingwithinthiscircuitmaydependonthetask retinas,isachievedbyaforwardmodelthatmakesuseoffeed-
athand.Forinstance,themoreconventionalcenter-out reach backofeyemovementcommands(VonHolstandMittelstaedt,
task produces approximately simultaneous activation in PRR 1950;vonHelmoltz,1866;Haarmeieretal.,2001).
andPMd(Pesaranetal.,2008).Inattentiontasks,FEFisactive InarecentstudyofPPC,monkeyslearnedtouseahand-oper-
firstforvisualsearchtasksthatarebelievedtorelyontop-down atedmanipulandum(‘‘joystick’’)tomoveacursoronacomputer
attention, but LIP is active first for pop-out stimuli that are screen.Itwasfoundthatwhenthemonkeysmovedthecursor
considered indicative of bottom-up attention (Buschman towardatargetthatnotonlytheeventualgoalofthemovement
andMiller,2007;butseeSchalletal.,2007).Itwillbeimportant but also the instantaneous direction of the cursor was repre-
todeterminethetimingofeffectorchoiceintheparietal-frontal sented(Mullikenetal.,2008a).Figure5Bshowsthestaticgoal
circuit. Interestingly, in trials where the effector is instructed angle—thevectorfromthefixationpointtothetarget.Therow
before the target, effector-specific activity diverges in parietal ofdotsrepresents15mssamplesofthecursoralongthetrajec-
cortexearlierthaninfrontalcortex(LawrenceandSnyder,2006). tory,andtheinstantaneousdirectionofmovementatonepoint
in the trajectory is shown and labeled the movement angle.
InternalModels:PPCPredictstheCurrent Figure 5C shows four center-out movements, and Figure 5D
StateofEffectors shows two movements around obstacles. The obstacles were
ReachDynamics usedtoincreasetherangeofmovementangles.Thetaskused
Clinical and transcranial magnetic stimulation studies provide 8 or 12 targets. All trials across all movement angles were
evidencethatthePPCinhumansisinvolvedinonlinecorrections used to construct space-time tuning functions for each cell.
of reach movements (Desmurget et al., 1999; Della-Maggiore This tuning function measures the instantaneous firing rate as
et al., 2004; Pisella et al., 2000). For reach movements, both a function of angle and lag times. The lag time is the relative
somatosensorystimulationfromthelimbmovementandvisual time difference between the instantaneous firing rate and the
stimulationfromwatchingthemovementareimportant.Somato- timethataparticularmovementangleoccurred.Asimilartuning
sensoryandvisualsignalsconvergeinthePPC,particularly in curvecanalsobeobtainedformutualinformation.Bothtypesof
PRR and adjoining Brodmann’s area5, and canprovide feed- measureproducedsinglepeakedtuningcurves.Figure5Eplots
backsignalsformakingcorrectionsduringreaching.However, theoptimallagtimeforthemostinformationaboutmovement
there is a considerable delay for these signals to reach PPC: angle for the population of neurons for the center-out and
(cid:3)30 ms for somatosensory and 90 ms for visual signals obstacle tasks. The optimal lag for the center-out task was
(Figure5A).Suchlongdelaysinfeedbacksystemscanleadto peakedat0msandfortheobstacletaskitledby30ms.Motor
instability.Toobtainanaccurateestimateofthecurrentstate, command signals would normally lead by 90 ms, and sensory
i.e., position, direction, and speed of a limb, requires more feedback would lag by 30 (somatosensory) or 90 ms (visual).
thansensorysignals.Ithasbeenproposedthatefferencecopy Although there are some cells that show these large lead and
signals, replicas of movement commands from motor areas, lag times, the population response is centered within the
arefedbacktoPPCtoeliminateanydelay(JordanandRumel- dynamicrangeinbetween.Thisisthedynamicrangeconsistent
hart, 1992; Wolpert et al., 1995; Shadmehr and Wise, 2005). withanefferencecopythatisusedforforwardstateestimation
Considerable evidence points to a ‘‘forward model’’ that uses (Figure 5E). Since the hand movement and cursor movement
574 Neuron63,September10,2009ª2009ElsevierInc.
Neuron
Review
werecorrelated,theresultsareconsistentwithbothaforward
modelpredictingthehandmovementandthecursormovement.
EyeMovements
Eyemovementactivitymayalsobeconsistentwithforwardstate
estimationinthePPC.Theeye-centeredlocationofatargetfor
asaccadeinthesuperiorcolliculusandLIP(MaysandSparks,
1980;GnadtandAndersen,1988)orareachmovementinPRR
(Batista et al., 1999) compensates for intervening saccades.
This compensation can occur as well for stimuli that are not
the targets for a subsequent saccade (Duhamel et al., 1992).
Althoughithasbeenproposedthattheresponsefieldsshiftto
take into account the eye movement, it is more parsimonious
to consider the activity shifting within the eye-centered map
(Xing and Andersen, 2000a). The shift of activity in LIP often
beginspriortotheeyemovement(Duhameletal.,1992).Since
thelocationcanbeidentifiedaftertheeyemovementbysensory
input, it has been proposed that this predictive shifting is a
signatureofaforwardmodelpredictingthelocationofastimulus
aftertheeyemovement(Vazirietal.,2006).
Eye-movement-related signals for pursuit, fixation, and
saccadeshavebeenreportedinPPC(Lynchetal.,1977;Mount-
castleetal.,1975).Itwouldbeinterestingtoexaminewhether
the pursuit signals indicate the instantaneous direction of the
eye movement with zero lag time, similar to the reach-related
signalsinPRR.MSTneuronsaretunedforthefocusofexpan-
sioninsimulatedopticflowstimuli,andthesesignalscompen-
sateforeyemovementsusinganefferencecopyofpursuiteye
movementsandheadmovements(Bradleyetal.,1996;Shenoy
etal.,1999;Leeetal.,2007).Suchcompensationisagainindic-
ativeofforwardmodelsforthepurposeofperceptualstability.
As has been shown for saccades, it may be possible that
the fixation-related activity is also predictive in PPC; in this
case, the predictive component would provide the current
fixation location of the eyes beginning at zero lag after an eye
movement has brought the eye to the fixation location. Such
afindingwouldpointtowardageneralizationthatmanymove-
Figure5. PPCNeuronsPredicttheCurrentStateofEffectorinthe
JoystickControlTask ment-relatedresponsesinPPCareefferencecopiesforforward
(A)DiagramofsensorimotorintegrationforonlinecontrolinPPC.Inputsto stateestimations.
PPCconsistofvisualandproprioceptivesensorysignalsand,potentially,an
efferencecopysignal.PlausiblePPCoutputsarethestatictargetdirection
CoordinateTransformations:RelativeEncoding
(goalangle)andthedynamiccursorstate(movementangle).
(B)Diagramofactualtrajectoryshowingthegoalangleandmovementangle forHand-EyeCoordination
andtheirrespectiveoriginsofreference.Thefilledgreenandredcirclesrepre- AreasinvolvedineyemovementssuchasLIPandFEFencode
sentthetargetandfixationpoint,respectively.
targets in predominantly eye-centered coordinates, although
(C)Exampletrajectoriesforcenter-outtask.Thedashedgreencircleisthe
starting location of the target and is not visible once the target has been the responses of LIP neurons can be modulated by eye and
jumpedtotheperiphery.Dotsrepresentcursorpositionsampledat15ms headpositionsignalsandFEFbyeyeposition(Andersenetal.,
intervalsalongthetrajectory(black,monkey1;magenta,monkey2).
1990b; Brotchie et al., 1995; Cassanello and Ferrera, 2007).
(D)Exampletrajectoriesfortheobstacletask.Targets,fixationpoints,and
cursorrepresentationsareidenticaltothecenter-outtask.Bluefilledcircles Thesemodulations,referredtoasgainfields,canbeeithermulti-
representtheobstacles. plicative or nonlinear additive effects and are believed to be a
(E)Histogramsummarizingtheoptimallagtimes(OLTs,thelagtimethatcon- first step in the transformation from eye coordinates to head
tainedthemaximalmutualinformation)formovement-angleneuronsforboth
andbodycoordinates(Andersenetal.,1985b;ZipserandAnder-
center-outandobstacletasks.Manyoftheseneurons’OLTswereconsistent
withaforwardestimateofthestateofthemovementangle,whichdidnot sen, 1988; Brozovic et al., 2008). Electrical stimulation of LIP
directlyreflectdelayedsensoryfeedbacktoPPCnorweretheycompatible producesfixed-vectorsaccadesineyecoordinates,consistent
with outgoing motor commands from PPC. Color-coded horizontal bars
withaneye-centeredrepresentation(ThierandAndersen,1996,
beneaththeabscissadenotetheapproximatelagtimerangesforsensory
(blue),forwardestimate(black),andmotor(red)representationsofthestate 1998;Constantinetal.,2007).
of the movement angle. Reproduced with permission from Mulliken et al. Visuallyguidedreachingrequirestransformationfromeye-to
(2008a).
limb-centered coordinates. The ventral premotor cortex in the
frontallobecodesvisualtargetsforreachinginbody-centered
coordinates(Grazianoetal.,1994).Interestingly,althoughPRR
Neuron63,September10,2009ª2009ElsevierInc. 575
Neuron
Review
Figure6. TheReferenceFrameDissociationTask
UsedtoExaminetheRelativeCodingofHandandEye
Positions
(A)Thereferenceframedissociationtask.Intheupperpanels
areachfromthesameinitialhandpositionismadetooneof
four target positions while fixation is maintained at one of
four eye positions. In the lower panels, four reaches are
madefromfourdifferentinitialhandpositions andasingle
eyeposition.Theoverallmatrixcontainsfourinitialhandposi-
tionsbyfoureyepositionsbyfourreachtargetsfor64trial
types.
(B)Geometryofrelativecoding.PMdneuronsrepresentthe
relativepositionsofthehand,eye,andreachtarget.Repro-
ducedwithpermissionPesaranetal.(2006).
isinvolved in reaching,itrepresents visualreach targets more tivetotheeye(eye-centered,Te),thetargetrelativetothehand
consistently in eye coordinates. This result was found in three (limb-centered, Th), the eye relative to the hand (eye-in-hand,
different studies using two very different analysis techniques Ehorequivalentlyhand-in-eye,He),aswellascombinationsof
(Batista et al., 1999; Buneo et al., 2002; Pesaran et al., 2006) twooreven all three(Figure 6B). Theseresults predict that,in
and with stimulus configurations that cannot bias for eye- some neurons, a unique relative spatial relation of all three
centeredcoordinates(Pesaranetal.,2006).Area5ontheother variables will produce the same activity for different absolute
handhasbeenshowntocodereachessimultaneouslyineyeand positions in space. Likewise, the relative position of these
limbcoordinates(Buneoetal.,2002).Itscellsshowpartialshifts parameters is encoded within the population activity of PMd.
intheirresponsefieldswitheithereyepositionorlimbposition. Thisformofencodinghasanadvantageforhand-eyecoordina-
Experimentsexaminingtherepresentationofauditorytargets tionindefiningacoordinateframebasedonthe‘‘workspace’’of
for saccades in LIP showed cells with response fields in eye thehand,eyes,andreachtarget.
coordinates,headcoordinates,and‘‘intermediatecells,’’which Inthesamestudyoutlinedabove,itwasfoundthatPRRcodes
showedonlypartialshiftingineyecoordinates(Stricanneetal., only in eye coordinates (Te), consistent with previous studies
1996). The cells with eye-centered coordinates are interesting (Batistaetal.,1999;Buneoetal.,2002).Anearlierstudyofthe
since sound localization begins as interaural differences in coordinatesofreachplanninginparietalarea5foundneurons
intensity,timing,andspectra(i.e.,head-referenced)butneeds coding simultaneously in eye (Te) and hand (Th) coordinates
to be converted to eye coordinates to saccade to auditory (Buneoetal.,2002).Itremainstobedeterminedifcellsinarea
stimuli. A similar distribution between head, eye, and interme- 5alsocodehand-in-eyecoordinates(He)andthushaveasimilar
diaterepresentationswasfoundforreachingtoauditorytargets relativecoordinatecodeasPMd.
(CohenandAndersen,2000).Intermediaterepresentations,i.e., IfPMdisinvolvedinhand-eyecoordination,thenthecellsin
partially shiftedresponsefields,havebeenexaminedinthree- this area should also code the plan to make saccades and
layerneuralnetworkmodelsthattransformcoordinatesbetween wouldbedistinctfromPRR,whichhasprimarilypostsaccadic
theinputandoutputlayers.Intermediaterepresentationsoccur responses(Snyderetal.,2000).IfPMdcellscodebothreaches
in the middle layer if there is strong recurrent feedback from andsaccades,thenanadditionalpredictioncanbemadethat
the output layer or if the network has separate output layers the saccade targets should be encoded in the same relative
that code in different coordinate frames (Deneve et al., 2001; coordinate frame as reaches, that is relative to the hand, eye,
XingandAndersen,2000b).Bothoftheseconditionsarebiolog- andsaccadetarget.
icallyplausibleforPPCareas. Attention
RecentstudiesbyMullette-Gillmanetal.(2005,2009)investi- Thecurrentreviewhasemphasizedplanning,decisionmaking,
gatedauditoryandvisualsaccadesfromdifferenteyepositions forward state estimation, and coordinate transformation roles
andreportedthatalmostallLIPandPRRneuronsintheirstudy forthePPCandareasofthefrontallobetowhichitconnects.
code in intermediate coordinates for both auditory and visual Ofcourse,anotherroleofPPCisinattention.Classically,atten-
targets.Likelyreasonsforsuchresultsarenoisydata,theprob- tion has been considered a sensory phenomenon in which
ableanalysisofmanyuntunedcellsgiventheirselectioncriteria stimuli are selected from the environment for further neuronal
and use of only saccades, and the widespread sampling and processing. However, the definition and scope of attention
lumping together of data with poor histological verification. havebeenexpandingintheliteraturetoembracesuchconcepts
Unfortunately,allofthesefactorswouldsumandstronglybias as‘‘motorattention’’thatisspecifictotherespondingeffector
theresultstowardasingleoverarchingcategory. (Rushworth et al., 2001) or that includes decision making by
Hand-eyecoordinationrequiresaninteractionbetweenbody attentional selection among motor plans (Goldberg and Seg-
parts,anditisofinteresttodetermineinwhatcoordinateframe raves, 1987). A distinction has been made between overt and
theseinteractionsareaccomplished.Recentlyaunique,relative covert orienting for attention with overt changes observed by
representationofcoordinateshasbeenfoundinPMd(Pesaran shifts of gaze and covert changes observed by other means
etal.,2006).Thecoordinateframeofreachtargetswasdeter- (Posner,1980).Whilevoluntaryeyemovementsareabehavioral
mined by independently changing the relative positions of the measureofshiftsofattention,itwouldbeerroneoustoarguethat
reachtarget,initialhandposition,andeyeposition(Figure6A). allneuralcorrelatesofvoluntaryeyemovementsshouldbecon-
WithinPMd,cellsarefoundthatcodethetargetofareachrela- sidered attention. For example, asserting that the oculomotor
576 Neuron63,September10,2009ª2009ElsevierInc.
Neuron
Review
neuronsinthebrainstemcontrolattentionratherthanmovethe suchrobustperformancecanbeconsideredstrongsupportthat
eyes would not be a useful construct. In this review, we have intentionsignalsexistintheseareasandcanbeharnessedby
adheredtothemoreclassicaldefinitionofattentionassensory the individual for controlling devices. Besides this proof-of-
selection for further processing. Of course, attention interacts concept, there are advantages to using these more high-level
withmovementplanningsinceattentionisdirectedtolocations and abstract intention signals, discussed below. The field of
of planned movements (Rizzolatti et al., 1987; Deubel and neural prosthetics israpidly evolving,anditisnotclearatthis
Schneider, 1996; Kowler et al., 1996). Attention may influence point which cortical and subcortical areas, or combination of
theinputstodecisionprocessesandmakeuseofforwardstate areas,willbethemostoptimalforparticulartypesofneuralinjury
estimations for prediction of where to reallocate attention to andparalysis.
compensate for eye and hand movements. The study of how GoalDecoding
attention interacts with other neural processing systems is The motor cortex provides control signals for the movement
amostimportantendeavor.However,wethinkthatover-gener- trajectory of the limbs. As such, motor-prosthetics have used
alizingattentiontoencompassalargevarietyofdifferentneural this activity to guide the trajectory of a cursor or a robot limb
processes weakens the concept and undercuts the ability to (Schwartz,2004;Vellisteetal.,2008;Nicolelis,2003;Donoghue,
developarobustunderstandingofothercognitivefunctions. 2002;Hochbergetal.,2006).However,toachieveagoalusing
SomestudieshaveproposedthatLIPinparticularhasthesole thismethodologytypicallytakesasecondormore.Thislength
purposeofcontrollingattention(Goldbergetal.,2002,2006;Bis- of time is required because the cursor (or robot limb) must be
ley and Goldberg, 2003). However, LIP has been reported to guided incrementally to the goal. Although intermediate steps
haveactivityrelatedtoavarietyoffunctions,includingtherepre- along the trajectory can be decoded quickly (Velliste et al.,
sentationofvalue,planningofeyemovements,prediction,cate- 2008), the attainment of the final goal requires considerable
gorization, cognitive set, shape recognition, decision making, time. On the other hand, the PPC and PMd provide signals
and timing (Platt and Glimcher, 1999; Snyder et al., 1997; relatedtothefinalgoalofthemovementratherthanthesteps
FreedmanandAssad,2006;ShadlenandNewsome,2001;Mai- togetthere.Thisgoal-related activityhasbeenusedin‘‘brain
monandAssad,2006;EskandarandAssad,1999;Janssenand control’’experimentstopositionacursordirectlyattheintended
Shadlen,2005;SerenoandMaunsell,1998;StoetandSnyder, location.Moreover,ithasbeenshownthatthesegoalscanbe
2004).Morerecently,proponentsofaprimarilyattentionalrole decoded in 100 ms (Musallam et al., 2004). Thus, in principle
forLIPhaveproposedthatLIPstillcontrolsattentionbutthatit many goals can be decoded in sequence, not unlike typing,
uses these various other functions to train the attentional and would have obvious advantages for communication and
controller(Gottliebetal.,2009).Anotherrecentproposalisthat other applications that require a high throughput of control
ofaprioritymap,whichdoesnotcontrolattentionoreyemove- signals. To this point, sequential goals have been decoded in
mentsbut highlights areas of interest that can be used by the braincontroltasksusingPMdactivity(Santhanametal.,2006).
oculomotororattentionsystems(Ipataetal.,2009).Thislatter DecisionVariableDecoding
ideafitsmorecloselywithourproposalthattheregionisgener- As reviewed above, the areas that code movement intentions
allyimportantasaninterfacebetweensensoryandmotorareas could be within the decision-making network since they show
for sensorimotor transformations, and its functions are neither activity for potential plans and outcomes of the decision. One
strictlysensorynormotor(Andersenetal.,1987;BuneoandAn- hallmark of thisinvolvement in decision makingfor the PPC is
dersen,2006;Ipataetal.,2009).Theprioritymapconceptdoes thecodingofexpectedvalueinLIP(PlattandGlimcher,1999).
accountforsomeaspectsofattentionandmovementplanning, TotestwhetherPRRalsoencodestheexpectedvalueforan
althoughtherearecaseswhereLIPandPRRdonotindicatethe action, the effect of expected reward on its activity was
locusofattention(Snyderetal.,1997;CuiandAndersen,2007; measured (Musallam et al., 2004). It was found that type of
GailandAndersen,2006;Baldaufetal.,2008).Alsothisideaof reward (juiceversuswater), amountof reward,and probability
aprioritymapdoesnotcapturetheintricaciesandessenceof ofrewardallstronglymodulatedPRRcellactivitywithincreased
forwardstateestimation,decisionmaking,orcoordinatetrans- responsiveness for the expectation of the preferred reward
formationsthatarecentralelementsofthisreview. (Figures 7A–7C). To test whether this expected value signal
could be used for practical neuroprosthetics applications,
AMedicalApplication:DecodingIntentionSignals decodings were performed for both ‘‘brain control’’ and reach
CognitiveNeuralProsthetics tasks.Itwasfoundthattheexpectationoftheanimalcouldbe
Arelativelynewandacceleratingfieldofresearchisneuralpros- decodedforbothtypesoftask(Figures7Dand7E).Moreover,
thetics.Thegoalofthisresearchistodecodeintentionsignalsin expectedvalue(preferredversusnonpreferredfortype,magni-
patients with movement disorders such as paralysis and use tude, or probability) and reach goals could be simultaneously
thesesignalstocontrolexternalassistivedevices.Mostefforts decoded with a mean accuracy of 51.2% ± 8.0% (mean, SD;
haveconcentratedonthemotorcortexforobtainingmovement chance12.5%).
intention signals (Donoghue, 2002; Schwartz, 2004; Nicolelis, This finding, that sucha high-level signal as expected value
2003;KennedyandBakay,2000;Moritzetal.,2008).However, canbedecodedinbraincontroltrials,isveryimportantforthe
some recent studies have focused on intention signals in pre- concept of cognitive prosthetics. There are some potential
motor and parietal cortex for neuroprosthetic applications practical advantages of decoding the expected value, since
(Musallam et al., 2004; Mulliken et al., 2008b; Santhanam it provides information about the preferences and mood of
etal.,2006).Thefactthattheseareas,particularlyPPC,provide the patient. After all, the first thing a doctor asks is ‘‘how are
Neuron63,September10,2009ª2009ElsevierInc. 577