higher level cognition: what’s missingski.clps.brown.edu/cogsim/cogsim.13exec.pdf · higher level...
TRANSCRIPT
Higher
Lev
elCognitio
n:What’s
Missin
g
Higher
Lev
elCognitio
n:What’s
Missin
g
•Plan
ning,p
roblem
solving,reaso
ning,co
mplex
decisio
n-m
aking
•What
doall
ofthese
hav
ein
common?
•Top-downcontro
lofbeh
avior:Instead
ofreactin
gin
abotto
m-up
fashionto
stimuli,b
ehav
iorisdriv
en(co
ntro
lled)byan
actively
main
tained
represen
tationofwhat
weare
supposed
tobedoing...
•Allo
wsusto
beh
avein
contex
tually
appropriate
fashioninstead
ofjust
givingthestro
ngest,m
ost
dominan
tresp
onse
•Also
gives
ustheab
ilityto
linkev
ents
across
timepoints,
andto
carry
outbeh
aviors
that
areexten
ded
across
time...
Whydoes
thishap
pen
?
A-not-B
Kai
A-not-B
Max
Card
Sort
Higher
Lev
elCognitio
n:What
WeKnow
Frontal
(andBG)dam
ageim
pairs
plan
ning,reaso
ning,d
ecision-m
aking,
self-initiated
actions,self-aw
areness,so
cialinteractio
n...
TheRan
geofFrontal
Functio
ns
Activation-based
working
mem
ory
Activ
ation-based
WorkingMem
ory
Monkey
electrophysio
logy
Cue
Delay
Response
ThePFCcan
main
taininform
ation(neu
ralfirin
g)over
time(activ
ation-based
mem
ory).
→Thiscan
beused
toguideatten
tionin
posterio
rreg
ions
(“guided
activatio
n”or“b
iasedcompetitio
n”).
Top-downvsbotto
m-upPFC
Busch
man
&Miller,
2007,Scien
ce
Top-downvsbotto
m-upPFC
Busch
man
&Miller,
2007,Scien
ce
•parietal
actfortarg
etlocatio
npreced
espfc
actforpop-out
•pfc
actpreced
esparietal
forsearch
Top-downvsbotto
m-upPFC
Busch
man
&Miller,
2007,Scien
ce
•greater
low
freqpfc-p
arietalsynch
ronizatio
nfortop-down
•greater
highfreq
synch
ronyforbotto
m-up
TheRan
geofFrontal
Functio
ns
Activation-based
working
mem
ory
InhibitionStro
op:Diffi
culty
inhibitin
gprep
oten
tresp
onse.
TheStro
opTask
TheStro
opTask
Pink
TheStro
opTask
Pink
Yello
w
TheStro
opTask
Pink
Yello
w
Green
TheStro
opTask
Pink
Yello
w
Green
Red
TheRan
geofFrontal
Functio
ns
Activation-based
working
mem
oryMonkey
electrophysio
logy.
InhibitionStro
op:Diffi
culty
inhibitin
gprep
oten
tresp
onse.
TheRan
geofFrontal
Functio
ns
Activation-based
working
mem
oryMonkey
electrophysio
logy.
InhibitionStro
op:Diffi
culty
inhibitin
gprep
oten
tresp
onse.
Flexibility
Contin
uewith
sameresp
onse
aftertask
chan
ges, p
erseveratio
n.
TheRan
geofFrontal
Functio
ns
Activation-based
working
mem
oryMonkey
electrophysio
logy.
InhibitionStro
op:Diffi
culty
inhibitin
gprep
oten
tresp
onse.
Flexibility
Contin
uewith
sameresp
onse
aftertask
chan
ges, p
erseveratio
n.
Fluency
Diffi
culty
gen
eratingvariety
ofresp
onses.
TheRan
geofFrontal
Functio
ns
Activation-based
working
mem
oryMonkey
electrophysio
logy.
InhibitionStro
op:Diffi
culty
inhibitin
gprep
oten
tresp
onse.
Flexibility
Contin
uewith
sameresp
onse
aftertask
chan
ges, p
erseveratio
n.
Fluency
Diffi
culty
gen
eratingvariety
ofresp
onses.
Executive
controlProbsw/goal-d
irectedplan
ning,coordinatin
g.
TheRan
geofFrontal
Functio
ns
Activation-based
working
mem
oryMonkey
electrophysio
logy.
InhibitionStro
op:Diffi
culty
inhibitin
gprep
oten
tresp
onse.
Flexibility
Contin
uewith
sameresp
onse
aftertask
chan
ges, p
erseveratio
n.
Fluency
Diffi
culty
gen
eratingvariety
ofresp
onses.
Executive
controlProbsw/goal-d
irectedplan
ning,coordinatin
g.
(shopping)
Monitoring/evaluation
e.g.,Erro
r-monito
ring.
TheRan
geofFrontal
Functio
ns
Activation-based
working
mem
oryMonkey
electrophysio
logy.
InhibitionStro
op:Diffi
culty
inhibitin
gprep
oten
tresp
onse.
Flexibility
Contin
uewith
sameresp
onse
aftertask
chan
ges, p
erseveratio
n.
Fluency
Diffi
culty
gen
eratingvariety
ofresp
onses.
Executive
controlProbsw/goal-d
irectedplan
ning,coordinatin
g.
(shopping)
Monitoring/evaluation
e.g.,Erro
r-monito
ring.
Stro
opEffect:
GREEN
Possib
leexplan
ation:differentialpathw
aystrength
•tw
opath
way
s:word
readingan
dcolornam
ing
•These
compete
togen
erateresp
onse
•Word
readingpath
way
ismuch
stronger
than
colornam
ing
•When
word
iden
tityinform
ationdoesn
’tmatch
color,itinterferes
stronglywith
colornam
ing
Stro
opEffect:
GREEN
Possib
leexplan
ation:differentialpathw
aystrength
•tw
opath
way
s:word
readingan
dcolornam
ing
•These
compete
togen
erateresp
onse
•Word
readingpath
way
ismuch
stronger
than
colornam
ing
•When
word
iden
tityinform
ationdoesn
’tmatch
color,itinterferes
stronglywith
colornam
ing
•Becau
secolorpath
way
isrelativ
elyweak
,incongruen
tcolorinfo
does
notinterfere
with
word
reading
Stro
opEffect:
GREEN
•Puzzle:
Ifthecolornam
ingpath
way
isweak
erthan
word
reading,h
ow
doweman
ageto
nam
ecoloroftheword
“green
”ab
ove?
•Solutio
n:Prefro
ntal
cortex
actively
main
tainsarep
resentatio
nofthe
taskthat
youare
supposed
tobedoing(co
lornam
ingorword
reading)
•Thisactiv
elymain
tained
taskrep
resentatio
nbiases
processingin
posterio
rcortex
byactiv
atingunits
inap
propriate
path
way
•e.g
.,colornam
ingtask
repin
PFCsen
dsactiv
ationto
theunits
incolor
nam
ingpath
way...
Stro
opModel
TheStro
opTask
:Model
Data
Contro
l:Red
.Conflict:
Red
.Congruen
t:Red
.
Control
Conflict
Congruent
Condition
500
600
700
800
900
Reaction Time (msec)
Stroop D
ata (Dunbar &
MacLeod, 84)
Color
Word
110−
115−
120−
125−
130−
135−
01
2
Color N
aming
Word R
eading
Conflict
Congruent
Control
Path
way
Stren
gth
vsProcessin
gSpeed
Theo
ries
•Model:
thekey
differen
cebetw
eenword
readingan
dcolornam
ingis
pathway
strength(read
ing>
colornam
ing).Thisresu
ltsin:
–Word
readingbein
gfaster
than
colornam
ing
–Asymmetric
interferen
ceeffects
•Other
(verb
al)theo
riesposit
Stro
opeffects
resultin
gfro
ma
horserace
rather
than
directcom
petition
Horse
Race
Theo
ries
•Colordoes
notaffect
word
readingbecau
setheword
readingprocess
runsto
completio
nbefo
recolorisprocessed
•Conversely,
word
iden
titydoes
affectcolornam
ingbecau
seword
readingprocess
completes
befo
recolorresp
onse
isgen
erated
•Thistheo
ry,statedas
such
,implies
that
itsh
ould
bepossib
leto
get
colorto
interfere
with
word
info
ifthecolornam
ingprocess
has
ahead
start
Horse
Race
Theo
ries
•Colordoes
notaffect
word
readingbecau
setheword
readingprocess
runsto
completio
nbefo
recolorisprocessed
•Conversely,
word
iden
titydoes
affectcolornam
ingbecau
seword
readingprocess
completes
befo
recolorresp
onse
isgen
erated
•Thistheo
ry,statedas
such
,implies
that
itsh
ould
bepossib
leto
get
colorto
interfere
with
word
info
ifthecolornam
ingprocess
has
ahead
start
Stro
opAcco
unts:
NotaHorse
Race
Stro
opAcco
unts:
NotaHorse
Race
⇒GREEN
Stro
opAcco
unts:
NotaHorse
Race
Stro
opAcco
unts:
NotaHorse
Race
Stro
opAcco
unts:
NotaHorse
Race
−400
−200
0200
400S
OA
400
450
500
550
600
650
700
Reaction Time (msec)
Stroop S
OA
(Glaser &
Glaser, 82)
Color C
onfC
olor Cong
Word C
onfW
ord Cong
80−
100−
120−
140−
160−
180−
200−
01
23
45
67
89
10
Color N
aming
Word R
eading
Conflict
Congruent
0−
8−
168
16S
OA
(cycles)
Stro
opAcco
unts:
Automaticity
•Early
accountsofStro
opfocu
sedonau
tomatic
vscontro
lledprocessin
g
•Acco
rdingto
these
theo
ries,word
readingis
automatic
andcolor
nam
ingisa
controlledprocess.
•Automatic
processes
don’tsu
fferfro
minterferen
ce(th
eyproceed
“automatically
”)butcontro
lledprocesses
do.
•categ
orical
distin
ction
Stro
opAcco
unts:
Automaticity
•Statu
s(w
heth
eraprocess
iscontro
lledorau
tomatic)
dep
endson
absolutepath
way
strength
Stro
opAcco
unts:
Automaticity
•Statu
s(w
heth
eraprocess
iscontro
lledorau
tomatic)
dep
endson
absolutepath
way
strength
•In
contrast,m
odel
focu
seson
relativepath
way
strength
-stro
nger
path
way
interferes
with
lessstro
ngpath
way
(butnotvice-v
ersa)
•Pred
iction:Ifwecould
comeupwith
atask
that
iseven
lesswell-learn
edthan
colornam
ing,w
ewill
findthat:
•New
taskwill
hav
enoeffect
oncolornam
ing
•Colornam
ingwill
interfere
with
new
task
Stro
opAcco
unts:
Automaticity
•Statu
s(w
heth
eraprocess
iscontro
lledorau
tomatic)
dep
endson
absolutepath
way
strength
•In
contrast,m
odel
focu
seson
relativepath
way
strength
-stro
nger
path
way
interferes
with
lessstro
ngpath
way
(butnotvice-v
ersa)
•Pred
iction:Ifwecould
comeupwith
atask
that
iseven
lesswell-learn
edthan
colornam
ing,w
ewill
findthat:
•New
taskwill
hav
enoeffect
oncolornam
ing
•Colornam
ingwill
interfere
with
new
task
•Asnew
taskispracticed
repeated
ly,effectssh
ould
reverse
Stro
opAcco
unts:
Contin
uum,n
otaDich
otomy
Stro
opAcco
unts:
Contin
uum,n
otaDich
otomy
"blue"
"red"
Shap
e-nam
ingfirst
likecolor-n
amingin
standard
Stro
op,
then
like-w
ord
reading.
Stro
opAcco
unts:
Contin
uum,n
otaDich
otomy
MacL
eod&
Dunbar,1988
TheStro
optask
model
...dem
onstrates
therole
ofatopdowninfluen
ceon
activatio
n-based
processin
gmed
iatedbyfro
ntal
cortex
.What
Ithinkisthe
downsid
eofmodels
likethisisthat
theprefro
ntal
cortex
isrep
resented
bya
layer
ofjusttw
ounits,m
eantto
conden
seits
functio
n.Tome,th
eseunits
seemsortoflik
eprep
rogram
med
gran
dmother
cells,in
thesen
sethat
the
rules
andrep
resentatio
nsfor”co
lornam
ing”an
d”w
ord
reading”aren
’t
reallylearn
edbythenetw
ork.Theuser
ofthemodel
understan
dshow
these
units
areworking,b
utthey
seemsortoflik
ean
unnatu
rallysim
plistic
way
torep
laceabigger,
more
powerfu
lsystem
.I’m
wonderin
gifthe
two-unitPFClay
erofthisnetw
ork
could
berep
lacedbyahidden
layer
(or
perh
apssometh
ingelse)
sothat
thecn
andwrtask
scould
first
belearn
ed
bythePFCan
dthen
applied
totheStro
optask
.Thiswould
mak
ethePFC
inthemodel
more
flexible
tolearn
new
rules
(ifsay
theinputhad
another
dim
ensio
nbesid
escolororword–size
orfontcould
beexam
ples).
Essen
tiallywhat
I’mask
ingis,can
thePFCin
thismodel
berep
lacedwith
a
layer(s)
that
would
mak
ethenetw
ork
better
able
tohan
dle
more
rules
and
tasksgiven
aset
ofinputs?
Buthow
doPFCunits
cometo
represen
ttask
rules??
•Stro
opmodel
isanice
simple
accountofPFCfunctio
n,butitsomeh
ow
assumes
that
PFC’knows’how
tomain
tainarule
forcolornam
ingan
d
tomag
icallybias
color-n
aminghidden
units
•Interestin
gquestio
nishow
these
rule-lik
erep
resentatio
nsdev
elopin
thefirst
place?
•Can
PFClearn
toassig
nab
stractrule-lik
erep
resentatio
nsthat
codefor
stimulusdim
ensio
ns(e.g
.,color)byexperien
ce(w
ithmultip
lecolors)??
PFCSpecializatio
ns→
Rule-L
ikeAbstract
Rep
s(R
ougier,N
oelle,B
raver,
Cohen
&O’Reilly,P
NAS)
Stim
uli
Netw
ork
color
size
shape
position
texture
Hidden (83 units)
Task Hidden (16 units)
Left Stim
ulusR
ight Stim
ulusT
ask
NF
MF
SF
LF
A1
A2
A3
A4
B1
B2
B3
B4
C1
C2
C3
C4
D1
D2
D3
D4
E1
E2
E3
E4
A1
A2
A3
A4
B1
B2
B3
B4
C1
C2
C3
C4
D1
D2
D3
D4
E1
E2
E3
E4A1
A2
A3
A4
B1
B2
B3
B4
C1
C2
C3
C4
D1
D2
D3
D4
E1
E2
E3
E4
No
AG
AB
CD
E
Dim
ension Cue
Response
Cue Hidden(16 units)
PFC Context(30 units)
Dev
elopingPFCRep
s(R
ougier,N
oelle,B
raver,
Cohen
&O’Reilly,P
NAS)
("Circle")
("Triangle")
("Square")
(blue, lg, circle..)(red, sm
, triangle..)(green, m
ed, square..)
tt+
1t+
2
Stim
ulus
Correct
Response
Netw
orkInput/T
arget
leftright
task
response
leftright
task
responseleft
righttask
response
Task
1=Nam
etheShap
e(orcolor,etc)
Dev
elopingPFCRep
s(R
ougier,N
oelle,B
raver,
Cohen
&O’Reilly,P
NAS)
("Circle")
("Triangle")
("Square")
(blue, lg, circle..)(red, sm
, triangle..)(green, m
ed, square..)
tt+
1t+
2
Stim
ulus
Correct
Response
Netw
orkInput/T
arget
leftright
task
response
leftright
task
responseleft
righttask
response
Task
1=Nam
etheShap
e(orcolor,etc)
Task
2:Dotw
ostim
ulimatch
alongsomedim
ensio
n?(yes/
no)
Task
3:W
hich
object
islarg
er?etc.
Dev
elopingPFCRep
s(R
ougier,N
oelle,B
raver,
Cohen
&O’Reilly,P
NAS)
("Circle")
("Triangle")
("Square")
(blue, lg, circle..)(red, sm
, triangle..)(green, m
ed, square..)
tt+
1t+
2
Stim
ulus
Correct
Response
Netw
orkInput/T
arget
leftright
task
response
leftright
taskresponse
leftright
task
response
Task
1=Nam
etheShap
e(orcolor,etc)
Task
2:Dotw
ostim
ulimatch
alongsomedim
ensio
n?(yes/
no)
Task
3:W
hich
object
islarg
er?etc.
Key
:Dorep
eatedtrials
ofsam
etask
–contin
uousatten
tionto
shap
es,etc
PFCSpecializatio
ns→
Rule-L
ikeAbstract
Rep
s(R
ougier,N
oelle,B
raver,
Cohen
&O’Reilly,P
NAS)
Weig
hts
from
PFCorHidden
tooutputresp
onse
units
Stim
uli
Posterio
rFullPFC
color
size
shape
position
texture
Rule
=Onestim
ulusdim
ensio
n(ro
w)relev
antat
atim
e.(e.g
.,card-so
rtingtask
s)
Abstractio
nderiv
esfro
msu
stained
main
tenan
ceover
trials!
PFCSpecializatio
ns→
Rule-L
ikeAbstract
Rep
s(R
ougier,N
oelle,B
raver,
Cohen
&O’Reilly, P
NAS)
Weig
hts
from
PFCorHidden
tooutputresp
onse
units
Stim
uli
Posterio
rFullPFC
color
size
shape
position
texture
Rule
=Onestim
ulusdim
ensio
n(ro
w)relev
antat
atim
e.(e.g
.,card-so
rtingtask
s)
Abstractio
nderiv
esfro
msu
stained
main
tenan
ceover
trials!
Posterio
rnet
’mem
orizes’
specifi
ccombinatio
nsoffeatu
res/resp
onses
foreach
task,
doesn
’tdev
elopsystem
aticrep
resentatio
ns
PFCSpecializatio
ns→
Rule-L
ikeAbstract
Rep
s(R
ougier,N
oelle,B
raver,
Cohen
&O’Reilly,P
NAS)
Stim
uli
PFCNoGate
FullPFC
color
size
shape
position
texture
Adap
tivegatin
giskey
:
PFCSpecializatio
ns→
Rule-L
ikeAbstract
Rep
s(R
ougier,N
oelle,B
raver,
Cohen
&O’Reilly,P
NAS)
Stim
uli
PFCNoGate
FullPFC
color
size
shape
position
texture
Adap
tivegatin
giskey
:
with
inblock
oftrials
feature
chan
ges
butgatin
gmech
learnsto
main
tain
constan
tPFCrep
(until
rule
switch
es,perfo
rman
cegoes
down→
update)
PFCSpecializatio
ns→
Rule-L
ikeAbstract
Rep
s(R
ougier,N
oelle,B
raver,
Cohen
&O’Reilly,P
NAS)
Stim
uli
FullPFC,T
askPairs
FullPFC,A
ll4Task
s
color
size
shape
position
texture
Asisbread
thofexperien
ce(sam
estim
uliacro
ssdifferen
ttask
s)
PFCSpecializatio
ns→
Rule-L
ikeAbstract
Rep
s(R
ougier,N
oelle,B
raver,
Cohen
&O’Reilly,P
NAS)
Stim
uli
FullPFC,T
askPairs
FullPFC,A
ll4Task
s
color
size
shape
position
texture
Asisbread
thofexperien
ce(sam
estim
uliacro
ssdifferen
ttask
s)
(increasin
gpressu
reto
use
samepfc
repsacro
sstask
s→
system
aticity);
with
small
#task
scan
get
bywith
mem
orizin
g)
Rule-L
ikeAbstract
Rep
s→
Gen
eralization
PosteriorP+RecP+Self
SRNSR
N−PFCN
oGate
Full
No P
FC
PF
C
25 50 75
Generalization % Correct
Task P
airsA
ll Tasks
Cross−T
ask Generalization
Training R
egimen
Netw
ork Configuration
FullP
FC
−P
air
0.250.50
0.75R
ule Representation M
easure
25 50 75
Generalization % Correct
Posterior−
All
NoG
ate−A
ll FullP
FC
−A
ll
Rule−Likeness P
redicts Generalization
r=.97
Abstractio
n→
better
gen
eralizationacro
sstask
s(accu
racyonstim
snotseen
inparticu
lartask
).
Rule-L
ikeAbstract
Rep
s→
Gen
eralization
PosteriorP+RecP+Self
SRNSR
N−PFCN
oGate
Full
No P
FC
PF
C
25 50 75
Generalization % Correct
Task P
airsA
ll Tasks
Cross−T
ask Generalization
Training R
egimen
Netw
ork Configuration
FullP
FC
−P
air
0.250.50
0.75R
ule Representation M
easure
25 50 75
Generalization % Correct
Posterior−
All
NoG
ate−A
ll FullP
FC
−A
ll
Rule−Likeness P
redicts Generalization
r=.97
Abstractio
n→
better
gen
eralizationacro
sstask
s(accu
racyonstim
snotseen
inparticu
lartask
).
Interactio
nofnatu
re(PFCmech
anism
s)an
dnurtu
re(bread
thof
experien
ce).
Stro
opPerfo
rman
ce(R
ougier
etal,P
NAS)
Stro
opPerfo
rman
ce(R
ougier
etal,P
NAS)
Neutral
Conflict
500
600
700
800
900
Reaction Time (msec)
People: C
olorP
eople: Word
Model: Lo F
reqM
odel: Hi F
req
Stroop T
ask
Neutral
Conflict
50
100
150
200
Reaction Time (s)
People: LF
Color
People: LF
Word
People: C
trl Color
People: C
trl Word
Model: LF
Clr
Model: LF
Wrd
Model: C
trl Clr
Model: C
trl Word
Stroop T
ask: Lesion Data
Sam
enetw
ork
¶m
eters:PFCcontro
lrep
resentatio
nsdev
eloped
entirely
throughlearn
ingfro
mran
dom
initial
weig
hts!
Stro
opPerfo
rman
ce(R
ougier
etal,P
NAS)
Neutral
Conflict
500
600
700
800
900
Reaction Time (msec)
People: C
olorP
eople: Word
Model: Lo F
reqM
odel: Hi F
req
Stroop T
ask
Neutral
Conflict
50
100
150
200
Reaction Time (s)
People: LF
Color
People: LF
Word
People: C
trl Color
People: C
trl Word
Model: LF
Clr
Model: LF
Wrd
Model: C
trl Clr
Model: C
trl Word
Stroop T
ask: Lesion Data
Sam
enetw
ork
¶m
eters:PFCcontro
lrep
resentatio
nsdev
eloped
entirely
throughlearn
ingfro
mran
dom
initial
weig
hts!
LF=left
frontal
(DLPFC)lesio
nsin
peo
ple
andmodel
(posttrain
ing,30%
dam
age)
TheRan
geofFrontal
Functio
ns
Activation-based
working
mem
oryMonkey
electrophysio
logy.
InhibitionStro
op:Diffi
culty
inhibitin
gprep
oten
tresp
onse.
Flexibility
Contin
uewith
sameresp
onse
aftertask
chan
ges, p
erseveratio
n.
Fluency
Diffi
culty
gen
eratingvariety
ofresp
onses.
Executive
controlProbsw/goal-d
irectedplan
ning,coordinatin
g.
Monitoring/evaluation
e.g.,Erro
r-monito
ring.
Dynam
icCateg
orizatio
nTask
s
Wisco
nsin
Card
Sort
AB
CD
Dynam
icCateg
orizatio
nTask
s
Wisco
nsin
Card
Sort
AB
CD
Dynam
icCateg
orizatio
nTask
s
Wisco
nsin
Card
Sort
AB
CD
Dynam
icCateg
orizatio
nTask
s
Wisco
nsin
Card
Sort
AB
CD
Dynam
icCateg
orizatio
nTask
s
Wisco
nsin
Card
Sort
AB
CD
Dynam
icCateg
orizatio
nTask
s
Wisco
nsin
Card
Sort
AB
CD
Dynam
icCateg
orizatio
nTask
s
Wisco
nsin
Card
Sort
AB
CD
Experim
ental
task(lik
eStro
op),b
utcap
tures
someessen
tialasp
ectsof
higher
level
cognitio
n.
Dynam
icCateg
orizatio
nTask
s
Wisco
nsin
Card
Sort
AB
CD
Experim
ental
task(lik
eStro
op),b
utcap
tures
someessen
tialasp
ectsof
higher
level
cognitio
n.
Frontal
patien
tspersev
eratewith
thefirst
rule.
→weig
ht-b
asedten
den
ciesbuild
upwhen
categorizin
gacco
rdingto
first
rule,an
dyouneed
toactiv
elymain
tainthenew
rule
tocounteract
these
weig
ht-b
asedten
den
cies
”Frontal
Task
s”
•Stro
op:Ability
tooverrid
eprep
oten
tresp
onse
(word
reading)in
favor
ofcu
rrently
relevan
ttask
(colornam
ing)–req
uires
top-downcontro
l.
•Activ
ationbased
directin
gofatten
tion.
”Frontal
Task
s”
•Stro
op:Ability
tooverrid
eprep
oten
tresp
onse
(word
reading)in
favor
ofcu
rrently
relevan
ttask
(colornam
ing)–req
uires
top-downcontro
l.
•Activ
ationbased
directin
gofatten
tion.
•”P
refrontal
contro
l”notjustforoverrid
inglongterm
associatio
nslik
e
word
reading,b
utalso
fortheab
ilityto
quick
lych
angeatten
tionin
an
onlin
efash
ionin
response
toch
angingtask
dem
ands:UPDATIN
G.
Rev
engeoftheDonuts...
Rev
engeoftheDonuts...
Twostrateg
iesforsolvingdonutcateg
orizatio
ntask
:
•Adjustweig
htsto
differen
tdonuttypes
•Activ
elymain
tainarep
resentatio
nofyourcu
rrentstrateg
y;d
eactivate
thisrep
andactiv
atean
other
ifyouget
neg
ativefeed
back
•Activ
emain
tenan
cedoes
notstro
ngly
ben
efitinitial
learningoftherule
•Howev
eritdoes
greatly
facilitateperfo
rman
cewhen
therule
switch
es
WCSTin
PFCmodel
(Rougier
etal)
Weig
ht-an
dActiv
ation-Based
Mem
ory
Interactio
ns
A-not-B
task
•Persev
erativesearch
ingat
A–also
seenin
patien
tswith
PFCdam
age
•Better
pefo
rman
cein
gaze/
expectatio
n
•Inhibitio
nproblem
?
•Model
dem
onstrates
main
tenan
ceproblem
.
•Sam
emodel
accounts
forvario
useffects
indifferen
tversio
nsof
A-not-B
tasknotexplain
edbyan
yother
unified
theo
ry(M
unak
ata,
1998).
A-not-B
Model
AB
C
AB
CH
idden
AB
CA
BC
Reach
C1
C2
T1
T2
LocationC
overT
oy
Gaze/
Expectation
Knowled
ge-actio
ndisso
ciationsin
card-so
rttask
•Kidscan
tellyouwhere
trucksgoin
thesh
apegam
e,even
aftersortin
g
accordingto
color!
•Butifyouask
“where
dored
trucksgoin
thesh
apegam
e”they
still
fail!(M
orto
n&
Munak
ata,2002)
•Explain
edbydifferen
tlev
elsofconflict
experien
cedwhen
facedwith
multip
lestim
uli-resp
onse
associatio
ns..
Card
Sortin
gTask
s
•Relev
antto
every
day
life,orjustto
thispecu
liartask
?
Card
Sortin
gTask
s
•Relev
antto
every
day
life,orjustto
thispecu
liartask
?
•Goodmeasu
reofonlin
ethinking&
problem
solving:Theab
ilityto
flexibly
consid
erdifferen
tpossib
ilitiesto
guidethinkingan
dbeh
avior.
Card
Sortin
gTask
s
•In
what
situatio
nsdoweneed
toto
consid
er/rep
resentdifferen
trules
inmindan
dhav
etheab
ilityto
flexibly
update/
main
tainthem
until
oneworkswell?
Card
Sortin
gTask
s
•In
what
situatio
nsdoweneed
toto
consid
er/rep
resentdifferen
trules
inmindan
dhav
etheab
ilityto
flexibly
update/
main
tainthem
until
oneworkswell?
•Rightnow!Thinking.I’m
askingyouaquestio
n,y
ouconsid
eran
alternativ
e(e.g
.,“N
ever:
cardsortin
gtask
sare
dumb”).
Card
Sortin
gTask
s
•In
what
situatio
nsdoweneed
toto
consid
er/rep
resentdifferen
trules
inmindan
dhav
etheab
ilityto
flexibly
update/
main
tainthem
until
oneworkswell?
•Rightnow!Thinking.I’m
askingyouaquestio
n,y
ouconsid
eran
alternativ
e(e.g
.,“N
ever:
cardsortin
gtask
sare
dumb”).
•Youthen
evalu
atethequality
ofwhat
you’re
holdingin
mind:does
it
mak
esen
se,isitlik
elyto
produce
agoodoutco
me?
•Ifyes,m
aintain
info
furth
erprocessin
g;ifnot,u
pdate.
Card
Sortin
gTask
s
•In
what
situatio
nsdoweneed
toto
consid
er/rep
resentdifferen
trules
inmindan
dhav
etheab
ilityto
flexibly
update/
main
tainthem
until
oneworkswell?
•Rightnow!Thinking.I’m
askingyouaquestio
n,y
ouconsid
eran
alternativ
e(e.g
.,“N
ever:
cardsortin
gtask
sare
dumb”).
•Youthen
evalu
atethequality
ofwhat
you’re
holdingin
mind:does
it
mak
esen
se,isitlik
elyto
produce
agoodoutco
me?
•Ifyes,m
aintain
info
furth
erprocessin
g;ifnot,u
pdate.
•Scien
ce:hypothesis
form
ulatio
nfro
mexperim
ental
data.
Dynam
icCateg
orizatio
nTask
s:ID
/ED
task
Dynam
icCateg
orizatio
nTask
s:ID
/ED
task
Sam
e(R
eversal)D
ifferent(S
hift)
IDR
ED
R
IDS
ED
S
Intra−dim
ensional(ID
)
Extra−
dimensional
(ED
)
Target
change
Stim
uli
Rule C
hange
Previous S
timuli
ID/ED
andFrontal
Dam
age
(Dias,R
obbins&
Roberts
(1997),JNeu
rosci)
IDS
IDR
ED
S
10 20 30 40 50
Errors to Criterion
Perseverations from
Frontal Lesions
Control
Orbital
Lateral
Previous S
timuli
Orig
inal
interp
retation:Orbital
=affectiv
einhibitio
n,
Lateral
=atten
tional
selection.
Altern
ativeAcco
unt
(O’Reilly,N
oelle,
Brav
er&
Cohen
(2002),Cereb
ralCortex
)
Orbital
PFCrep
resents
detailed
features.
Lateral
PFCrep
resents
abstract
dim
ensio
ns.
Altern
ativeAcco
unt
(O’Reilly,N
oelle,
Brav
er&
Cohen
(2002),Cereb
ralCortex
)
Orbital
PFCrep
resents
detailed
features.
Lateral
PFCrep
resents
abstract
dim
ensio
ns.
Activ
ation-based
PFCprocessin
gfacilitates
rule
switch
:
Orbital
=sw
itchto
new
features
(IDR).
Lateral
=sw
itchto
new
dim
ensio
n(EDS).
Altern
ativeAcco
unt
(O’Reilly,N
oelle,
Brav
er&
Cohen
(2002),Cereb
ralCortex
)
Orbital
PFCrep
resents
detailed
features.
Lateral
PFCrep
resents
abstract
dim
ensio
ns.
Activ
ation-based
PFCprocessin
gfacilitates
rule
switch
:
Orbital
=sw
itchto
new
features
(IDR).
Lateral
=sw
itchto
new
dim
ensio
n(EDS).
Persev
erations=weig
ht-b
asedprocessin
gin
absen
ceofPFC.
ID/ED
Model
Input
Hidden
PF
C_F
eat
Output P
FC
_Dim
Dim
AD
im B
LR
LR
Dim
AD
im B
Dim
AD
im B
LR
Gate (D
A)
Twodim
ensio
ns,A
andB(sh
apes
&lin
es)
Oneach
trial,fourstim
sare
presen
ted:
Dim
Aleft,D
imA
right,D
imBleft,D
imBrig
ht...
ID/ED
Model
shapeline
LR
LL
RR
Output
Input
Posterior
Cortex
shapeline
shapeline
VT
A
PF
C_F
eat
PF
C_D
im
ID/ED
Model
shapeline
LR
LL
RR
Output
Input
Posterior
Cortex
shapeline
shapeline
VT
A
PF
C_F
eat
PF
C_D
im
Activ
ationlim
itedin
cortex
:atten
tion.
PFCprovides
top-downbias,
with
DA/updatin
gunit.
PFC
Feat
=stim
features
indep
oflocatio
nPFC
Dim
=ab
stractdim
ensio
ns
PFCupdatin
gbased
onunexpected
reward
san
derro
rs
•When
there
isan
increase
inDA
activity
(e.g.themodel
gotthean
swer
rightbutwasn
’texpectin
garew
ard):
–hidden
unitactiv
ityisgated
into
PFC
–connectio
nsfro
mhidden
units
toDA
areincreased
–PFCserv
esto
amplify
theinfluen
ceofhidden
units
associated
with
correct
responding
•When
there
isadecrease
inDA
activity
(themodel
was
expectin
ga
reward
butgav
ethewrongresp
onse):
–PFCactiv
ityiswiped
clean
–connectio
nsfro
mhidden
units
toDA
aredecreased
•Also
,there
issome“g
atingnoise”:
trialan
derro
rsearch
Sim
ilarities/Differen
ceswith
Store-Ig
nore-R
ecall
•With
S-I-R
,themodel
had
togate
the“S
tore”
stimulusinto
PFC(an
d
carryitforw
ardin
time)
inorder
toresp
ondcorrectly
;S-I-R
canonly
be
solved
with
thehelp
ofactiv
emain
tenan
ce(w
orkingmem
ory)
•TheID
/ED
taskcan
besolved
with
outactiv
emain
tenan
ce;butPFC
canhelp
byfocu
singthemodel’s
attentio
nonusefu
lparts
oftheinput
butitisn
’tnecessary..
IDR,E
DSin
theModel
shapeline
LR
LL
RR
Output
Input
Posterior
Cortex
Orbital
PF
C
LateralP
FC
shapeline
shapeline
VT
Ashape
line
LR
LL
RR
Output
Orbital
PF
C
LateralP
FC
shapeline
shapeline
VT
Ashape
line
LR
LL
RR
Output
Orbital
PF
C
LateralP
FC
shapeline
shapeline
VT
A
a) Initialb) ID
Rc) E
DS
Model
Data
IDS
IDR
ED
S
10 20 30 40 50Errors to Criterion
Perseverations from
Frontal Lesions
Control
Orbital
Lateral
Previous S
timuli
IDS
IDR
ED
S0 2 4 6 8 10 12 14 16
Errors to Criterion
Perseverations in the M
odel
IntactF
eatD
im
Explan
ationofLesio
nData:
IDS
•Intrad
imen
sional
shift
(IDS):d
ifferentstim
ulipre
andpost-sh
ift;the
relevan
tdim
ensio
n(A
)stay
sthesam
e
•Noeffect
ofPFClesio
ns
•PFCisunnecessary
becau
sethere
arenostro
ng,in
appropriate
tenden
ciesto
overco
me(new
stimuli)
Model
Data
IDS
IDR
ED
S0 2 4 6 8 10 12 14 16
Errors to Criterion
Perseverations in the M
odel
IntactF
eatD
im
Explan
ationofLesio
nData:
IDR
•Intrad
imen
sional
reversal
(IDR):sam
estim
ulipre
andpost-sh
ift;
initially
A1=targ
et;afterthesh
iftA2=targ
et
•Perfo
rman
ceisim
paired
afterPFC
Feat
lesionsbutnotPFC
Dim
lesions
•It’s
clearwhyPFC
Dim
isnotim
portan
there:
Itinvolves
ash
iftof
attentio
nwith
inadim
ensio
n,notacro
ssdim
ensio
ns..
•How
does
PFC
Feat
help
perfo
rman
ce?
Explan
ationofLesio
nData:
IDR
•Befo
resh
ift,somehidden
units
learnto
gen
eratetheA1resp
onse
•After
shift,th
esehidden
units
pointto
thewrongresp
onse
•PFChelp
sthemodel
focu
sonother
hidden
units,w
hich
canthen
be
associated
with
thenew
response
•Thisway
themodel
avoidshav
ingto
fully
unlearn
theasso
ciation
betw
eentheorig
inal
hidden
units
andA1resp
onse
Model
Data
IDS
IDR
ED
S0 2 4 6 8 10 12 14 16
Errors to Criterion
Perseverations in the M
odel
IntactF
eatD
im
Explan
ationofLesio
nData:
EDS
•Extrad
imen
sional
shift
(EDS):d
ifferentstim
ulipre-
andpost-sh
ift;
initially
A1=targ
et;afterthesh
ift,B3=targ
et
•Perfo
rman
ceisim
paired
afterPFC
Dim
lesionsbutnotPFC
Feat
lesions
•It’s
clearwhyPFC
Dim
isim
portan
t:Ithelp
sfocu
satten
tiononthe
new
lyrelev
antdim
ensio
n
•Whycan
’tPFC
Feat
servethesam
efunctio
n?sh
ould
beab
leto
bias
new
repsas
befo
re...
Explan
ationofLesio
nData:
EDS
•Extrad
imen
sional
shift
(EDS):d
ifferentstim
ulipre-
andpost-sh
ift;initially
A1=targ
et;afterthesh
ift,B3=targ
et
•Perfo
rman
ceisim
paired
afterPFC
Dim
lesionsbutnotPFC
Feat
lesions
•It’s
clearwhyPFC
Dim
isim
portan
t:Ithelp
sfocu
satten
tiononthe
new
lyrelev
antdim
ensio
n
•Whycan
’tPFC
Feat
servethesam
efunctio
n?sh
ould
beab
leto
bias
new
repsas
befo
re...
•With
outPFC
Dim
,PFC
Feat
has
nosen
seofwhat
constitu
tesa
“dim
ensio
n”,ju
stupdates
toran
dom
new
pattern
offeatu
resfro
mboth
Aan
dBdim
ensio
ns...
After
EDS:P
FC
Dim
lesion
Advan
tages
ofID
/ED
model
•PFCrep
sare
notclam
ped
asin
Stro
op–updated
inresp
onse
to
chan
gingtask
dem
ands.
•Nice
fitan
dexplan
ationofcomplex
monkey
data.
•Showshow
workingmem
ory
andcognitiv
econtro
lmay
betw
osid
es
ofthesam
ecoin:activ
ation-based
mem
ory
isnotjustmem
ory
butalso
biases
activity
elsewhere
inthebrain
.
•Showsthat
ID/ED
data
canbeexplain
edin
termsoflateral
andorbital
PFCcarry
ingoutthesam
efunctio
n(biasin
gcompetitio
nin
posterio
r
cortex
),ap
plied
todifferen
tkindsofconten
t(featu
resvsdim
ensio
ns)
•Also
provides
eviden
ceforahierarch
yofPFCrep
resentatio
ns
Lim
itationsofID
/ED
model
•Rep
snotclam
ped
,butstill
notlearn
ed–oneto
oneconnectiv
ityfro
m
HL.
Lim
itationsofID
/ED
model
•Rep
snotclam
ped
,butstill
notlearn
ed–oneto
oneconnectiv
ityfro
m
HL.
•Distin
ctionbetw
eenOFC=featu
res,DLPFC=dim
ensio
nsmay
betoo
conven
ient:observ
eddisso
ciation;n
otmuch
eviden
ceofOFC-featu
res
(seeFran
k&
Clau
s,2006).
•New
ermodels
address
theissu
eofhow
PFCrep
resentatio
nscan
dev
elopin
child
hoodan
dlead
tohigher
level
abstractio
nan
d
gen
eralizationto
new
tasks(Rougier
etal,2005,P
NAS)
Lim
itationsofID
/ED
model
•Doesn
’tdistin
guish
b/w
updatin
gan
dmain
tenan
cesystem
s.
•Goal/
Subgoal
requires
selectiveupdatin
gwith
concu
rrent
main
tenan
ceoftask
relevan
tinfo.
•Updatin
gsystem
thoughtto
involvetheBG
andDA,d
amag
edin
PD,SZan
dlead
to”fro
ntal-lik
e”im
pairm
ents
inStro
op,WCST,etc.
•New
erBG
models
address
these
issues
inmore
complex
tasks(eg
.
O’Reilly
&Fran
k,2006)
Goal/
Subgoal
Hierarch
icalStru
cture
1.Open
fridge.
2.Get
fooditem
s.
3.Close
fridge.
4.Get
bread
from
cupboard
Update
these
subgoals
toguideactio
ns,b
utto
guidetheorderin
gof
subgoals
them
selves,
need
tomain
tainoverall
goal
oftask
(Mak
e
sandwich
)
AUnified
Activ
ation-based
Acco
unt
Cen
tralfro
ntal
mech
anism
s:
Activation-based
working
mem
oryFrontal
neu
ronsmain
tainactiv
ely
over
delay
s.
Monitoring/evaluation
e.g.,Erro
r-monito
ring,critical
fordopam
inerg
ic
modulatio
n.
AUnified
Activ
ation-based
Acco
unt
InhibitionNeed
tomain
taintop-downactiv
ationforweak
ertask
.
AUnified
Activ
ation-based
Acco
unt
InhibitionNeed
tomain
taintop-downactiv
ationforweak
ertask
.
Flexibility
Dynam
icsofactiv
ation-based
more
rapid
than
weig
ht-b
ased.
AUnified
Activ
ation-based
Acco
unt
InhibitionNeed
tomain
taintop-downactiv
ationforweak
ertask
.
Flexibility
Dynam
icsofactiv
ation-based
more
rapid
than
weig
ht-b
ased.
Fluency
Only
problem
w/novel
categories
ofresp
onses
—need
top-downsu
pportto
overco
meprep
oten
tcateg
ories.
AUnified
Activ
ation-based
Acco
unt
InhibitionNeed
tomain
taintop-downactiv
ationforweak
ertask
.
Flexibility
Dynam
icsofactiv
ation-based
more
rapid
than
weig
ht-b
ased.
Fluency
Only
problem
w/novel
categories
ofresp
onses
—need
top-downsu
pportto
overco
meprep
oten
tcateg
ories.
Executive
controlMain
tain&
update
plan
s/goals
over
time,av
oid
distractio
n.
Higher
Lev
elCognitio
n:What’s
Missin
g
•Plan
ning
•Reaso
ning
•Decisio
n-m
aking
•Emotio
n
•Conscio
usn
ess,sen
seofself
•Free
will
•Social
interactio
n
Higher
Lev
elCognitio
n:What’s
Missin
g
•Plan
ning
•Reaso
ning
•Decisio
n-m
aking
•Emotio
n
•Conscio
usn
ess,sen
seofself
•Free
will
•Social
interactio
n
Bey
ondthePFC
Bias
&Bindingin
thePFCan
dHippocam
pus:
Bias
Familiar(Hippocampus Independent)
Novel(Hippocampus Dependent)
Binding
Autom
atic
Weak/S
ustained(P
FC
Dependent)
Strong/T
ransient(P
FC
Independent)
Novel,
transient
Routine,
transient
Novel,
sustainedor w
eak
Fam
iliar,sustainedor w
eak
Controlled