nurture in development: a parallel distributed …stanford.edu/~jlmcc/papers/mcclelland94.pdfthe...

17
(1994). In P. Bertelson, P. Eelen , G. d' Ydewalle (Eds. International Perspectives on Psychological Scienc~ Volume 1: Leading Jhemes. Erlbaum: Hillsdale , NJ. The Interaction of Nature and Nurture in Development: A Parallel Distributed Processing Perspective James L. McClelland Department of Psychology, Carnegie Mellon University, Pittsburgh, USA Parallel distributed processing (PDP) models provide a rich set of resources for exploring issues of nature, nurture and their interaction in cognition development. I present the essential aspects of the PDP(orconnectionist) framework, and I draw parallels between the child as learner and the mechanisms of learning in connectionist systems. The remaining sections discuss some of the implications of this framework for our understanding of the acquisition of knowledge. I point outlhat many lines ofargumenllhal have typically been given in support ofnalivist approaches need to be reconsidered in the light of the characteristics of PDP models of learning and development. The first of these sections points out that connectionist models offer a dramatic advance over classical associationist approaches to learning. The second illustrates how stage- like progressions can be understood in terms of the typical learning trajectories seen in connectionist models. The third section considers the meaning and possible sources of early competence from- a PDP perspective, and the fourth considers how connectionist models may shed light on the fact that some of the structure of human behaviour appears to be imposed by the learner. In all, the chapter amounts to an argument that connectionist models allow us to see ways in which experience might lead to the rich and interesting cognitive structures and developmental progressions that have often been taken as supportive of nativist approaches. INTRODUCTION Where do cognitive abilities come from? Are they born in us, innate endowments of nature? Are they products of experience. plain and simple? Or do they arise through the interaction of the characteristics of the organism and the

Upload: others

Post on 10-Jul-2020

6 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Nurture in Development: A Parallel Distributed …stanford.edu/~jlmcc/papers/McClelland94.pdfThe Interaction of Nature and Nurture in Development: A Parallel Distributed Processing

(199

4). I

n P. Bertelson, P. Eelen, G. d'

Yde

wal

le (

Eds

.International Perspectives on Psychological Scienc~

Volume 1: Le

adin

g Jh

emes

. Erl

baum

: Hill

sdal

e, NJ.

The

Int

erac

tion

of N

atur

e an

dN

urtu

re in

Dev

elop

men

t:A

Par

alle

l Dis

trib

uted

Processing Perspective

Jam

es L

. McC

lella

ndD

epar

tmen

t of P

sych

olog

y, C

arne

gie

Mel

lon

Uni

vers

ity,

Pitt

sbur

gh, U

SA

Para

llel d

istr

ibut

ed p

roce

ssin

g (P

DP)

mod

els

prov

ide

a ri

ch s

et o

f re

sour

ces

for

expl

orin

g is

sues

of

natu

re, n

urtu

re a

nd th

eir

inte

ract

ion

in c

ogni

tion

deve

lopm

ent.

I pr

esen

t the

ess

entia

l asp

ects

of

the PDP(orconnectionist) framework, and I draw

para

llels

bet

wee

n th

e ch

ild a

s le

arne

r an

d th

e m

echa

nism

s of

lear

ning

inco

nnec

tioni

st s

yste

ms.

The

rem

aini

ng s

ectio

ns d

iscu

ss s

ome

of th

e im

plic

atio

nsof

this

fra

mew

ork

for

our

unde

rsta

ndin

g of

the

acqu

isiti

on o

f kn

owle

dge.

I p

oint

outlh

at m

any

lines

ofa

rgum

enllh

al h

ave

typi

cally

bee

n gi

ven

in s

uppo

rt o

fnal

ivis

tap

proa

ches

nee

d to

be

reco

nsid

ered

in th

e lig

ht o

f th

e ch

arac

teri

stic

s of

PD

P m

odel

sof

lear

ning

and

dev

elop

men

t. T

he f

irst

of

thes

e se

ctio

ns p

oint

s ou

t tha

t con

nect

ioni

stm

odel

s of

fer

a dr

amat

ic a

dvan

ce o

ver

clas

sica

l ass

ocia

tioni

st a

ppro

ache

s to

lear

ning

. The

sec

ond

illus

trat

es h

ow s

tage

- lik

e pr

ogre

ssio

ns c

an b

e un

ders

tood

inte

rms

of th

e ty

pica

l lea

rnin

g tr

ajec

tori

es s

een

in c

onne

ctio

nist

mod

els.

The

thir

dse

ctio

n co

nsid

ers

the

mea

ning

and

pos

sibl

e so

urce

s of

ear

ly c

ompe

tenc

e fr

om-

aPD

P pe

rspe

ctiv

e, a

nd th

e fo

urth

con

side

rs h

ow c

onne

ctio

nist

mod

els

may

she

d lig

hton

the

fact

that

som

e of

the

stru

ctur

e of

hum

an b

ehav

iour

app

ears

to b

e im

pose

dby

the

lear

ner.

In

all,

the

chap

ter

amou

nts

to a

n ar

gum

ent t

hat c

onne

ctio

nist

mod

els

allo

w u

s to

see

way

s in

whi

ch e

xper

ienc

e m

ight

lead

to th

e ri

ch a

nd in

tere

stin

gco

gniti

ve s

truc

ture

s an

d de

velo

pmen

tal p

rogr

essi

ons

that

hav

e of

ten

been

take

n as

supp

ortiv

e of

nat

ivis

t app

roac

hes.

INT

RO

DU

CT

ION

Whe

re d

o co

gniti

ve a

bilit

ies

com

e fr

om?

Are

they

bor

n in

us,

inna

te e

ndow

men

tsof

nat

ure?

Are

they

pro

duct

s of

exp

erie

nce.

pla

in a

nd s

impl

e? O

r do

they

ari

sethrough the interaction of the characteristics of the or

gani

sm a

nd th

e

Page 2: Nurture in Development: A Parallel Distributed …stanford.edu/~jlmcc/papers/McClelland94.pdfThe Interaction of Nature and Nurture in Development: A Parallel Distributed Processing

McC

LE

LL

AN

D4. NATURE AND NURTURE

IN D

EV

EL

OPM

EN

T

envi

ronm

ent?

The

se q

uest

ions

hav

e st

ood

at th

e ce

ntre

of

the

stud

y of

min

d fo

rce

ntur

ies.

Som

e of

the

mos

t pro

min

ent r

esea

rche

rs o

f ou

r ce

ntur

y-C

hom

sky,

Skinner and Piaget-ha

ve ta

ken

each

of

thes

e vi

ews,

and

eac

h ha

s sp

awne

d la

rge

follo

win

gs w

ithin

psy

chol

ogy

and

the

larg

er s

cien

tific

com

mun

ity~

Whi

ch p

ositi

on is

rig

ht?

The

pur

e em

piri

cist

trad

ition

no

long

e~ h

olds

muc

hsw

ay; a

nd th

ere

is c

erta

inly

a v

ery

activ

e m

ovem

ent i

n th

e fi

eld

toda

y th

at f

avou

rsth

e na

tivis

t pos

ition

. Yet

I th

ink

ther

e is

a f

eelin

g in

man

y qu

aner

s th

at e

xper

ienc

em

ust p

laya

larg

er r

ole than the mere setting of parameters in an otherwise

inna

tely

pre

dete

nnin

ed c

ogni

tive

syst

em.

In th

is c

hapt

er, I

will

add

ress

this

que

stio

n. I

will

arg

ue f

or a

n in

tera

ctio

nist

posi

tion

, sim

ilar

in s

ome

way

s to

Pia

get's

. My

mai

n ai

m, t

houg

h, "

Viii

not

be

tochampion interactionism

per se.

Rat

her,

it w

ill b

e to

sug

gest

that

the

para

llel

dist

ribu

ted

proc

essi

ng f

ram

ewor

k (a

ka th

e co

nnec

tioni

st f

ram

ewor

k) h

as b

road

impl

icat

ions

for

our

und

erst

andi

ng o

f co

gniti

ve d

evel

opm

ent.

Thi

s fr

amew

ork,

I w

ill a

rgue

, pro

vide

s m

echa

nism

s an

d id

eas

that

allo

w u

s to

exp

lore

the

inte

rpla

yof

nat

ure

and

nunu

re, a

nd th

at s

ugge

st h

ow e

xper

ienc

e m

ay b

e th

e en

gine

that

driv

es d

evel

opm

ent,

thro

ugh

chan

nels

sha

ped

by b

oth

inna

te c

onst

rain

ts a

nd th

est

ruct

ure

of th

e en

viro

nmen

t.I

begi

n w

ith a

pre

sent

atio

n of

the

esse

ntia

l asp

ects

of

para

llel d

istr

ibut

edpr

oces

sing

that

are

rel

evan

t to

the

poin

ts I

hope

to m

ake

abou

t cog

nitiv

ede

velo

pmen

t , a

nd I

dra

w p

aral

lels

bet

wee

n th

e ch

ild a

s le

arne

r/ex

peri

ence

r an

dth

e m

echa

nism

s of

lear

ning

in c

onne

ctio

nist

sys

tem

s. T

he r

emai

ning

sec

tions

disc

uss

som

e of

the

impl

icat

ions

of

this

fra

mew

ork

for

our

unde

rsta

ndin

g of

the

repr

esen

tatio

n an

d ac

quis

ition

of

know

ledg

e.Fi

rst,

I w

ill p

oint

out

that

con

nect

ioni

st m

odel

s of

fer

a dr

amat

ic a

dvan

ce o

ver

othe

r fo

nns

of le

arni

ng, p

anic

ular

ly c

lass

ical

ass

ocia

tioni

st a

ppro

ache

s. T

his

poin

tis

cru

cial

, sin

ce it

rel

ates

qui

te s

tron

gly

to th

e qu

estio

n of

wha

t is

lear

nabl

e.Se

cond

, I w

ill e

xplo

re th

e tim

e-co

urse

of

deve

lopm

ent.

In p

artic

ular

, I w

illco

nsid

er s

tage

- lik

e pr

ogre

ssio

ns. H

ere

I w

ill d

emon

stra

te h

ow c

onne

ctio

nist

mod

els

impl

emen

t mec

hani

sms

of c

ogni

tive

chan

ge v

ery

clos

e to

som

e of

the

mec

hani

sms

prop

osed

by

Piag

et, a

nd h

ow th

ey c

an a

ddre

ss s

ome

of th

e pu

zzle

sth

at h

ave

plag

ued

othe

r ex

perie

ntia

l acc

ount

s of

the

sour

ces

of p

rogr

ess

from

stag

e to

sta

ge. T

hird

, I w

ill c

onsi

der

the

mea

ning

and

pos

sibl

e so

urce

s of

ear

lyco

mpe

tenc

e , f

rom

a c

onne

ctio

nist

poi

nt o

f vi

ew. F

inal

ly, I

will

con

side

r ho

wco

nnec

tioni

st m

odel

s m

ay s

hed

light

on

the

fact

that

som

e of

the

stru

ctur

e of

hum

an b

ehav

iour

app

ears

to b

e im

pose

d by

the

lear

ner.

A f

ew o

f th

e po

ints

mad

e in

the

chap

ter

have

bee

n m

ade

prev

ious

ly in

McC

lella

nd (

1989

). B

ates

and

Elm

an (

1993

), K

arm

iloff-

Smith

(19

92a,

I 992

b) a

nd P

lunk

ett a

nd S

inha

(19

91)

have

writ

ten

on r

elat

ed to

pics

and

man

yof

thei

r ar

gum

ents

hav

e co

ntri

bute

d to

the evolution of the viewpoints

expr

esse

d he

re.

TH

E P

DP

FR

AM

EW

OR

K

Para

llel d

istr

ibut

ed p

roce

ssin

g (P

DP)

pro

vide

s us

with

a f

ram

ewor

k fo

r th

inki

ngab

out t

he m

echa

nism

s th

at r

epre

sent

, acq

uire

and

use

kno

wle

dge.

The

fra

mew

ork

is d

escr

ibed

in d

etai

l in

the

firs

t fou

r ch

apte

rs o

f R

umel

han,

McC

lella

nd a

nd th

ePD

P R

esea

rch

Gro

up (

1986

c). H

ere

I to

uch

only

on

poin

ts r

elev

ant t

o th

e pr

esen

tdi

scus

sion

.Fi

rst,

PDP

assu

mes

that

cog

nitiv

e pr

oces

ses

aris

e fr

om th

e in

tera

ctio

ns o

f la

rge

num

bers

of

sim

ple

proc

essi

ng u

nits

, org

anis

ed in

to m

odul

es. A

ver

y ge

neri

cex

ampl

e of

suc

h a

netw

ork

is s

how

n in

Fig

. 4. 1

. With

in e

ach

mod

ule,

eac

h un

itco

mpu

tes

a si

mpl

e fu

nctio

n of

the

inpu

ts it

rec

eive

s fr

om o

ther

uni

ts. C

ruci

al to

us w

ill b

e th

e fa

ct th

at th

is is

a c

ontin

uous

but

non

- lin

ear

func

tion,

like

the

one

show

n in

Fig

. 4,1

. Sec

ond,

the

PDP

appr

oach

ass

umes

that

the

know

ledg

e th

atgo

vern

s pr

oces

sing

is s

tore

d in

the

stre

ngth

s of

the

conn

ectio

ns a

mon

g th

e un

its.

Such

kno

wle

dge

allo

ws

the

patte

rn p

rese

nt o

n on

e se

t of

units

to g

ive

rise

to o

ther

patte

rns

on o

ther

set

s of

uni

ts; o

r fo

r th

e pa

ttern

of

activ

atio

n at

one

poi

nt in

tim

eto

giv

e ri

se to

a s

ucce

ssor

at t

he n

ext m

omen

t. T

hird

, the

PD

P ap

proa

ch a

ssum

esth

at a

cqui

sitio

n of

kno

wle

dge

occu

rs th

roug

h th

e ad

just

men

t of

conn

ectio

nst

reng

ths.

The

se a

djus

tmen

ts, i

n tu

rn, a

re d

rive

n by

sig

nals

ari

sing

ulti

mat

ely

from

exte

rnal

inpu

ts to

the

netw

ork.

I d

iscu

ss th

is a

spec

t mor

e fu

lly b

elow

.

The

Rol

e of

the

Inna

te E

ndow

men

t

Let

us

now

con

side

r ho

w th

is f

ram

ewor

k al

low

s us

to e

xplo

re th

e in

tera

ctio

n of

inna

te e

ndow

men

t and

exp

erie

nce

as d

eten

nina

nts

of th

e co

urse

and

out

com

e of

know

ledg

e ac

quis

ition

. The

inna

te e

ndow

men

t has

sev

eral

asp

ects

. Am

ong

them

is th

e gr

oss

arch

itect

ure

of th

e sy

stem

- the

mod

ules

, the

num

ber

of p

roce

ssin

gun

its in

eac

h, a

nd th

eir

initi

al c

onne

ctiv

ity (

Rum

elha

n, H

into

n, &

McC

lella

nd,

;..,:: -0(

Inpu

t

FIG

. 4. 1 A

gene

ric

conn

eciio

nisl

nel

wor

k, lo

gelh

er w

ilh il

s ac

lival

ion

func

lion.

Page 3: Nurture in Development: A Parallel Distributed …stanford.edu/~jlmcc/papers/McClelland94.pdfThe Interaction of Nature and Nurture in Development: A Parallel Distributed Processing

McC

LE

LL

AN

D4. NATURE AND NURTURE

IN D

EV

EL

OPM

EN

T

1986

a). W

e kn

ow o

f co

urse

that

the

brai

n is

not

at a

ll fu

lly c

onne

cted

and

the

orga

nisa

tion

of th

is c

onne

ctiv

ity d

eten

nine

s w

hich

uni

ts w

ill b

e in

line

to r

ecei

vere

lativ

ely

dire

ct a

udito

ry in

put,

whi

ch to

rec

eive

vis

ual i

nput

, whi

ch to

com

bine

thes

e tw

o so

urce

s of

info

nnat

ion,

and

so

on.

A s

econ

d as

pect

of

the

inna

te e

ndow

men

t con

cern

s th

e ba

sic

rule

s th

at g

over

nth

e pr

opag

atio

n of

act

ivat

ion

and

the

adju

stm

ent o

f co

nnec

tion

stre

ngth

s. T

hese

rule

s, to

the

exte

nt th

at th

ey d

iffe

r in

dif

fere

nt p

arts

of

the

syst

em, p

rovi

de a

mea

nsw

here

by m

odal

ity-

or d

omai

n-sp

ecifi

c co

nstr

aint

s ca

n be

impo

sed

on th

eco

gniti

ve s

yste

m. T

o th

e ex

tent

that

they

are

gen

eral

thro

ugho

ut th

e sy

stem

, the

ypr

ovid

e a

set o

f pe

rvas

ive

com

mon

mec

hani

sms

for

proc

essi

ng a

nd le

arni

ng,

A th

ird

aspe

ct o

f th

e in

nate

end

owm

ent c

once

rns

the

deta

iled

para

met

ers

ofdi

ffer

ent p

arts

of

the

syst

em. E

ven

whe

n th

e ba

sic

rule

s of

pro

cess

ing

and

conn

ectio

n st

reng

th a

re h

eld

cons

tant

, dif

fere

nces

in p

aram

eter

s ca

n le

ad to

mec

hani

sms

with

ver

y di

ffer

ent c

hara

cter

istic

s. A

s on

e ex

ampl

e, O

'Rei

lly (

1992

)to

ok a

net

wor

k (i

llust

rate

d in

Fig

. 4.2

) co

nsis

ting

of tw

o se

ts o

f in

tern

al u

nits

,ea

ch r

ecei

ving

inpu

t fro

m th

e sa

me

set o

f in

puts

, and

eac

h se

ndin

g ou

tput

to a

llof

the

outp

ut u

nits

. The

inte

rnal

uni

ts in

the

two

pool

s di

ffer

ed o

nly

in th

e ra

teat

whi

ch th

eir

activ

atio

ns c

hang

ed o

ver

time.

Tho

se o

n th

e le

ft c

hang

ed a

ctiv

atio

nsl

owly

whe

n th

eir

inpu

ts c

hang

ed, w

hile

thos

e on

the

righ

t cha

nged

thei

rac

tivat

ion

very

rap

idly

in r

espo

nse

to c

hang

ing

inpu

ts. T

he n

etw

ork

was

ask

ed

to p

erfo

nn th

e ta

sk o

f lea

rnin

g to

indi

cate

the

loca

tion

and

iden

tity

of in

put

patte

rns.

Som

e of

the

outp

ut u

nits

wer

e us

ed to

spe

cify

iden

tity,

and

the

othe

rsw

ere

used

to s

peci

fy s

hape

, with

bot

h se

ts o

f hi

dden

uni

ts c

onne

cted

to a

ll of

the

outp

ut u

nits

. Eac

h in

put p

atte

rn a

rose

at a

dis

cret

e po

int i

n tim

e an

d th

en p

ersi

sted

for

seve

ral t

ime

step

s, m

ovin

g ab

out f

rom

pla

ce to

pla

ce w

ithin

the

inpu

t arr

ay,

as if

the

netw

ork

wer

e re

ceiv

ing

inpu

t at d

iffe

rent

ret

inal

loca

tions

ove

r a

seri

esof

eye

fix

atio

ns. T

hus,

the

iden

tity

of a

pat

tern

tend

ed to

per

sist

for

sev

eral

tim

est

eps,

whi

le th

e lo

catio

n of

the

patte

rn m

oved

aro

und.

The

res

ult w

as th

at th

epo

ol o

f un

its w

ith th

e sl

ow r

ate

of c

hang

e be

cam

e sp

ecia

lised

for

det

ectin

g th

eid

entit

y of

the

obje

ct, w

hile

the

units

with

the

fast

er r

ate

of c

hang

e be

cam

esp

ecia

lised

for

det

ectin

g its

loca

tion.

Thi

s is

but

one

exa

mpl

e of

a b

urge

onin

gbo

dy o

f w

ork

show

ing

how

mod

ules

with

slig

htly

dif

fere

nt in

itial

par

amet

ers

orbi

ases

can

bec

ome

diff

eren

tially

spe

cial

ised

thro

ugh

expe

rien

ce. T

his

part

icul

arca

se a

lso

emph

asis

es th

e jo

int r

oles

of

envi

ronm

ent a

nd in

nate

str

uctu

re. I

t is

the

com

bina

tion

of th

e fa

ct th

at o

bjec

t ide

ntity

rem

ains

the

sam

e as

the

retin

allo

catio

n ch

ange

s w

ith th

e fa

ct th

at th

e un

its in

dif

fere

nt m

odul

es d

iffe

r in

thei

rte

mpo

ral p

aram

eter

s th

at d

eten

nine

s th

e ou

tcom

e in

this

cas

e.

The

Rol

e of

Exp

erie

nce

Now

let u

s fo

cus

on th

e ro

le o

f ex

peri

ence

. Her

e I

will

beg

in b

y pr

opos

ing

age

nera

l way

of

thin

king

abo

ut h

ow e

xper

ienc

e m

ight

pro

pel c

ogni

tive

deve

lopm

ent,

and

then

I w

ill r

elat

e th

is id

ea b

ack

to p

aral

lel d

istr

ibut

edpr

oces

sing

(M

cCle

lland

, 198

9). W

e be

gin

by th

inki

ng o

f th

e ch

ild e

xper

ienc

ing

an e

nvir

onm

ent i

nvol

ving

an

ongo

ing

sequ

ence

of

even

ts. W

e as

sum

e th

at th

echild, while alert and attending, is always making implicit predictions

for

wha

tm

ight

hap

pen

next

. Let

me

emph

asis

e th

e im

plic

it na

ture

of

the

proc

ess.

I d

onm

ean

that

the

child

is c

onsc

ious

ly a

war

e of

ask

ing,

and

pos

ing

answ

ers

to th

equ

estio

n

, "

Wha

ts

com

ing

next

? W

hat

s co

min

g ne

xt?"

I d

o m

ean,

thou

gh, t

hat

his

or h

er c

ogni

tive

syst

em is

in f

act a

ntic

ipat

ing

the

futu

re, a

nd th

at a

rea

ctio

nca

n oc

cur

if th

ese

expe

ctat

ions

are

vio

late

d, B

ehav

iour

ally

, suc

h re

actio

ns a

reac

com

pani

ed b

y or

ient

ing

resp

onse

s--e

ye m

ovem

ents

, pup

il di

latio

n, in

crea

ses

in g

aze

dura

tion,

etc

. The

y al

so, a

t lea

st in

lang

uage

pro

cess

ing,

gen

erat

e la

rge

and

robu

st e

voke

d po

tent

ials

, suc

h as

the

N40

0 of

Kut

as a

nd H

illya

rd (

1980

).N

ote

that

mat

chin

g or

mis

mat

chin

g is

alw

ays

a m

atte

r of

deg

ree

in P

OP

syst

ems-

it is

a m

atte

r of

the

cons

iste

ncy

betw

een

one

patte

rn o

f ac

tivat

ion

and

anot

her-

and

in g

ener

al a

ctua

l eve

nts

will

alw

ays

diff

er to

som

e de

gree

fro

mw

hat i

s an

ticip

ated

.In

any

cas

e, a

s ev

ents

occ

ur, t

hey

perm

it th

e ch

ild to

com

pare

impl

icit

pred

ictio

ns w

ith w

hat a

ctua

lly h

appe

ns n

ext ,

and

ther

eby

to le

arn

to m

ake

bette

rpr

edic

tions

fo

r th

e fu

ture

. The

ess

entia

l ass

umpt

ion

is th

at th

e ch

ild u

ses

the

follo

win

g pr

oced

ure

to a

djus

t the

pre

dict

ions

:

Inpu

t

FIG

. 4.2 O'

Rei

llys network. The slow units in the module

on

the

lefl

bec

ome

"wha

t" u

nits

and

the fast units in the module

on

the

righ

t bec

ome

"whe

re"

units

.

Page 4: Nurture in Development: A Parallel Distributed …stanford.edu/~jlmcc/papers/McClelland94.pdfThe Interaction of Nature and Nurture in Development: A Parallel Distributed Processing

4. NATURE AND NURTURE

IN D

EV

EL

OPM

EN

TM

cCL

EL

LA

ND

Adj

ust e

ach

para

met

er o

f th

e m

ind

in p

ropo

rtio

n to

the

exte

nt th

at it

s ad

just

men

tw

ill r

educ

e th

e di

scre

panc

y be

twee

n pr

edic

ted

and

obse

rved

eve

nts.

Seco

nd, t

he k

now

ledg

e is

inhe

rent

ly g

rade

d in

nat

ure.

Thi

s is

a m

ost i

mpo

rtan

tpo

int,

and

one

that

will

per

vade

muc

h of

the

rest

of

this

cha

pter

. Whe

n a

netw

ork

is f

irst

pla

ced

in a

n en

viro

nmen

t, its

con

nect

ion

wei

ghts

mig

ht b

e in

itial

ised

tosm

all,

rand

om v

alue

s, w

ithin

bro

ad c

onne

ctiv

ity c

onst

rain

ts. I

ts p

redi

ctio

ns,

then

, mig

ht b

e in

itial

ly w

eak

and

rand

om. A

s ex

peri

ence

s oc

cur ,

pre

dict

ions

grad

ually

bec

ome

attu

ned

to s

ituat

ions

, and

aft

er m

ore

and

mor

e tim

e ha

s pa

ssed

,th

e ne

twor

k be

com

es s

truc

ture

d in

acc

orda

nce

with

the

cont

ents

of

the

dom

ain.

At f

irst

we

wou

ld s

ay th

e ne

twor

k kn

ows

noth

ing,

and

at t

he e

nd w

e w

ould

say

the

netw

ork

know

s th

e st

ruct

ure

of th

e do

mai

n, b

ut in

the

mid

dle

our

ordi

nary

tenn

inol

ogy

and

way

s of

thin

king

abo

ut w

hat i

t mea

ns to

"kn

ow"

som

ethi

ng b

reak

dow

n. T

he n

etw

ork

has

a ki

nd o

f pa

rtia

l kno

wle

dge-

it ge

nera

tes

pred

ictio

nsth

at in

clin

e to

war

ds b

eing

cor

rect

, or

are

corr

ect o

nly

in c

erta

in c

ases

, or

that

capt

ure

the

gros

s re

gula

ritie

s w

ithou

t cap

turi

ng th

e su

btle

r de

tails

. The

pro

cess

is c

ompl

etel

y gr

adua

l and

ther

e is

no

spec

ial d

iscr

ete

poin

t at w

hich

we

wou

ldsa

y th

e ne

twor

k no

w k

now

s, a

nd b

efor

e th

is, i

t did

not

kno

w.

Thi

rd, a

cqui

sitio

n is

gra

dual

and

incr

emen

tal.

Gra

dual

ness

is im

port

ant,

asw

e w

ill s

ee in

our

con

side

ratio

n of

an

exam

ple

belo

w, b

ecau

se it

allo

ws

the

over

all

dire

ctio

n of

cha

nge

to b

e de

tenn

ined

by

the

over

all s

truc

ture

in th

e do

mai

n th

atit

is le

arni

ng r

athe

r th

an s

peci

fic

indi

vidu

al s

timul

i. It

is in

crem

enta

l, in

that

eac

hch

ange

bui

lds

on e

arlie

r ch

ange

s. A

s in

Pia

get's

dev

elop

men

tal t

heor

y, w

here

cogn

itive

str

uctu

re e

mer

ges

from

acc

omm

odat

ion

and

assi

mila

tion,

eac

h ne

wad

apta

tion

is b

ut a

slig

ht v

aria

tion

and

exte

nsio

n of

the

syst

em th

at r

esul

ted

from

prev

ious

act

s of

ada

ptat

ion

(see

Fla

vell,

196

3, f

or d

iscu

ssio

n).

Giv

en th

e im

plic

it na

ture

of

the

know

ledg

e em

bodi

ed in

PO

P sy

stem

s, it

may

seem

to m

any

rese

arch

ers

at f

irst

gla

nce

that

PO

P sy

stem

s ha

ve r

elat

ivel

y lit

tleto

off

er a

ny e

ffor

t to

unde

rsta

nd c

ogni

tive

deve

lopm

ent.

Thi

s w

ill b

e es

peci

ally

true

for

res

earc

hers

who

thin

k of

"re

al"

know

ledg

e as

exp

licit

know

ledg

e, o

rw

ho th

ink

of d

isco

ntin

uous

cha

nge,

suc

h as

insi

ght,

as th

e ha

llmar

k de

velo

pmen

t. T

here

is c

erta

inly

no

deny

ing

that

exp

licit

know

ledg

e ex

ists

. or

that

the

deve

lopm

ent o

f ex

plic

it kn

owle

dge

is a

maj

or a

spec

t of

cogn

itive

deve

lopm

ent.

But

I b

elie

ve th

at im

plic

it kn

owle

dge,

of

the

kind

that

is b

uilt

into

conn

ectio

nist

sys

tem

s th

roug

h ex

perie

nce,

is m

ore

fund

amen

tal ,

in th

at it

stru

ctur

es th

e ve

ry r

epre

sent

atio

ns o

f ex

peri

ence

itse

lf th

at p

rovi

de th

e in

put t

oex

plic

it th

ough

t pro

cess

es. C

erta

inly

, we

all a

ppre

ciat

e fr

om li

ngui

stic

s th

esu

btle

ty th

at im

plic

it kn

owle

dge

can

have

, and

the

role

that

it c

an p

lay

in s

hapi

ngou

r re

pres

enta

tions

of

lingu

istic

stim

uli.

I sh

are

the

view

of

Kar

milo

ff- S

mith

(199

2a)

that

suc

h im

plic

it kn

owle

dge

is a

s im

port

ant i

n ot

her

dom

ains

, and

serv

es, a

s it

does

for

lang

uage

, as

the

subs

trat

e on

whi

ch e

xplic

it co

gniti

on is

ultim

atel

y bu

ilt. O

n th

is v

iew

, our

und

erst

andi

ng o

f ex

plic

it co

gniti

on w

ill b

een

rich

ed b

y a

deep

er u

nder

stan

ding

of

this

sub

stra

te, T

here

fore

, the

res

t of

this

chap

ter

will

foc

us o

n th

e im

plic

atio

ns o

f PO

P fo

r th

inki

ng a

bout

the

natu

re a

ndac

quis

ition

of

know

ledg

e of

this

impl

icit

kind

.

The

mos

t str

aigh

tfor

war

d an

d ty

pica

l way

of

rela

ting

this

idea

to c

onne

ctio

nist

netw

orks

is a

s fo

llow

s (s

ee, e

.g. E

lman

, 199

0): F

irst,

we

imag

ine

that

the

inpu

tto

the

syst

em r

epre

sent

s th

e cu

rren

t situ

atio

n, f

rom

whi

ch a

pre

dict

ion

will

be

mad

e. T

his

inpu

t may

be

sign

als

aris

ing

from

out

side

the

syst

em, o

r th

e sy

stem

own

inte

rnal

sta

te. S

econ

d, w

e im

agin

e th

at th

e ou

tput

rep

rese

nts

the

pred

ictio

nth

at th

e sy

stem

mak

es f

or w

hat m

ay h

appe

n ne

xt. F

inal

ly, w

e im

agin

e th

at th

eac

tual

nex

t eve

nt a

rise

s as

a p

atte

rn o

f ac

tivity

on

thes

e un

its. T

he r

ule

for

lear

ning

bec

omes

:

Adj

ust e

ach

conn

ectio

n w

eigh

t in

the

netw

ork

in p

ropo

rtio

n to

the

exte

nt th

atits

adj

ustm

ent w

ill r

educ

e th

e di

ffer

ence

bet

wee

n th

e ou

tput

of

the

netw

ork

and

the

actu

al n

ext e

vent

.

The

alg

orith

m s

tand

ardl

y us

ed f

or a

djus

ting

conn

ectio

n st

reng

ths,

the

back

-pr

opag

atio

n le

arni

ng a

lgor

ithm

(R

umel

hart

, Hin

ton

, & W

illia

ms,

198

6b),

issimply an algorithm for calculating the relevant quantities-in

par

ticul

ar, t

heex

tent

to w

hich

an

adju

stm

ent t

o ea

ch w

eigh

t will

red

uce

a m

easu

re o

f th

edi

ffer

ence

s ac

ross

the

outp

ut u

nits

.G

iven

suc

h a

lear

ning

rul

e, e

xper

ienc

e dr

ives

cha

nges

in c

onne

ctio

n w

eigh

ts,

ther

eby

caus

ing

the

netw

ork

to le

arn

how

to r

epre

sent

and

use

inpu

ts to

mak

ebe

tter

pred

ictio

ns f

or th

e fu

ture

. Jus

t thi

s ap

proa

ch h

as b

een

used

to tr

ain

netw

orks

to d

o a

num

ber

of v

ery

impo

rtan

t and

inte

rest

ing

thin

gs, w

hich

we

will

turn

toin

a m

omen

t. F

or n

ow, t

houg

h, I

wan

t to

high

light

bri

efly

thre

e ke

y fe

atur

es o

fth

e co

nnec

tioni

st a

ppro

ach

to k

now

ledg

e an

d kn

owle

dge

acqu

isiti

on. -

Pro

pert

ies

of C

onne

ctio

nist

Kno

wle

dge

Firs

t, th

e kn

owle

dge

is im

plic

it, in

the

very

sen

se th

at th

is te

nn h

as b

een

used

in th

e de

velo

pmen

tal l

itera

ture

by

Kan

nilo

ff-Smith (1986,

1991

, 199

2a),

and

inth

e m

emor

y lit

erat

ure

by a

num

ber

of in

vest

igat

ors

(see

Sch

acte

r, 1

987)

. It i

skn

owle

dge

that

is e

mbe

dded

in s

peci

fic

proc

essi

ng m

echa

nism

s. I

t is

know

ledg

eth

at c

an b

e us

ed to

mak

e pr

edic

tions

, perceptual anticipations and pattern

com

plet

ions

, and

to s

ubse

rve

the

grad

ual a

cqui

sitio

n of

dom

ain-

spec

ific

ski

lls.

Vio

latio

ns o

f pr

edic

tions

mad

e on

the

basi

s of

suc

h kn

owle

dge

can

serv

e as

the

basi

s of

mak

ing

vari

ous

kind

s of

judg

emen

ts, a

s w

hen

a pe

rson

is a

sked

to m

ake

a ju

dgem

ent a

bout

the

gram

mat

ical

ity o

f a

pote

ntia

l exa

mpl

e of

a n

atur

al o

rar

tific

ial l

angu

age

(Kno

wlto

n, R

amus

, & S

quir

e, 1

992;

Reb

er, 1

976,

198

9). B

utth

e kn

owle

dge

itsel

f is

not

dir

ectly

ava

ilabl

e as

suc

h, e

ither

for

rep

ort,

for

expl

icit

reas

onin

g, o

r an

y ot

her

purp

ose.

Page 5: Nurture in Development: A Parallel Distributed …stanford.edu/~jlmcc/papers/McClelland94.pdfThe Interaction of Nature and Nurture in Development: A Parallel Distributed Processing

McC

LE

LL

AN

D4. NATURE AND NURTURE

IN D

EV

EL

OPM

EN

T

IMPL

ICA

TIO

NS

OF

TH

E P

DP

APP

RO

AC

HA

lthou

gh h

e ac

know

ledg

es th

at e

xper

ienc

e pl

ays

a ro

le in

thes

e de

velo

pmen

ts,

Kei

l arg

ues

in s

ever

al p

lace

s (1

981,

198

7, 1991 a, 1991 b) for the idea that the

abili

ty to

use

exp

erie

nce

in th

e se

rvic

e of

the

cons

truc

tion

of s

uch

rela

tiona

lsystems of meaning depends on the use of

a priori,

dom

ain-

spec

ific

kno

wle

dge.

As

pan

of h

is a

rgum

ent h

e po

ints

out

a c

halle

nge

for

any

lear

ning

- bas

ed a

ccou

ntthat does not rely on such

a priori

know

ledg

e. H

e ar

gues

that

cla

ssic

alas

soci

atio

nist

acc

ount

s-ac

coun

ts h

e ta

kes

as p

arad

igm

atic

of

expe

rien

ce-b

ased

appr

oach

es to

kno

wle

dge

acqu

isiti

on-a

ppea

r to

lack

the

cruc

ial a

bilit

y to

shi

ftfr

om a

rel

ianc

e on

mer

ely

corr

elat

ed f

eatu

res

to a

rel

ianc

e on

dee

per

rela

tiona

lpr

open

ies.

The

rea

sons

for

this

, he

argu

es, l

ie in

the

basi

s on

whi

ch th

e le

arni

ngm

echa

nism

s pr

ovid

ed b

y su

ch a

ccou

nts

lear

n an

d us

e w

hat t

hey

have

lear

ned,

Kei

l (19

91a)

not

es tw

o ke

y pr

oper

ties

of a

ssoc

iatio

nist

acc

ount

s: le

arni

ng o

ccur

sby

con

tigui

ty (

for

exam

ple,

con

tigui

ty o

f a

situ

atio

n w

ith a

n ou

tcom

e, o

r of

acu

rren

t eve

nt w

ith th

e ne

xt e

vent

) an

d ge

nera

lisat

ion

occu

rs b

y si

mila

rity,

perm

ittin

g re

spon

ding

to n

ovel

inpu

ts b

ecau

se o

f th

eir

sim

ilari

ty to

fam

iliar

case

s, T

hus,

for

exa

mpl

e, I

ass

ocia

te th

e si

ght o

f a

rose

with

the

smel

l of

a ro

seby

con

tigui

ty, l

earn

ing

to p

redi

ct th

e sm

ell f

rom

the

visu

al a

ppea

ranc

e. I

kno

wth

at a

noth

er r

ose

will

sm

ell a

s sw

eet b

ecau

se it

look

s si

mila

r to

the

firs

t ros

e,In essence, K

eil's

arg

umen

t is

that

chi

ldre

n-an

d ce

rtai

nly adults-do not

alw

ays

gene

ralis

e on

the

basi

s of

sim

ilari

ty. R

athe

r, th

ey c

ome

to r

ely,

som

etim

esal

mos

t exc

lusi

vely

, on

deep

er, r

elat

iona

l inf

orm

atio

n. L

acki

ng a

n ex

peri

ence

- bas

edal

tern

ativ

e to

ass

ocia

tioni

sm th

at o

verc

omes

this

lim

itatio

n, h

e po

ints

to th

e w

ork

of S

pelk

e an

d ot

hers

(e.

g. S

pelk

e, 1

991;

Spe

lke,

Bre

inlin

ger,

Mac

ombe

r, &

Jaco

bson

, 199

2) f

or e

vide

nce

of v

ery

earl

y (a

nd th

eref

ore

puta

tivel

y in

nate

)kn

owle

dge

in a

num

ber

of d

omai

ns, a

nd h

e su

gges

ts th

at c

hild

ren

may

hav

e a

hand

ful o

f in

nate

"pr

oto-

theo

ries

" th

at a

llow

them

to d

eter

min

e ju

st w

hat t

heap

prop

riat

e ba

ses

are

or g

ener

alis

atio

n, in

eac

h of

sev

eral

dom

ains

,N

ow le

t us

take

a lo

ok a

t thi

s ar

gum

ent,

in li

ght o

f PD

P m

odel

s, a

nd s

ee h

owit

hold

s up

. At f

irst

gla

nce,

it s

eem

s th

at P

DP

mod

els

may

not

cha

nge

thin

gs m

uch,

sinc

e th

ese

mod

els

do le

arn

from

con

tigui

ty o

f si

tuat

ions

and

out

com

es, a

s I

have

alre

ady

expl

aine

d. F

unhe

rmor

e, it

is a

lso

true

that

they

tend

to g

ener

alis

e by

sim

ilari

ty: s

imila

r pa

ttern

s of

act

ivat

ion

on o

ne la

yer

of u

nits

tend

to p

rodu

cesi

mila

r pa

ttern

s of

act

ivat

ion

at th

e ne

xt. T

his

may

be

why

so

man

y co

gniti

vesc

ient

ists

and

phi

loso

pher

s w

ith o

ther

wis

e w

idel

y di

verg

ent v

iew

s es

sent

ially

dism

iss

conn

ectio

nist

mod

els

of le

arni

ng. P

utna

m h

as s

tate

d, f

or e

xam

ple,

that

thin

gs th

at c

anno

t be

expr

esse

d in

term

s of

cor

rela

tions

bet

wee

n ha

rd-c

oded

vari

able

s ca

nnot

be

foun

d by

the

algo

rith

m, a

nd a

re n

ot a

ble

to r

ecov

er th

est

ruct

ure

that

lies

beh

ind

the

surf

ace.

Fod

or h

as s

tate

d, "

The

que

stio

n is

whe

ther

you

can

get s

ome

succ

ess

that

you

cou

ld n

ot g

et b

y ju

st d

oing

sta

tistic

s. .

I th

ink

of th

em (

c onn

ectio

nist

mod

els)

as

anal

og s

tatis

tics

pack

ages

,I

The

quo

tatio

ns f

rom

Fod

or a

nd P

utna

m h

ave

been

use

d w

ith p

erm

issi

on o

f th

e au

thor

s (p

erso

nal

com

mun

icat

ion,

199

2). T

he s

tate

men

ts w

ere

orig

inal

ly m

ade

in in

terv

iew

s w

ith P

. Bau

mga

nner

,w

ho c

onve

yed

thes

e qu

otat

ions

to m

e fo

r co

mm

ent.

My

mai

n po

int i

n th

is c

hapt

er is

to s

how

that

the

PDP

appr

oach

has

pro

foun

dim

plic

atio

ns f

or o

ur u

nder

stan

ding

of

deve

lopm

ent.

In g

ener

al, m

y ar

gum

ent i

sth

at w

ith th

e to

ols

prov

ided

by

the

PDP

fram

ewor

k w

e ca

n ca

ll in

to q

uest

ion

som

e of

the

tene

ts o

f na

tivis

t app

roac

hes

on th

e on

e ha

nd, a

nd b

reat

he n

ew li

fein

to th

e in

tera

ctio

nist

pos

ition

on

the

othe

r.T

he r

est o

f th

is c

hapt

er w

ill p

ursu

e th

is p

oint

thro

ugh

four

spe

cifi

c ar

gum

ents

.Fi

rst,

I w

ill s

how

that

con

nect

ioni

st le

arni

ng r

ules

allo

w u

s to

ext

end

the

scop

eof

wha

t is

lear

nabl

e fa

r be

yond

wha

t was

thou

ght p

ossi

ble

unde

r ea

rlier

appr

oach

es, t

here

by r

educ

ing

the

need

man

y au

thor

s ha

ve s

een

to b

uild

spe

cifi

ckn

owle

dge

in f

rom

the

star

t. Se

cond

, I w

ill a

rgue

that

PD

P al

low

s us

to u

nder

stan

dst

age-

like

prog

ress

ions

in d

evel

opm

ent m

ore

clea

rly

than

was

pos

sibl

e un

der

Piag

et's

inte

ract

ioni

st a

ccou

nt. T

hird

, I w

ill p

oint

out

how

PD

Ppro

vide

s us

with

a la

ngua

ge in

whi

ch to

ret

hink

the

mea

ning

of m

any

of th

e in

tere

stin

g ph

enom

ena

of e

arly

com

pete

nce

in in

fant

s, F

ourt

h, I

will

sug

gest

how

PD

P al

low

s us

toun

ders

tand

how

a le

arni

ng p

roce

ss m

ay s

truc

ture

the

dom

ain

that

is b

eing

lear

ned.

The

cha

pter

doe

s no

t lay

out i

ts o

wn

deta

iled

theo

ry o

f de

velo

pmen

t, no

r do

esit

even

use

the

PDP

appr

oach

to m

ake

spec

ific

pre

dict

ions

. Rat

her,

it u

ses

exam

ple

sim

ulat

ions

to d

emon

stra

te s

peci

fic

poin

ts th

at c

ontr

ast w

ith v

iew

poin

ts o

ften

take

n in

mor

e na

tivis

t app

roac

hes.

Tak

en to

geth

er, t

he e

xam

ples

sup

port

the

mor

ege

nera

l cla

im th

at P

DP

can

lead

us

to r

ethi

nk m

uch

of th

e ev

iden

ce th

at h

as b

een

take

n in

sup

port

of

the

view

that

muc

h es

sent

ial c

onte

nt k

now

ledg

e is

inna

te a

ndth

at e

xper

ienc

e pl

ays

only

an

elab

orat

ing

or p

aram

eter

-set

ting

role

in d

evel

opm

ent.

Ext

endi

ng th

e D

omai

n of

the

Lea

rnab

leIn

ord

er to

bri

ng o

ut th

e w

ay in

whi

ch c

onne

ctio

nist

lear

ning

allo

ws

us to

ext

end

the

dom

ain

of th

e le

arna

ble,

and

to u

nder

scor

e its

rel

evan

ce to

bas

ic q

uest

ions

abou

t the

sou

rces

of

our

conc

eptu

al k

now

ledg

e, I

will

foc

us o

n an

exa

mpl

e us

edby

Kei

l (19

87)

in h

is r

esea

rch

on c

once

ptua

l dev

elop

men

t. K

eil c

onsi

ders

how

child

ren

com

e to

shi

ft th

eir

cate

gori

satio

n be

havi

our

from

an

earl

y re

lianc

e on

an e

nsem

ble

of c

hara

cter

istic

fea

ture

s to

a la

ter

relia

nce

on c

ore

rela

tiona

lpr

oper

ties.

Am

ong

the

exam

ples

he

uses

is th

e us

e of

kin

ship

term

s su

ch a

sun

cle

. He

repo

rts

that

chi

ldre

n sh

ift f

rom

an

initi

al a

ccep

tanc

e of

des

crip

tions

such

as

"a f

ello

w w

ho w

as n

ot r

elat

ed to

any

one

in y

our

fam

ily b

ut w

as a

big

pal o

f you

r da

d's

and

brou

ght y

ou p

rese

nts

on y

our

birt

hday

and

Chr

istm

as"

toa

will

ingn

ess

to a

ccep

t des

crip

tions

suc

h as

"a

two

year

old

who

was

you

r m

ombr

othe

r. K

eil s

tres

ses

that

the

deve

lopm

enta

l pat

tern

indi

cate

s th

at th

e ch

ild'

conc

eptu

al d

evel

opm

ent r

efle

cts

the

emer

genc

e or

a "r

elat

iona

l sys

tem

" re

leva

ntto

the

spec

ific

dom

ain

of k

now

ledg

e, r

athe

r th

an p

iece

mea

l acq

uisi

tion

ofkn

owle

dge

abou

t, sa

y, u

ncle

s bu

t not

aun

ts, o

r an

y ki

nd o

f do

mai

n ge

nera

lch

ange

in o

vera

ll ap

proa

ch,

Page 6: Nurture in Development: A Parallel Distributed …stanford.edu/~jlmcc/papers/McClelland94.pdfThe Interaction of Nature and Nurture in Development: A Parallel Distributed Processing

McC

LE

LL

AN

D4. NATURE AND NURTURE

IN D

EV

EL

OPM

EN

T

But

ther

e is

som

ethi

ng v

ery

wro

ng w

ith a

ll th

ese

argu

men

ts. A

lthou

ghco

nnec

tioni

st m

odel

s ar

e si

mila

r in

som

e w

ays

to a

ssoc

iatio

nist

mod

els,

and

alth

ough

they

are

ext

ensi

ons

of e

xist

ing

stat

istic

al te

chni

ques

, the

re is

a c

ruci

aldi

ffer

ence

. In

conn

ectio

nist

mod

els,

the

very

bas

is o

f de

term

inin

g w

hat i

s si

mila

rto

wha

t is

part

of

wha

t is

disc

over

ed in

the

cour

se o

f le

arni

ng. T

his

is n

ot ju

st a

mat

ter

of e

mph

asis

ing

som

e di

men

sion

s an

d de

-em

phas

isin

g ot

hers

(al

thou

ghthis is what happens in some cases), but of

tota

lly r

ecod

ing

each

inpu

t so

as to

map

it in

to a

sim

ilari

ty s

pace

who

se s

truc

ture

depe

nds,

not on the surface

properties of the inputs, but

on th

e de

man

ds th

at a

re m

ade

by th

e la

sk o

fpredicting outcomesfrom inputs.

Thi

s se

nsiti

vity

of

the

inte

rnal

rep

rese

ntat

ions

to th

e ta

sk th

ey a

re r

equi

red

to p

erfo

rm d

ram

atic

ally

incr

ease

s th

e ra

nge

of w

hat

such

mod

els

are

able

to a

chie

ve c

ompa

red

to a

ssoc

iatio

nist

mod

els

or s

tand

ard

stat

istic

al te

chni

ques

.T

his

poin

t has

bee

n ill

ustr

ated

man

y tim

es in

con

nect

ioni

st le

arni

ng r

esea

rch.

The

impo

rtan

ce o

f co

nstr

uctin

g re

pres

enta

tions

was

the

them

e of

Rum

elha

rt e

tal

. (I9

86b)

and

alm

ost e

very

inte

rest

ing

appl

icat

ion

of c

onne

ctio

nist

lear

ning

sinc

e th

eir

pape

r ha

s m

ade

esse

ntia

lly th

is p

oint

. The

exa

mpl

e I

have

cho

sen

tous

e he

re is

one

of

the

firs

t to

dem

onst

rate

this

poi

nt, b

y H

into

n (1

986)

. It h

appe

nsto

use

an

exam

ple

that

rel

ates

spe

cifi

cally

to K

eil's

kin

ship

exa

mpl

es, a

s is

ther

efor

e pa

rtic

ular

ly a

ppro

pria

te. I

spe

nd s

ome

time

revi

ewin

g th

is e

xam

ple

here

beca

use

its p

oint

app

ears

to h

ave

been

lost

on

man

y ve

ry im

port

ant r

esea

rche

rsan

d co

mm

enta

tors

.H

into

n tr

aine

d a

conn

ectio

nist

net

wor

k to

ans

wer

que

stio

ns a

bout

the

kins

hip

rela

tions

am

ong

two

grou

ps o

f in

divi

dual

s, o

ne c

alle

d E

nglis

h an

d th

e ot

her

calle

dIt

alia

n (F

ig. 4

.3).

With

in e

ach

grou

p, th

ere

wer

e pa

ralle

l set

s of

kin

ship

rel

atio

ns,

but t

here

wer

e no

rel

atio

ns b

etw

een

the

grou

ps. H

into

n tr

aine

d a

conn

ectio

nist

netw

ork

on th

e re

latio

ns a

mon

g th

ese

indi

vidu

als

in th

e fo

llow

ing

way

. The

task

of th

e ne

twor

k w

as to

take

que

ries

that

can

be

glos

sed

"Per

sonl

' s R

elat

ion

is" (e.

g. "

Col

ins Uncle is -

) an

d to

then

res

pond

with

the

corr

ect

com

plet

ion

or c

ompl

etio

ns o

f th

e pr

opos

ition

(Pe

rson

2). A

s sh

own

in F

ig. 4

.4he

set

up

two

pool

s of

inpu

t uni

ts, o

ne c

onsi

stin

g of

a u

nit f

or e

ach

poss

ible

filt

erof

the

Pers

on I

rol

e, a

nd th

e ot

her

cons

istin

g of

a u

nit f

or e

ach

of th

e ki

nshi

pre

latio

ns. T

his

allo

wed

him

' to

pres

ent q

uerie

s by

sim

ply

activ

atin

g th

eap

prop

riat

e Pe

rson

I u

nit a

nd th

e ap

prop

riat

e R

elat

ion

unit.

He

also

set

up

a se

tof

out

put u

nits

, for

Per

son2

, aga

in c

onsi

stin

g of

one

uni

t for

eac

h pe

rson

, to

allo

wth

e ne

twor

k to

act

ivat

e a

unit

corr

espo

ndin

g to

the

indi

vidu

al o

r in

divi

dual

s at

the

othe

r en

d of

the

rela

tion.

In

betw

een

ther

e w

ere

sepa

rate

gro

ups

of h

idde

nun

its, o

ne f

or th

e ne

twor

k to

use

for

an

inte

rnal

representation of PI, one to

repr

esen

t the

rel

atio

n, o

ne to

rep

rese

nt th

e co

mbi

natio

n of

P I

and

the

rela

tion

and

one

to r

epre

sent

P2.

The

impo

rtan

t poi

nt is

that

, at t

he in

put a

nd o

utpu

t lev

els,

the

pers

ons

and

the

rela

tions

are

eac

h re

pres

ente

d by

dis

tinct

inpu

t uni

ts. I

nitia

lly. t

he c

onne

ctio

nsfr

om th

e in

put u

nits

to th

e co

rres

pond

ing

repr

esen

tatio

n un

its a

re r

ando

m, s

o

Chr

isto

pher

= P

enel

ope

And

rew

= C

hris

tine

Margaret = Arthur V

icto

ria =

Jam

es Je

nnif

er =

Cha

rles

Col

inC

harl

otte

Rob

erto

= M

arla

Pie

rro

= F

ranc

esca

Gina = Emilio Lu

cia

= M

arco

Ang

ela

= T

omas

o

Alf

onso

Soph

ia

FIG

. 4.

The

fam

ily tr

ees

unde

rlyi

ng th

e tr

aini

ng c

orpu

s us

ed b

y H

into

n (J

989

). T

he s

ymbo

l "m

eans

"married 10

. From

Para

llel D

istr

ibut

ed P

roce

ssin

g: I

mpl

icat

ions

for

psy

chol

ogy

and

neur

obio

logy

(p

. 49)

, by

R,

M. M

orri

s. N

ew Y

ork:

Oxf

ord

Uni

vers

ily P

ress

. Cop

yrig

ht 1

989

M. M

orri

s. R

epro

duce

d by

pen

niss

ion.

Pers

on2

Person 1

Rel

atio

n

FIG

.4 The

netw

ork

used

by

Hin

ton

(198

6) to

lear

n th

e fa

mily

tree

s sh

own

in F

ig, 4

. 3. R

epro

duce

dby

pen

niss

ion.

that

the

sim

ilari

ty r

elat

ions

am

ong

the

initi

al in

tern

al r

epre

sent

atio

ns a

re a

rbitr

ary

and

unre

late

d to

the

task

. Dur

ing

the

cour

se o

f tr

aini

ng, t

he c

onne

ctio

n w

eigh

tsth

roug

hout

the

netw

ork

chan

ge. O

n ea

ch tr

aini

ng tr

ial,

the

netw

ork

uses

its

Page 7: Nurture in Development: A Parallel Distributed …stanford.edu/~jlmcc/papers/McClelland94.pdfThe Interaction of Nature and Nurture in Development: A Parallel Distributed Processing

4. NATURE AND NURTURE

IN D

EV

EL

OPM

EN

TM

cCL

EL

LA

ND

exis

ting

conn

ectio

n w

eigh

ts to

pre

dict

the

corr

ect c

ompl

etio

n or

com

plet

ions

;th

e ou

tput

is c

ompa

red

to th

e co

rrec

t res

pons

e; a

nd th

e co

nnec

tions

are

adj

uste

dto

red

uce

the

diff

eren

ce u

sing

bac

k-pr

opag

atio

n, T

he w

eigh

ts f

rom

the

Pers

on I

inpu

t uni

ts to

the

Pers

on I

rep

rese

ntat

ion

units

det

erm

ine

how

eac

h pe

rson

isre

pres

ente

d. T

he r

epre

sent

atio

ns c

aptu

re w

hat m

ight

be

calle

d a

theo

ry o

f th

eun

derl

ying

str

uctu

re b

ehin

d th

e se

t of

spec

ific

pro

posi

tions

on

whi

ch th

e ne

twor

kw

as tr

aine

d. T

hey

repr

esen

t as

sim

ilar

thos

e in

divi

dual

s w

ho p

lay

sim

ilar

role

s,w

ithin

and

acr

oss

the

two

fam

ily tr

ees

show

n in

Fig

. 4.

We

can

try

to v

isua

lise

the

disc

over

ed s

imila

rity

str

uctu

re2

by im

agin

ing

the

patte

rns

of a

ctiv

atio

n on

thes

e pe

rson

-rep

rese

ntat

ion

units

as

poin

ts in

a s

ix-

dim

ensi

onal

spa

ce (

whe

re e

ach

dim

ensi

on c

orre

spon

ds to

the

activ

atio

n of

one

of th

e si

x un

its).

Sin

ce s

uch

spac

es a

re d

iffi

cult

to v

isua

lise.

and

sin

ce u

nits

inthe network tend to be used redundantly anyway, w

e ca

n us

e pr

inci

pal

com

pone

nts

anal

ysis

to p

ick

out t

he im

port

ant d

imen

sion

s. W

hen

we

do th

is, w

efi

nd th

at th

ere

are

thre

e st

rong

com

pone

nts.

We

can

then

plo

t the

indi

vidu

als

ina

thre

e-di

men

sion

al s

pace

, rep

rese

ntin

g w

here

eac

h in

divi

dual

fal

ls w

ith r

espe

ctto

eac

h of

thes

e th

ree

com

pone

nts.

The

thre

e-di

men

sion

al s

truc

ture

is s

till

som

ewha

t dif

ficu

lt to

gra

sp in

the

proj

ecte

d vi

ew p

rese

nted

in F

ig, 4

.5, b

ut it

can

be b

roug

ht o

ut b

y us

ing

the

fact

that

the

repr

esen

tatio

ns o

f th

e in

divi

dual

s

can

be s

epar

ated

per

fect

ly b

y a

plan

e th

roug

h th

e sp

ace

show

n in

Fig

. 4. 5

that

leav

es a

ll of

the

Ital

ians

on

one

side

of

the

plan

e, a

nd a

ll of

the

Eng

lish

on th

eot

her.

The

dim

ensi

on p

erpe

ndic

ular

to th

is p

lane

can

be

calle

d th

e " n

atio

nalit

ydi

men

sion

. The

net

wor

k di

scov

ers

this

dim

ensi

on b

ecau

se it

pla

ys a

ver

y st

rong

role

in r

estr

ictin

g th

e po

ssib

le c

ompl

etio

ns o

f a

quer

y, s

ince

Per

son

I an

d Pe

rson

2al

way

s ha

ve th

e sa

me

natio

nalit

y.W

e ca

n th

en lo

ok a

t the

pro

ject

ion

of th

e E

nglis

h an

d th

e It

alia

n gr

oups

ont

oth

e se

para

ting

plan

e, a

nd w

hat w

e se

e is

sho

wn

in F

ig. 4

. 6. F

or b

oth

the

Eng

lish

and

the

Ital

ians

, the

old

er g

ener

atio

n (s

quar

es)

is a

t the

top

of th

e pl

ane ,

the

youn

ger

gene

ratio

n (X

' s)

at th

e bo

ttom

, and

the

mid

dle

gene

ratio

n (t

rian

gles

) is

in b

etw

een.

Als

o, f

or b

oth

the

Eng

lish

and

Ital

ians

, the

indi

vidu

als

in th

e le

ftbr

anch

es o

f th

e tr

ees

show

n in

Fig

. 4. 3

(op

en s

ymbo

ls)

are

loca

ted

to th

e le

ft o

fth

e pl

ane,

and

the

indi

vidu

als

in th

e ri

ght b

ranc

hes

in F

ig. 4

.3 (

fille

d sy

mbo

ls)

are

loca

ted

to th

e ri

ght.

In g

ener

al, w

ithin

gen

erat

ions

, cor

resp

ondi

ng in

divi

dual

s

Eng

lish

Fam

ily T

ree

Rec

over

ed

0 0

Italia

n' .

English' +

+ +

(It

Ital

ian

Fam

ily T

ree

Rec

over

ed

-0.

-0.

-0.

-0.

-O.

-0.

-0.

-0,

-0.

.... 0 0

(QI

FIG

. 4.

The

inte

rnal

rep

rese

ntat

ions

of

the

indi

vidu

als

in th

e tw

o fa

mily

tree

s, d

ispl

ayed

as

poin

ts in

a m

ullid

imen

sion

al s

pace

.

FIG

.P

roje

ctio

ns o

f the

rep

rese

ntat

ions

sho

wn

in F

ig. 4

, 5 o

nto

the

plan

e th

at s

epar

ates

the

Eng

lish

from

the

lIal

ians

. See

text

for

exp

lana

tion.

, The

pal

lern

s us

ed f

or th

e an

alys

is d

iscu

ssed

in th

is a

nd th

e fo

llow

ing

para

grap

hs w

ere

kind

lysu

pplie

d by

Geo

ff H

into

n. w

ho r

e-ra

n th

e ne

twor

k de

scri

bed

in h

is 1

986

pape

r to

reg

ener

ate

the

data

.

Page 8: Nurture in Development: A Parallel Distributed …stanford.edu/~jlmcc/papers/McClelland94.pdfThe Interaction of Nature and Nurture in Development: A Parallel Distributed Processing

McC

LE

LL

AN

D4. NATURE AND NURTURE

IN D

EV

EL

OPM

EN

T

in th

e tw

o tr

ees

occu

py c

orre

spon

ding

loca

tions

, with

onl

y sl

ight

dif

fere

nces

betw

een

the

posi

tion

of e

ach

Eng

lish

and

the

corr

espo

ndin

g It

alia

n.T

he p

oint

of

revi

ewin

g th

is d

emon

stra

tion

is to

mak

e cl

ear

that

the

netw

ork

has

lear

ned

to tr

eat a

s si

mila

r, b

oth

with

in a

nd b

etw

een

the

two

fam

ily tr

ees,

thos

ein

divi

dual

s w

ho o

ccup

y th

e m

ost s

imila

r po

sitio

ns in

eac

h tr

ee. T

he r

easo

nin

divi

dual

s at

sim

iliar

loca

tions

in th

e tr

ee a

re g

iven

sim

ilar

repr

esen

tatio

ns is

that

thes

e re

pres

enta

tions

are

use

ful f

or a

llow

ing

the

netw

ork

to s

olve

the

prob

lem

pose

d by

the

quer

y-co

mpl

etio

n ta

sk, s

ince

the

com

plet

ions

of

prop

ositi

ons

invo

lvin

gin

divi

dual

s lo

cate

d at

sim

ilar

plac

es in

the

fam

ily tr

ees

occu

py s

imila

r tr

ee p

ositi

ons.

Giv

en th

e us

e of

this

bas

is f

or a

ssig

ning

sim

ilari

ty to

the

Pers

on I

inpu

ts, t

hene

twor

k ca

n gc

..ner

alis

e pr

oper

ly to

cas

es th

at w

ere

not p

rese

nted

in tr

aini

ng. I

nth

is c

ase,

eig

ht o

f th

e re

latio

ns w

ere

held

bac

k fr

om th

e tr

aini

ng c

orpu

s an

d on

lyap

pear

ed d

urin

g te

stin

g. T

he n

etw

ork

used

the

repr

esen

tatio

ns it

had

der

ived

for

each

indi

vidu

al d

urin

g th

e co

urse

of

trai

ning

to g

ener

alis

e pr

oper

ly a

t the

tim

eof

test

. It w

as a

ble

to d

o th

is in

a v

ery

sim

ple

way

. The

nat

iona

lity

dim

ensi

on o

fth

e re

pres

enta

tion

of P

erso

n I

dete

rmin

ed th

e na

tiona

lity

of th

e re

spon

se te

rm.

Oth

erw

ise,

the

repr

esen

tatio

ns o

f co

rres

pond

ing

indi

vidu

als

in th

e di

ffer

ent

fam

ily tr

ees

wer

e id

entic

al. T

here

fore

, wha

t it l

earn

ed a

bout

an

indi

vidu

al in

one

tree

tran

sfer

red

to in

divi

dual

s in

the

othe

r.In

sho

rt, t

he n

etw

ork

lear

ns b

y co

ntig

uity

and

gen

eral

ises

by

sim

ilari

ty, b

ut it

diff

ers

in a

cru

cial

res

pect

fro

m a

ssoc

iatio

nist

mod

els

and

stan

dard

sta

tistic

alte

chni

ques

. Unl

ike

thes

e m

echa

nism

s, it

ass

igns

the

sim

ilari

ty r

elat

ions

am

ong

the

patte

rns

as p

art o

f w

hat i

t lea

rns

thro

ugh

the

use

of a

sim

ple

grad

ual a

ndin

crem

enta

l pro

cedu

re fo

r ad

just

ing

the

stre

ngth

s of

the

conn

ectio

ns a

mon

g un

its.

One

of

the

key

feat

ures

of

wha

t hap

pens

in th

e process of assigning

repr

esen

tatio

ns is

the

crea

tion

of c

ateg

orie

s-gr

oups

of p

atte

rns

cont

aini

ngsi

mila

r m

embe

rs-a

nd d

istin

ctiv

e fe

atur

es th

at d

iffe

rent

iate

mem

bers

of

diff

eren

tca

tego

ries

. As

in th

e fa

mily

tree

s ex

ampl

e, s

uch

cate

gori

es s

uppo

rt g

ener

alis

atio

n.E

lman

(19

90)

cons

ider

ed th

is is

sue

as it

rel

ates

to s

ynta

x an

d S

t. Jo

hn a

ndM

cCle

lland

(19

90; M

cCle

lland

, St.

John

, & T

arab

an, 1

989)

con

side

red

this

mat

ter

in a

con

nect

ioni

st m

odel

of

sent

ence

com

preh

ensi

on. H

uman

lang

uage

user

s ca

n pa

rse

unfa

mili

ar b

ut g

ram

mat

ical

sen

tenc

es b

ecau

se th

ey k

now

the

cate

gori

es a

nd s

ubca

tego

ries

' of n

ouns

and

ver

bs. S

o ca

n co

nnec

tioni

st n

etw

orks

that

dis

cove

r th

ese

cate

gori

es th

roug

h gr

adua

l lea

rnin

g. I

n th

e St

. Joh

n an

dM

cCle

lland

cas

e, a

s in

Hin

ton

s ne

twor

k, it

is th

e di

scov

ered

sim

ilari

ty s

truc

ture

,ba

sed

on th

e ro

les

each

wor

d pl

ays

in c

onst

rain

ing

the

inte

rpre

tatio

n of

the

mea

ning

of

vari

ous

sent

ence

s, th

at g

over

ns g

ener

alis

atio

n, r

athe

r th

an a

ny s

urfa

cesi

mila

rity

am

ong

part

icul

ar in

put p

atte

rns.

The

fac

t the

con

nect

ioni

st m

odel

s di

scov

er th

eir

own

inte

rnal

rep

rese

ntat

ions

,an

d th

e fa

ct th

at th

ese

repr

esen

tatio

ns c

an s

erve

as

a pr

inci

pled

bas

is f

or n

on-t

rivi

alge

nera

lisat

ions

, app

ears

to h

ave

been

igno

red

in d

iscu

ssio

ns, s

uch

as th

e on

e by

Keil (l99Ia), where the point is

to m

otiv

ate

a na

tivis

t app

roac

h by

app

ealin

g to

the insufficiency of experience-ba

sed

appr

oach

es. T

he f

act t

hat a

ssoc

iatio

nist

mod

els,

and

not

con

nect

ioni

st m

odel

s, a

re d

iscu

ssed

by

Kei

l , s

impl

y in

dica

tes

that

the

thin

king

und

erly

ing

nativ

ist a

ppro

ache

s is

bas

ed in

par

t on

an in

suff

icie

ntap

prec

iatio

n of

the

pow

er o

f w

hat c

an b

e ac

com

plis

hed

by th

e ne

wer

and

mor

epo

wer

ful m

echa

nism

s of

lear

ning

that

are

ava

ilabl

e in

the

conn

ectio

nist

fra

mew

ork.

I ha

ve c

hose

n to

foc

us o

n K

eil's

pre

sent

atio

n of

thes

e is

sues

bec

ause

of

his

clar

ity in

layi

ng h

is p

ositi

on o

ut a

nd b

ecau

se o

f th

e sp

ecif

ic r

elev

ance

of

Hin

ton

dem

onst

ratio

n to

his

ana

lysi

s of

the

acqu

isiti

on o

f ki

nshi

p kn

owle

dge.

But

the

poin

t is

a ve

ry g

ener

al o

ne, a

nd I

bel

ieve

that

ther

e ha

s be

en a

fai

lure

in th

e w

ork

of o

ther

key

aut

hors

(am

ong

them

Fod

or a

nd P

utna

m)

to a

ttend

to th

e di

ffer

ence

sbe

twee

n co

nnec

tioni

st le

arni

ng m

echa

nism

s an

d pr

e-ex

istin

g id

eas

such

as

asso

ciat

ioni

st m

odel

s or

sta

ndar

d st

atis

tical

tech

niqu

es. T

his

is u

nfor

tuna

te o

ntw

o gr

ound

s. O

ne is

the

poin

t alr

eady

mad

e , th

at c

onne

ctio

nist

mod

els

are

far

mor

e po

wer

ful t

han

asso

ciat

ioni

st m

odel

s or

sta

ndar

d st

atis

tical

tech

niqu

es. T

heot

her

is th

at th

e us

e of

con

nect

ioni

st m

odel

s do

es n

ot e

ntai

l the

cla

im o

ften

attributed to associationists that the organism starts as a

tabula rasa

with

not

hing

mor

e th

an d

omai

n-in

depe

nden

t pri

ncip

les

to g

uide

it. O

n th

e co

ntra

ry, a

s I

have

alre

ady

indi

cate

d, it

is q

uite

eas

y to

inco

rpor

ate

dom

ain-

spec

ific

var

iatio

n. T

his

allows an exploration of the full range of possibilities, ranging from pure

tabu

lara

sa

formulations, to strongly pr

e-st

ruct

ured

acc

ount

s. T

he a

ppea

l of

conn

ectio

nist

mod

els

is n

ot th

at th

ey c

omm

it th

e re

sear

cher

to o

ne f

orm

ulat

ion

or th

e ot

her ,

but

that

they

pro

vide

the

oppo

rtun

ity to

exp

lore

a w

ide

rang

e of

poss

ible

acc

ount

s of

the

role

s of

nat

ure

and

nurt

ure

in d

evel

opm

ent.

Whi

le it

seem

s to

me

we

can

take

the

Hin

ton

wor

k as

rai

sing

que

stio

ns a

bout

the

need

for

inna

te d

omai

n-sp

ecif

ic k

now

ledg

e of

rel

evan

ce to

suc

h th

ings

as

kins

hip

rela

tions

, I w

ould

not

wan

t to

sugg

est t

hat i

t pro

ves

the

suff

icie

ncy

of a

bla

nket

dom

ain-

gene

ral a

ppro

ach.

Aft

er a

ll, t

here

is c

onsi

dera

ble

pre-

stru

ctur

ing

ofH

into

ns

arch

itect

ure

for

the

disc

over

y of

the

rele

vant

rel

atio

nal i

nfor

mat

ion.

The

ext

ent t

o w

hich

this

pre

-str

uctu

ring

is c

ruci

al is

not

yet

kno

wn.

The

Tim

e-co

urse

of

Dev

elop

men

tal C

hang

eA

noth

er d

omai

n no

w a

vaila

ble

for

expl

orat

ion

is th

e st

udy

of th

e tim

e-co

urse

of

chan

ge. I

n th

is d

omai

n, o

ne q

uest

ion

we

may

con

side

r is

, why

is d

evel

opm

ent

so s

low

? A

s Fl

avel

l (19

63, p

. 49)

put

in h

is b

ook

abou

t Pia

get,

" Wha

t pre

vent

sth

e or

gani

sm f

rom

mas

teri

ng in

one

fel

l sw

oop

all t

hat i

s co

gniz

able

in a

giv

ente

rrai

n?"

The

sam

e qu

estio

n ca

n be

ask

ed o

f co

nnec

tioni

st n

etw

orks

. Her

e, in

case

s lik

e th

e H

into

n ex

ampl

e , w

e ca

n se

e on

e re

ason

why

lear

ning

is g

radu

al.

The

rea

son

is th

at th

e ne

twor

k m

ust i

nteg

rate

the

chan

ges

it m

akes

to it

sco

nnec

tions

ove

r a

suff

icie

nt s

ampl

e of

the

envi

ronm

ent,

so th

at th

e ch

ange

s ca

nbe

gui

ded

by th

e st

ruct

ure

that

is p

rese

nt in

it. I

t is

only

whe

n th

e ch

ange

s oc

cur

slow

ly e

noug

h th

at th

eir

over

all d

irec

tion

is g

over

ned

by th

e st

ruct

ure

of th

een

viro

nmen

t tha

t the

abi

lity

to r

epre

sent

that

str

uctu

re c

an e

mer

ge.

Page 9: Nurture in Development: A Parallel Distributed …stanford.edu/~jlmcc/papers/McClelland94.pdfThe Interaction of Nature and Nurture in Development: A Parallel Distributed Processing

McC

LE

LL

AN

D4. NATURE AND NURTURE

IN D

EV

EL

OPM

EN

T

But

now

a p

uzzl

e se

ems

to a

rise.

If d

evel

opm

ent i

s a

slow

and

gra

dual

pro

cess

,w

hy is

it th

at it

oft

en a

ppea

rs to

be

mar

ked

by lo

ng p

erio

ds o

f st

atis

, pun

ctua

ted

by b

rief

per

iods

of

rela

tivel

y ra

pid

chan

ge?

Wha

t acc

ount

s, in

Fla

vell'

s w

ords

,

for

its "velocity and acceleration" (Flavell, 1

963,

p. 4

9)?

How

is it

, mor

esp

ecif

ical

ly, t

hat i

ncre

men

tal m

echa

nism

s gi

ve r

ise

to a

ppar

ently

qua

litat

ive

chan

ge?

It is

cru

cial

to d

istin

guis

h be

twee

n th

e se

nse

of s

tage

s in

tend

ed in

this

cha

pter

and

the

sens

e of

sta

ges

that

com

es f

rom

the

deve

lopm

enta

l the

ory

of P

iage

t. In

this

cha

pter

, sta

ges

are

only

mea

nt to

ref

er to

per

iods

of

rela

tive

stat

is w

ithin

dom

ains

, and

sta

ge tr

ansi

tions

are

mea

nt a

s re

fere

nces

to th

e tr

ansi

tions

bet

wee

nth

ese

dom

ain-

spec

ific

fea

ture

s of

the

deve

lopm

enta

l pro

file

. Thi

s co

ncep

tion

ofst

ages

con

tras

ts s

harp

ly w

ith th

e Pi

aget

ian

conc

eptio

n of

bro

ad s

tage

s cu

tting

acro

ss a

ll co

gniti

ve d

omai

ns. I

t may

be

that

Pia

get

s in

sist

ence

on

the

cent

ralit

yof

thes

e br

oad

stag

es c

ontr

ibut

ed to

the

diff

icul

ty in

und

erst

andi

ng h

ow h

isvi

ews

of th

e so

urce

s of

cog

nitiv

e ch

ange

-acc

omm

odat

ion

and

assi

mila

tion-

coul

d gi

ve r

ise

to c

ogni

tive

deve

lopm

ent.

In a

ny c

ase.

the

ques

tion

of h

ow a

nd w

hy d

omai

n-sp

ecif

ic s

tage

tran

sitio

nsoc

cur

rem

ains

puz

zlin

g fo

r ex

peri

ence

-bas

ed th

eori

es o

f al

l typ

es. I

f w

e ta

ke a

nin

crem

enta

l vie

w, i

t is

not a

t all

obvi

ous

why

we

shou

ld s

ee a

ppar

ent s

tabi

lity

for

long

per

iods

at a

ll, e

spec

ially

if, a

s of

ten

happ

ens,

thes

e lo

ng p

erio

ds o

fst

abili

ty a

re f

ollo

wed

by

rela

tivel

y ab

rupt

tran

sitio

ns to

a n

ew p

late

au.

One

res

pons

e to

sta

ge tr

ansi

tions

has

bee

n to

arg

ue th

at th

ey d

o no

t rea

llyex

ist-

that

they

rep

rese

nt, p

erha

ps, t

he c

rude

ness

of

our

met

hods

for

ass

essi

ngch

ilden

s ca

paci

ties

at p

artic

ular

poi

nts

in d

evel

opm

ent.

Ano

ther

app

roac

h ha

sbe

en to

thin

k of

them

as

aris

ing

from

som

e m

atur

atio

nal p

roce

ss. O

n th

is v

iew

,th

e ch

ild p

rogr

esse

s fr

om s

tage

to s

tage

bas

ed o

n so

me

exog

enou

s pr

oces

s, m

uch

as a

but

terf

ly p

rogr

esse

s fr

om s

tage

to s

tage

. Bot

h of

thes

e ki

nds

of th

ings

may

be p

art o

f th

e st

ory.

In

addi

tion,

how

ever

, ins

ight

s fr

om th

e st

udy

ofle

arni

ng in

conn

ectio

nist

mod

els

may

als

o sh

ed s

ome

light

on

the

mat

ter.

I ex

amin

ed th

is is

sue

in a

stu

dy o

f de

velo

pmen

tal c

hang

e in

the

bala

nce

scal

etask introduced by Inhelder and Piaget (1958), and subsequently studied

exte

nsiv

ely

by S

iegl

er (

e.g.

197

6, 1

981)

and

man

y ot

hers

. The

task

, qui

te s

impl

y,

is to

look

at b

alan

ce s

cale

pro

blem

s of

the

kind

sho

wn

in F

ig, 4

.7, a

nd to

indi

cate

whi

ch s

ide

will

go

dow

n. T

he s

peci

fic

prob

lem

con

sist

s of

pro

vidi

ng s

ome

num

ber

of u

nit w

eigh

ts o

n ea

ch s

ide,

with

the

wei

ghts

all

on o

ne p

eg, s

ome

num

ber

of s

teps

fro

m th

e fu

lcru

m. S

iegl

er id

entif

ied

seve

ral s

tage

s in

the

deve

lopm

ent

of th

e ab

ility

to s

olve

this

kin

d of

pro

blem

. In

wha

t I w

ill c

all s

tage

0 (

typi

cal o

fch

ildre

n be

low

the

age

of 5

), h

e fo

und

that

chi

ldre

n te

nded

to p

erfo

rm r

ando

mly

in th

e ta

sk. I

n st

age

I, c

hild

ren

resp

onde

d to

the

num

ber

of w

eigh

ts o

n ea

ch s

ide,

but i

gnor

ed d

ista

nce

from

the

fulc

rum

. In

stag

e 2,

chi

ldre

n to

ok d

ista

nce

into

acco

unt o

nly

whe

n th

e w

eigh

ts o

n bo

th s

ides

wer

e th

e sa

me.

In

stag

e 3,

they

took

both

dis

tanc

e an

d w

eigh

t int

o ac

coun

t, bu

t beh

aved

inco

nsis

tent

ly w

hen

thes

etw

o cu

es w

ere

plac

ed in

con

flic

t. St

age

4, w

hich

is n

ot a

lway

s ac

hiev

ed e

ven

by

FIG

. 4.7 A

bala

nce

scal

e of

the

kind

stu

died

by

Sieg

ler

(197

6, 1

981)

. Rep

rodu

ced

by p

enni

ssio

n.A

cade

mic

Pre

ss. (

nc,

adul

ts, r

equi

res

mul

tiply

ing

wei

ght t

imes

dis

tanc

e w

hen

wei

ght i

s gr

eate

r on

one

side

and

the

dist

ance

is g

reat

er o

n th

e ot

her.

The

mai

n in

tere

st f

or n

ow is

in th

etr

ansi

tions

thro

ugh

the

earl

y st

ages

, par

ticul

arly

into

and

out

of

stag

e I,

and

inth

e fa

ct th

at c

hild

ren

typi

cally

sta

y in

sta

ge I

for

sev

eral

yea

rs. T

he q

uest

ion

isho

w is

it th

at th

is s

tage

- lik

e ch

arac

ter

of d

evel

opm

ent c

an a

rise

fro

m a

nin

crem

enta

l lea

rnin

g m

echa

nism

?In

fac

t, th

is k

ind

of s

tage

- lik

e pr

ogre

ssio

n is

exa

ctly

wha

t we

typi

cally

see

inco

nnec

tioni

st le

arni

ng m

odel

s, T

he p

heno

men

on r

equi

res

a m

ulti-

laye

rne

twor

k--o

ne c

onsi

stin

g of

at l

east

one

laye

r of

hid

den

units

bet

wee

n in

put a

ndou

tput

. In

the

sim

ples

t pro

blem

s th

at r

equi

re a

two-

laye

r ar

chite

ctur

e-lik

e th

eex

clus

ive-

or p

robl

em, s

how

n in

Fig

. 4.

ther

e ar

e tw

o st

ages

, In

this

pro

blem

,it

is s

tand

ard

to u

se a

net

wor

k co

nsis

ting

of tw

o in

put u

nits

, tw

o hi

dden

uni

tsan

d on

e ou

tput

, and

the

task

the

netw

ork

face

s is

to le

arn

that

the

outp

ut u

nit

shou

ld b

e on

if e

ither

inpu

t uni

t is

on, b

ut s

houl

d be

off

if n

eith

er o

r bo

th a

re o

n.In

trai

ning

suc

h ne

twor

ks w

ith m

ultip

le s

wee

ps th

roug

h th

e se

t of

trai

ning

exam

ples

, we

see

an e

arly

sta

ge in

whi

ch le

arni

ng p

rogr

esse

s ve

ry s

low

ly, a

ndth

e ne

twor

k is

per

fonn

ing

very

poo

rly,

pro

duci

ng th

e sa

me

resp

onse

in e

very

case

-an

activ

atio

n of

0. 5

. The

re th

en fo

llow

s a

rapi

d tr

ansi

tion

, in

whi

chpe

rfon

nanc

e ra

pidl

y im

prov

es, u

ntil

it le

vels

off

at a

hig

h le

vel o

f ac

cura

cy, i

nw

hich

the

outp

ut o

f th

e ne

twor

k ap

proa

ches

the

righ

t ans

wer

for

all

four

trai

ning

case

s. I

n m

ore

com

plex

cas

es, w

here

som

e of

the

trai

ning

exa

mpl

es a

re e

asie

rto master than others, there can be multiple plateaus

, aga

in p

unct

uate

d by

rela

tivel

y ra

pid

tran

sitio

ns.

The

beh

avio

ur in

the

sim

ples

t cas

es c

an b

e un

ders

tood

in th

e fo

llow

ing

way

.E

arly

on,

bot

h in

put a

nd o

utpu

t con

nect

ions

in th

e ne

twor

k ar

e ra

ndom

. A c

hang

eto

a c

onne

ctio

n w

eigh

t on

the

inpu

t sid

e of

the

netw

ork

has

no u

sefu

l eff

ect,

sinc

eth

e ou

tput

con

nect

ions

are

ran

dom

; and

a c

hang

e to

a c

onne

ctio

n w

eigh

t on

the

outp

ut s

ide

of th

e ne

twor

k ha

s no

use

ful e

ffec

t, si

nce

the

inpu

t con

nect

ions

are

rand

om. G

radu

ally

, tho

ugh,

thro

ugh

the

cour

se o

f ex

peri

ence

, the

con

nect

ion

wei

ghts

beg

in to

bec

ome

coor

dina

ted.

The

inpu

t wei

ghts

beg

in to

pro

duce

use

ful

repr

esen

tatio

ns, a

nd th

e ou

tput

wei

ghts

beg

in to

pro

duce

use

ful p

redi

ctio

nsba

sed

on th

ese

repr

esen

tatio

ns, A

s th

is h

appe

ns, i

t tur

ns o

ut th

at s

mal

l cha

nges

to th

e w

eigh

ts b

egin

to p

rodu

ce b

igge

r an

d bi

gger

cha

nges

in p

erfo

nnan

ce.

Page 10: Nurture in Development: A Parallel Distributed …stanford.edu/~jlmcc/papers/McClelland94.pdfThe Interaction of Nature and Nurture in Development: A Parallel Distributed Processing

McC

LE

LL

AN

D

XOR Network

Inpu

tO

utpu

t

4. NATURE AND NURTURE

IN D

EV

EL

OPM

EN

T

used

to a

sses

s ch

ildre

ns

perf

onna

nce

at e

ach

poin

t alo

ng th

e w

ay. I

ran

the

netw

ork

seve

ral t

imes

, usi

ng d

iffe

rent

sta

rtin

g w

eigh

ts a

nd d

iffe

rent

ran

dom

sequ

ence

s of

trai

ning

tria

ls, a

nd f

ound

that

it c

aptu

red

seve

ral a

spec

ts o

fch

ildre

ns

perf

onna

nce

very

nic

ely.

In g

ener

al, I

foun

d th

at th

e si

mul

atio

nco

nfon

ned

to o

ne o

f Sie

gler

s st

ages

abo

ut 8

6% o

f th

e tim

e-in

chi

ldre

n, th

efi

gure

is a

roun

d 93

%. M

ore

grad

ual l

earn

ing

prod

uces

a h

ighe

r le

vel o

fco

nfon

nity

(Sc

hmid

t & S

chul

tz, 1

992)

. The

net

wor

k al

way

s pr

ogre

ssed

fro

mst

age

0 to

sta

ge I

, and

then

aft

er a

per

iod

of s

tabi

lity

it ex

hibi

ted

a no

isy

tran

sitio

nth

roug

h st

age

2 to

a s

tage

rep

rese

ntin

g so

met

hing

inte

nned

iate

bet

wee

n Si

egle

rst

ages

3 a

nd 4

. In

so d

oing

, it c

aptu

red

a nu

mbe

r of

asp

ects

of

the

data

, inc

ludi

ngso

me

of th

e fe

atur

es th

at in

dica

te th

at c

hild

ren

may

use

a c

ontin

uous

fun

ctio

n,ra

ther

than

dis

cret

e ru

les,

in m

akin

g th

eir

judg

emen

ts in

the

bala

nce

scal

e ta

sk(W

ilken

ing

& A

nder

son,

199

1; s

ee S

chm

idt &

Sch

ultz

, 199

2, a

nd M

cCle

lland

& Jenkins, 1991

, for

fur

ther

dis

cuss

ion)

.I

did

find

that

in o

rder

to s

imul

ate

the

resu

lts, i

t was

nec

essa

ry to

mak

e tw

oas

sum

ptio

ns. F

irst,

I ha

d to

ass

ume

that

wei

ght a

nd d

ista

nce

wer

e en

code

dse

para

tely

, as

show

n in

the

sim

ple

netw

ork

, and

com

bine

d on

ly a

fter

sep

arat

een

codi

ng. S

econ

d, I

had

to p

rovi

de-

som

e ba

sis

for

the

use

of th

e di

stan

ce c

ue to

deve

lop

mor

e sl

owly

than

the

use

of th

e w

eigh

t cue

. I g

ot th

e re

sults

just

sho

wn

usin

g th

e as

sum

ptio

n th

at w

eigh

t var

ies

mor

e of

ten

in th

e re

leva

nt e

xper

ienc

e

-""'

"

FIG

. 4.8 XOR:

The

net

wor

k. th

e in

put-

()ut

put p

atte

rns

and

the

time-

cour

se o

f ac

quis

ition

.

Dis

tonc

e

Whi

le it

see

ms

like

no p

rogr

ess

is b

eing

mad

e at

all

at f

irst

, in

fact

ther

e is

som

epr

ogre

ss a

ll al

ong.

The

ear

ly p

hase

s pr

oduc

e on

ly v

ery

smal

l cha

nges

in o

vert

perf

onna

nce,

but

they

are

nec

essa

ry to

bri

ng th

e ne

twor

k to

the

poin

t whe

re it

is c

apab

le o

f m

ore

rapi

d an

d no

ticea

ble

prog

ress

. Lat

er r

apid

cha

nge

depe

nds

upon

the

earl

ier,

som

etim

es a

lmos

t inf

inite

sim

ally

sm

all a

dapt

atio

ns.

In a

pply

ing

thes

e id

eas

in a

n ef

fort

to u

nder

stan

d de

velo

pmen

t in

the

bala

nce

scal

e pr

oble

m, w

e ca

n en

visi

on a

pro

gres

sion

in w

hich

the

kind

of

tran

sitio

n w

ese

e in

the

XO

R n

etw

ork

happ

ens,

firs

t for

one

cue

-the

wei

ght c

ue-a

nd th

enfo

r th

e di

stan

ce c

ue. I

ndee

d, i

t is

poss

ible

to s

imul

ate

the

prog

ress

ion

of s

tage

sin

just

this

way

(M

cCle

lland

, 198

9; M

cCle

lland

& J

enki

ns, 1

991;

Sch

mid

t &Sc

hultz

, 199

2). I

n th

e si

mul

atio

ns r

epor

ted

in M

cCle

lland

(19

89),

I s

et u

p a

netw

ork

in w

hich

the

inpu

ts w

ere

a re

pres

enta

tion

of a

bal

ance

sca

le p

robl

em,

indi

catin

g so

me

num

ber

of w

eigh

ts s

ome

num

ber

of s

teps

fro

m th

e fu

lcru

m o

nea

ch s

ide

(Fig

. 4.9

). T

he ta

sk w

as s

impl

y to

turn

on

one

of tw

o ou

tput

s, in

dica

ting

whi

ch s

ide

shou

ld g

o do

wn.

The

net

wor

k w

as tr

aine

d on

a r

ando

m s

eque

nce

ofba

lanc

e sc

ale

prob

lem

s, a

nd w

as te

sted

with

the

sam

e pr

oble

ms

that

Sie

gler

(19

81)

Wei

ght

0008

008

000

FIG

. 4.

The

con

nect

ioni

st n

etw

ork

used

to s

tudy

sta

ge tr

ansi

tions

by

McC

lella

nd (

1989

).R

epro

duce

d by

per

mis

sion

, Oxf

ord

Uni

vers

ity P

ress

, New

Yor

k.

Page 11: Nurture in Development: A Parallel Distributed …stanford.edu/~jlmcc/papers/McClelland94.pdfThe Interaction of Nature and Nurture in Development: A Parallel Distributed Processing

NA

TU

RE

AN

D N

UR

TU

RE

IN D

EV

ELO

PM

EN

T

Hid

den

to O

utpu

t

1\1

C)I

100

Epo

ch

Inpu

t to

Hid

den

~ N

(.)

C)I

100

Epo

ch

McC

LE

LL

AN

D

of th

e ch

ild th

an d

ista

nce

does

-in

the

sim

ulat

ions

, thi

s am

ount

s to

usi

ng m

ore

freq

uent

pre

sent

atio

ns o

f ca

ses

in w

hich

wei

ght d

iffe

rs o

n th

e tw

o si

des

of th

esc

ale

but d

ista

nce

is th

e sa

me,

by

anal

ogy

to c

hild

ren

s ex

peri

ence

with

see

-saw

s,w

here

dis

tanc

e fr

om th

e fu

lcru

m is

rel

ativ

ely

fixe

d by

the

situ

atio

n. M

ore

rece

ntly

, I h

ave

exam

ined

the

idea

that

dis

tanc

e is

a m

ore

com

plex

rel

atio

n th

anw

eigh

t, W

eigh

t is

(at l

east

in a

fix

ed g

ravi

tatio

nal f

ield

) a

one-

plac

e pr

edic

ate-

a pr

oper

ty o

f an

indi

vidu

al o

bjec

t. In

con

tras

t, di

stan

ce b

etw

een

two

obje

cts

isa

two-

plac

e pr

edic

ate,

a r

elat

ion

betw

een

an o

bjec

t and

ano

ther

obj

ect-in this

case

, the

wei

ght a

nd th

e fu

lcru

m. T

o ca

ptur

e th

is id

ea, I

use

d a

mor

e co

mpl

exin

put,

in w

hich

the

posi

tion

of th

e weights and of the fulcrum could vary

(McC

lella

nd, i

n pr

ess)

. Thi

s m

ade

it ne

cess

ary

for

the

netw

ork

to le

arn

to c

ompu

teth

e m

ore

com

plex

dis

tanc

e cu

e fr

om th

e si

mpl

er p

ositi

on c

ues.

In

that

cas

e, e

ven

if w

eigh

t and

dis

tanc

e va

ry e

qual

ly o

ften

in th

e tr

aini

ng e

xam

ples

, the

net

wor

kle

arns

the

wei

ght c

ue m

uch

mor

e qu

ickl

y, a

nd th

ere

is a

long

pha

se th

atco

rres

pond

s to

Sie

gler

s ru

le I

, fol

low

ed e

vent

ually

by

a tr

ansi

tion

to a

sta

gelik

e Si

egle

rs

rule

3, i

n w

hich

wei

ght a

nd d

ista

nce

are

both

con

side

red.

In th

e w

ork

with

the

sim

pler

rep

rese

ntat

ion

(McC

lella

nd, 1

989)

, it i

s ea

sy to

exam

ine

the

time-

cour

se o

f le

arni

ng a

bout

the

wei

ght a

nd d

ista

nce

cues

, in

term

s

of th

e st

reng

ths

of th

e co

nnec

tion

wei

ghts

into

and

out

of

the

hidd

en u

nits

(Fi

g.10

). F

or b

oth

cues

, we

see

roug

hly

the

sam

e pa

ttern

. Ini

tially

, the

net

wor

k is

inse

nsiti

ve to

the

cue

at b

oth

leve

ls. T

hen

ther

e is

a r

elat

ivel

y st

eep

tran

sitio

n,fo

llow

ed b

y a

leve

lling

off

, The

tran

sitio

n is

mor

e ra

pid

for

the

wei

ght c

ue th

anth

e di

stan

ce c

ue, b

ecau

se o

f di

ffer

entia

l fre

quen

cy o

f ex

posu

re to

var

iatio

ns in

the

two

cues

. The

re a

re a

ctua

lly m

any

empi

rica

l res

ults

sup

port

ing

the

idea

that

the

acqu

isiti

on o

f th

e di

stan

ce c

ue is

rel

ativ

ely

grad

ual i

n ch

ildre

n, a

s it

is in

the

mod

el (

e,g.

Wilk

enin

g &

And

erso

n, 1

991)

. The

key

poi

nt th

ough

is th

e no

n-ho

mog

enei

ty o

f th

e de

velo

pmen

tal p

roce

ss, b

oth

for

the

wei

ght c

ue a

nd th

edi

stan

ce c

ue. I

n bo

th c

ases

, the

re is

an

earl

y ph

ase

of n

o ov

ert c

hang

e, f

ollo

wed

by a

mor

e ra

pid

tran

sitio

n to

a n

ew le

vel.

Of

cour

se, t

his

sim

ulat

ion

vast

ly o

vers

impl

ifie

s th

e ch

alle

nges

a le

arne

r fa

ces

duri

ng d

evel

opm

ent.

Chi

ldre

n do

not

lear

n ab

out b

alan

ce s

cale

s in

isol

atio

nfr

om o

ther

kin

ds o

f ex

peri

ence

with

dis

tanc

e an

d w

eigh

t inf

orm

atio

n. D

ista

nces

and

wei

ghts

do

not c

ome

nice

ly p

repa

ckag

ed in

uni

t qua

ntiti

es, a

s th

ey d

o in

the

inpu

t to

the

netw

ork.

Nev

erth

eles

s, I

thin

k th

e m

odel

doe

s ca

ptur

e on

e im

port

ant

aspe

ct o

f de

velo

pmen

t, na

mel

y its

sta

ge-l

ike

char

acte

r in

man

y di

ffer

ent d

omai

ns.

In s

umm

ary,

then

, we

see

that

con

nect

ioni

st m

odel

s ca

n sh

ed s

ome

light

on

the

time-

cour

se o

f de

velo

pmen

t. T

he w

ork

does

not

trea

t the

pro

cess

as

one

that

ari

ses

stri

ctly

fro

m e

nvir

onm

enta

l inf

luen

ces.

Ind

eed,

the

mod

el il

lust

rate

s th

e ge

nera

lim

port

ance

' of ar

chite

ctur

al c

onst

rain

ts, s

ince

it is

cle

ar th

at th

e su

cces

s of

the

mod

elde

pend

s on

the

sepa

rate

enc

odin

g of

wei

ght a

nd d

ista

nce

info

rmat

ion.

Exp

erie

nce

is n

ot th

e on

ly f

acto

r th

at p

lays

a r

ole

in s

hapi

ng d

evel

opm

ent.

But

exp

erie

nce

isth

e en

gine

that

dri

ves

deve

lopm

ent f

orw

ard,

and

the

wor

k m

akes

it c

lear

how

sta

ge-

like

beha

viou

r ca

n em

erge

fro

m in

crem

enta

l lea

rnin

g m

echa

nism

s.

F~

G. 1

4.10

Tim

e-co

urse

of

acqu

isiti

on o

f co

nnec

tion

stre

ngth

s re

leva

nt to

the

role

of

wei

ght a

nddi

stan

ce in

the

bala

nce

scal

e si

mul

atio

n of

McC

lella

nd (

1989

), R

epro

duce

d by

pen

niss

ion.

Oxf

ord

Uni

vers

ity P

ress

, New

Yor

k.

It is

wor

th s

topp

ing

for

a m

omen

t to

note

how

muc

h th

e co

nnec

tioni

st a

ppro

ach

adva

nces

our

und

erst

andi

ng o

f th

e m

echa

nism

s of

dev

elop

men

tal c

hang

e, r

elat

ive

to p

ropo

sals

adv

ance

d, f

or e

xam

ple,

by

Piag

et. P

iage

t pre

sent

ed h

is th

inki

ng in

Page 12: Nurture in Development: A Parallel Distributed …stanford.edu/~jlmcc/papers/McClelland94.pdfThe Interaction of Nature and Nurture in Development: A Parallel Distributed Processing

4, NATURE AND NURTURE

IN D

EV

EL

OPM

EN

TM

cCL

EL

LA

ND

tenn

s of

the

idea

s of

equ

ilibr

atio

n, a

ccom

mod

atio

n an

d as

sim

ilatio

n (s

ee F

lave

ll,19

63, f

or a

con

cise

pre

sent

atio

n). T

hese

idea

s su

gges

t pro

cess

es o

f co

ntin

ual,

grad

ual a

dapt

atio

n, a

nd a

s su

ch c

aptu

re th

e ge

nera

l fac

t tha

t cog

nitiv

ede

velo

pmen

t is

a ve

ry s

low

pro

cess

, spa

nnin

g ve

ry m

any

year

s. B

ut th

ey d

o no

tre

ally

pro

vide

any

cle

ar o

r ex

plic

it ba

sis

for

unde

rsta

ndin

g w

hy th

ere

are

plat

eaus

in d

evel

opm

ent i

nter

sper

sed

with

per

iods

of

rapi

d ch

ange

. Per

haps

in p

art f

orth

is r

easo

n, e

xper

ienc

e-ba

sed

acco

unts

of

deve

lopm

enta

l cha

nge

have

tend

ed to

fall

from

fav

our,

with

man

y in

vest

igat

ors

pref

erri

ng to

pro

pose

inst

ead

dist

inct

lym

atur

atio

nal a

ccou

nts

(e.g. Case, 1978; Pascual-L

eone

, 198

7). M

y po

int i

sno

t tha

t the

se m

atur

atio

nal a

ccou

nts

are

wro

ng, b

ut th

at w

e ar

e no

w in

a p

ositi

onth

at a

llow

s ex

perie

nce-

base

d al

tern

ativ

es to

be

give

n an

exp

licit

mec

hani

stic

bas

isth

at a

llow

s fo

r fu

ller

and

mor

e de

taile

d ex

plor

atio

n.

capa

bilit

ies,

over and above object pennanence

per se.

It h

as b

een

prop

osed

(Bai

llarg

eon

, 199

3; B

ailla

rgeo

n, G

rabe

r, D

eVos

, & B

lack

, 199

0; D

iam

ond,

1991

; Spe

lke

et a

I., 1

992)

that

you

ng c

hild

ren

know

that

obj

ects

that

they

can

no lo

nger

see

are

stil

l pre

sent

, but

lack

the

abili

ty to

per

form

mea

ns-e

nds

anal

ysis

-to

real

ise

that

they

mus

t per

fonn

an

indi

rect

ly r

elev

ant a

ctio

n, s

uch

as r

emov

ing

the

cove

ring

, fir

st. T

hink

ing

ofth

is k

ind

is o

f co

urse

con

sist

ent w

ithna

tivis

t app

roac

hes,

like

thos

e pr

opos

ed b

y Sp

elke

et a

I., i

n w

hich

it is

sup

pose

dth

at th

e ch

ild h

as a

n in

nate

cap

acity

to r

epre

sent

and

rea

son

abou

t the

con

tinue

dex

iste

nce

of in

visi

ble

obje

cts,

but

has

faile

d to

acq

uire

som

e an

cilla

ry s

kill

need

ed f

or th

e ex

pres

sion

of

this

kno

wle

dge

in c

erta

in ta

sks.

The

fra

mew

ork

for

thin

king

abo

ut d

evel

opm

ent i

n co

nnec

tioni

st te

rms,

how

ever

, lea

ds to

a v

ery

diff

eren

t way

of

thin

king

abo

ut w

hat m

ay b

e go

ing

onin

this

cas

e. F

irst

, it s

eem

s en

tirel

y na

tura

l to

supp

ose

that

the

abili

ty to

per

fonn

in ta

sks

like

Bai

llarg

eon

s ar

ises

fro

m th

e im

plic

it pr

edic

tive

kind

of

lear

ning

that

we

have

bee

n co

nsid

erin

g th

roug

hout

this

cha

pter

. In

the

cour

se o

f ob

serv

ing

the

wor

ld a

roun

d th

em, y

oung

chi

ldre

n co

uld

deve

lop

the

abili

ty to

pre

dict

fut

ure

even

ts b

ased

on

the

curr

ent i

nput

and

on

prec

edin

g co

ntex

t. Su

ch p

redi

ctio

nsca

n co

me

to r

efle

ct a

sen

sitiv

ity to

obj

ects

that

are

not

in f

act v

isib

le f

or a

t lea

sta

shor

t per

iod

of ti

me

befo

re th

e pr

edic

ted

even

t, as

in B

ailla

rgeo

ns

task

. Ind

eed,

Mun

akat

a (p

ers.

com

m.)

has

trai

ned

a re

curr

ent n

etw

ork

(sho

wn

in F

ig. 4

.11) to

antic

ipat

e th

e fu

ture

pos

ition

s of

obj

ects

mov

ing

acro

ss a

sim

ulat

ed d

ispl

ay, a

ndto

pre

dict

the

reap

pear

ance

of

thes

e ob

ject

s w

hen

they

dis

appe

ar b

ehin

d ba

rrie

rs,

base

d on

thei

r po

sitio

n an

d di

rect

ion

of m

otio

n at

the

poin

t whe

n th

eydi

sapp

eare

d. T

he id

ea is

that

low

er le

vels

of

visu

al p

roce

ssin

g pr

ojec

t the

dis

play

,

The

Sou

rce

and

Mea

ning

of E

arly

Com

pete

nce

Let

us

now

con

side

r th

e ph

enom

enon

of

earl

y co

mpe

tenc

e-fo

r ex

ampl

e, th

e

fact

that

a y

oung

chi

ld c

an o

ften

exh

ibit

surp

rise

whe

n sh

own

an e

vent

that

vio

late

sba

sic

phys

ical

reg

ular

ities

suc

h as

obj

ect p

erm

anen

ce. W

hat i

s th

e m

eani

ng o

fth

is k

ind

of b

ehav

iour

? Sp

elke

(19

91; S

pelk

e et

aI.

, 199

2) a

nd o

ther

s (e

.g. K

eil,

1991

a, 1

991b

) ha

ve u

sed

it to

sug

gest

that

the

child

pos

sess

es c

erta

in in

nate

prin

cipl

es c

hara

cter

isin

g th

e ph

ysic

al w

orld

. But

is th

is a

nec

essa

ry in

terp

reta

tion?

Let

us

cons

ider

the

obje

ct p

enna

nenc

e ca

se a

s an

exa

mpl

e. W

e w

ill s

ee th

at th

ena

tivis

t app

roac

h an

d th

e PD

P ap

proa

ch p

rovi

de v

ery

diff

eren

t way

s to

ana

lyse

the

beha

viou

ral e

vide

nce.

Piag

et s

how

ed, i

n so

me

of h

is m

ost f

amou

s w

ork,

six

sta

ges

of d

evel

opm

ent

ofth

e ob

ject

con

cept

. His

met

hod

invo

lved

sho

win

g th

e ch

ild a

n ob

ject

, and

then

hidi

ng it

, by

putti

ng it

und

er s

omet

hing

like

a b

lank

et. Y

oung

chi

ldre

n, b

elow

the

age

of 9

mon

ths

or s

o, w

ill n

ot r

etri

eve

such

a h

idde

n ob

ject

, and

the

clas

sica

lin

terp

reta

tion

was

that

they

no

long

er r

epre

sent

its

exis

tenc

e w

hen

it is

out

of

sigh

t. A

fter

9 m

onth

s, th

e ch

ild is

cap

able

of

succ

essf

ul r

etri

eval

und

er th

ese

cond

ition

s, d

emon

stra

ting,

acc

ordi

ng to

the

Piag

etia

n vi

ew, s

ucce

ssfu

l mas

tery

of th

e fo

urth

sta

ge in

the

deve

lopm

ent o

f the

obj

ect c

once

pt. S

ince

Pia

get'

s w

ork,

how

ever

, man

y re

sear

cher

s ha

ve c

halle

nged

his

vie

w th

at c

hild

ren

lack

a c

once

ptof

obj

ect p

enna

nenc

e be

fore

9 m

onth

s. I

t has

bee

n sh

own

in v

ario

us w

ays

that

child

ren

are

sens

itive

to o

bjec

ts th

ey c

an n

o lo

nger

see

. For

exa

mpl

e, B

ailla

rgeo

n(1

987;

Bai

llarg

eon,

Spe

lke,

& W

asse

nnan

, 198

5) h

abitu

ated

chi

ldre

n to

the

site

of a

dra

wbr

idge

ope

ning

and

clo

sing

. The

chi

ldre

n th

en s

aw a

blo

ck p

lace

d in

the

path

of

the

draw

brid

ge, p

ositi

oned

in s

uch

a w

ay th

at a

s th

e dr

awbr

idge

sw

ung

the

bloc

k di

sapp

eare

d fr

om th

e in

fant

s' lo

ok lo

nger

if th

e dr

awbr

idge

sw

ung

thro

ugh

the

loca

tion

occu

pied

by

the

bloc

k, a

s th

ough

the

bloc

k ha

d ce

ased

toex

ist,

sugg

estin

g th

ey a

re s

ensi

tive

to th

e co

ntin

ued

exis

tenc

e of

the

bloc

k.O

ne p

ossi

ble

conc

lusi

on f

rom

this

res

earc

h is

that

the

obje

ct c

once

pt is

act

ually

pres

ent i

n ve

ry y

oung

chi

ldre

n, a

nd th

at P

iage

ts

task

s si

mpl

y re

quir

ed a

dditi

onal

Dis

play

FIG

. 4.11 Munakata

s re

curr

ent n

etw

ork

for

lear

ning

to a

ntic

ipat

e th

e fu

ture

pos

ition

s of

obj

ects

.

Page 13: Nurture in Development: A Parallel Distributed …stanford.edu/~jlmcc/papers/McClelland94.pdfThe Interaction of Nature and Nurture in Development: A Parallel Distributed Processing

McC

LE

LL

AN

D4. NATURE AND NURTURE

IN D

EV

EL

OPM

EN

T

in w

hich

the

obje

ct m

oves

, int

o a

perc

ept,

whi

ch is

like

ly to

be

rich

ly s

truc

ture

dve

ry e

arly

to p

rovi

de u

sefu

l rep

rese

ntat

ions

of

obje

cts

and

thei

r re

lativ

e lo

catio

nin

thre

e-di

men

sion

al s

pace

. The

net

wor

k is

trai

ned

by s

how

ing

it si

mpl

e ob

ject

sm

ovin

g ac

ross

a f

ield

of

view

, som

etim

es c

ol1i

ding

with

oth

er o

bjec

ts a

ndst

oppi

ng, a

nd s

omet

imes

mov

ing

eith

er in

fro

nt o

f or

beh

ind

thes

e ob

ject

s. I

nth

e la

tter

case

, the

mov

ing

obje

ct b

ecom

es in

visi

ble,

but

the

netw

ork

can

lear

nto

trac

k it,

usi

ng it

s pr

evio

us in

tern

al r

epre

sent

atio

n as

inpu

t to

the

form

atio

n of

subs

eque

nt in

tern

al r

epre

sent

atio

ns. T

he n

etw

ork

can

also

use

thes

e in

tern

alre

pres

enta

tions

to p

redi

ct th

e m

omen

t of

the

obje

cts

reap

pear

ance

fro

m b

ehin

dth

e ba

rrie

r. I

f th

e ob

ject

doe

s no

t rea

ppea

r, th

ere

is a

dis

crep

ancy

bet

wee

n th

epr

edic

tion

and

the

obse

rved

eve

nt. S

uch

a di

scre

panc

y m

ay b

e th

e si

gnal

that

caus

es c

hild

ren

to in

crea

se lo

okin

g w

hen

surp

risi

ng e

vent

s oc

cur,

eith

er in

rea

llif

e or

in d

evel

opm

enta

l exp

erim

ents

. In

any

case

, the

key

poi

nt is

that

the

netw

ork

com

es to

acq

uire

inte

rnal

rep

rese

ntat

ions

, not

par

t of

the

perc

ept i

tsel

f,th

at tr

ack

the

mov

emen

t of t

he o

bjec

t beh

ind

the

barr

ier,

and

to p

redi

ct it

sre

appe

aran

ce o

n th

e ot

her

side

.In

the

cont

ext o

f su

ch n

etw

orks

, the

re is

no

reas

on w

hy th

e ab

ility

to m

ake

perc

eptu

al p

redi

ctio

ns w

ould

nec

essa

rily

impl

y an

abi

lity

to r

each

for

an

obje

ctth

e ch

ild c

an n

o lo

nger

see

, Tw

o po

ssib

ilitie

s ar

e sh

own

in F

ig. 4

.12.

Acc

ordi

ng

to th

e fir

st, r

elev

ant p

arts

of t

he m

otor

sys

tem

may

rec

eive

inpu

t ini

tial1

y fr

omth

e pe

rcep

tual

rep

rese

ntat

ion

itsel

f, a

nd n

ot f

rom

the

inte

rnal

rep

rese

ntat

ion

need

ed f

or m

akin

g pe

rcep

tual

pre

dict

ions

. Whe

n th

e ob

ject

is h

idde

n, th

e in

tern

alre

pres

enta

tion

may

ref

lect

its continued existence, but the object is not

repr

esen

ted

in th

e pe

rcep

t its

elf

at th

is ti

me,

and

so

wou

ld n

ot b

e in

a p

ositi

onto

gov

ern

reac

hing

, On

this

acc

ount

, lat

er d

evel

opm

ent o

f co

nnec

tions

bet

wee

nth

e in

tern

al r

epre

sent

atio

ns a

nd th

e m

otor

sys

tem

wou

ld b

e ne

cess

ary

befo

re th

ech

ild c

ould

rea

ch f

or u

nsee

n ob

ject

s. T

he s

econ

d po

ssib

ility

is th

at r

each

ing

isba

sed

only

on

the

inte

rnal

rep

rese

ntat

ions

, but

the

inte

rnal

rep

rese

ntat

ions

for

visi

ble

obje

cts

are

sim

ply

stro

nger

than

the

repr

esen

tatio

ns o

f ob

ject

s th

at a

reno

t dir

ectly

vis

ible

. Aft

er a

l1, w

hen

the

obje

ct is

vis

ible

ther

e ar

e tw

o so

urce

s of

inpu

t, th

e pe

rcep

t and

the

prio

r in

tern

al s

tate

, whe

reas

whe

n th

e ob

ject

is in

visi

ble

there is only the prior internal state to sustain the representation. Su

ch a

repr

esen

tatio

n m

ight

eas

ily. b

e st

rong

eno

ugh

to s

usta

in p

erce

ptua

l pre

dict

ions

stro

ngly

eno

ugh

to e

licit

a m

ild s

urpr

ise

reac

tion

in ta

sks

like

Bai

llarg

eon

with

out b

eing

str

ong

enou

gh to

sup

port

str

uctu

red

reac

hing

beh

avio

ur.

Mun

akat

a (1

992)

car

ried

out

an

expe

rim

ent t

o tr

y to

dis

tingu

ish

poss

ibili

ties

such

as

thes

e fr

om th

e m

eans

~nd

s ty

pe o

f int

erpr

etat

ion,

She

firs

t atte

mpt

ed to

trai

n 7-

mon

th-o

ld c

hild

ren

to r

etrie

ve o

bjec

ts p

lace

d on

top

of a

tow

el, a

t adi

stan

ce o

f a

coup

le o

f fe

et f

rom

them

. Man

y ch

ildre

n le

arne

d to

per

form

the

task

suc

cess

ful1

y. S

he th

en g

ave

the

succ

essf

ul c

hild

ren

test

tria

ls, i

n w

hich

the

obje

ct w

as f

irst

pla

ced

on th

e to

wel

, and

then

a s

cree

n w

as p

ul1e

d in

fro

nt o

f it.

In o

ne c

ase,

the

scre

en w

as tr

ansp

aren

t. In

the

othe

r ca

se, t

he s

cree

n w

as o

paqu

e,an

d pr

even

ted

the

child

fro

m s

eein

g th

e ob

ject

. Acc

ordi

ng to

the

clai

m th

at

Inac

cess

ible

rep

rese

ntat

ions

:

Inte

rnal

Rep

rese

ntat

ion

crl

Perc

ept

Res

pons

eSy

stem

s

Dis

play

Wea

ker

repr

esen

tatio

n of

invi

sibl

e ob

ject

s:

Inte

rnal

Rep

rese

ntat

ion

Res

pons

eSy

stem

s

Perc

ept

Dis

play

FIG

.12 Two

poss

ible

net

wor

k ac

coun

ts o

f fa

ilure

to r

each

for

obj

ects

who

se r

eapp

eara

nce

from

behi

nd a

bar

rier

can

be

antic

ipat

ed.

child

ren

s fa

ilure

to r

etrie

ve h

idde

n ob

ject

s is

due

to a

failu

re o

f mea

ns~

nds

anal

ysis

, the

se c

hild

ren

, who

hav

e no

w b

een

give

n th

e re

leva

nt m

eans

, sho

uld

have

bee

n ab

le to

ret

riev

e th

e ob

ject

s ju

st a

s w

el1

whe

ther

it w

as v

isib

le o

r no

t.In

fact

, tho

ugh,

Mun

akat

a fo

und

that

chi

ldre

n of

7 m

onth

s w

ere

muc

h m

ore

likel

yto

ret

riev

e th

e ob

ject

whe

n th

e sc

reen

was

tran

spar

ent t

han

whe

n it

was

opa

que.

Whe

n th

e sc

reen

was

opa

que,

they

wer

e no

mor

e lik

ely

to p

ul1

the

tow

el w

hen

a to

y w

as p

rese

nt th

an w

hen

it w

as a

bsen

t.T

he r

esul

ts p

rese

nt a

puz

zle

from

a p

oint

of

view

in w

hich

we

trea

t the

abi

lity

to r

epre

sent

obj

ects

that

are

no

long

er v

isib

le a

s a

disc

rete

abi

lity

that

the

child

Page 14: Nurture in Development: A Parallel Distributed …stanford.edu/~jlmcc/papers/McClelland94.pdfThe Interaction of Nature and Nurture in Development: A Parallel Distributed Processing

4. NATURE AND NURTURE

IN D

EV

EL

OPM

EN

TM

cCL

EL

LA

ND

eith

er h

as o

r do

es n

ot h

ave

(Spe

lke

et a

I., 1

992)

, Chi

ldre

n de

mon

stra

te s

ensi

tivity

to o

bjec

ts th

ey c

ant s

ee in

set

tings

like

Bai

llarg

eon

s, b

ut y

et d

o no

t exh

ibit

this

sens

itivi

ty in

a ta

sk li

ke M

unak

ata

s, w

here

the

mea

ns-e

nds

inte

rpre

tatio

n ha

s be

enru

led

out.

Thu

s th

e re

sults

see

m o

n th

e on

e ha

nd to

sup

port

, and

on

the

othe

r to

cont

radi

ct, t

he c

laim

of

Spel

ke e

t aI.

(19

92, p

. 606

) th

at y

oung

chi

ldre

n "c

an

repr

esen

t sta

tes

of th

e w

orld

that

they

can

no

long

er p

erce

ive

. But

the

resu

lts c

an

be a

ccou

nted

for

in te

nns

of id

eas

that

see

m p

erfe

ctly

str

aigh

tfor

war

d w

hen

we

thin

k ab

out t

he p

robl

em f

rom

a P

DP

pers

pect

ive.

The

kno

wle

dge

that

sub

serv

es

perc

eptu

al a

ntic

ipat

ions

is s

tore

d in

the

fonn

of

conn

ectio

n w

eigh

ts, a

nd a

s su

ch

may

be

imbe

dded

in p

arts

of

the

syst

em th

at p

rodu

ce r

epre

sent

atio

ns th

at a

re n

otin

itial

ly u

sed

to g

over

n re

achi

ng. E

ven

if th

ey a

re u

sed

to g

over

n re

achi

ng, t

hest

reng

th o

f th

ese

conn

ectio

ns m

ight

not

be

stro

ng e

noug

h to

sup

port

a c

ompl

exbe

havi

our

such

as

read

ing,

eve

n if

they

can

sup

port

a m

ild s

urpr

ise

reac

tion

whe

nat

ypic

al p

erce

ptua

l seq

uela

e oc

cur.

Mun

akat

as

expe

rim

ent d

oes

not,

of c

ours

e,de

mon

stra

te d

irec

tly th

at th

e ab

ility

to p

erfo

nn p

erce

ptua

l ant

icip

atio

ns is

acq

uire

d.B

ut it

is c

onsi

sten

t with

the

idea

that

see

ms

quite

nat

ural

fro

m a

PD

P pe

rspe

ctiv

eth

at c

apab

ilitie

s ex

hibi

ted

in p

erce

ptua

l exp

erim

ents

do

not n

eces

sari

ly r

efle

ct th

efu

ll-bl

own

inna

te p

rese

nce

of c

once

pts

such

as

obje

ct p

enna

nenc

e.T

o su

mm

aris

e th

is s

ectio

n, I

hav

e tr

ied

to s

how

how

it m

ight

be

frui

tful

toth

ink

of e

arly

com

pete

nce

as r

efle

ctin

g th

e ea

rlie

st d

evel

opin

g as

pect

s of

aca

paci

ty to

pre

dict

the

futu

re b

ased

on

the

pres

ent a

nd th

e pa

st, a

nd I

hav

e tr

ied

to in

dica

te h

ow s

uch

an id

ea m

ay b

e at

leas

t as

plau

sibl

e as

an

appr

oach

bas

edon

inna

te k

now

ledg

e of

gen

eral

pri

ncip

les.

Aga

in, t

houg

h, I

wou

ld a

rgue

that

the

role

of

the

PDP

pers

pect

ive

is n

ot s

o m

uch

that

it le

ads

us to

a p

ositi

on o

nknowledge as innate

vs

lear

ned

per se.

Rat

her,

its

role

is to

pro

vide

us

with

aco

ncre

te a

nd e

xplic

it fr

amew

ork

in w

hich

to e

xplo

re th

e is

sue

thro

ugh

the

fonn

ulat

ion

of m

odel

s of

the

fact

ors

that

mig

ht b

e in

volv

ed,

Thi

s ar

gum

ent h

as to

be

righ

t up

to a

poi

nt. T

he a

bilit

y to

ext

ract

str

uctu

refr

om a

set

of

inpu

ts a

nd a

pply

it to

nov

el it

ems

clea

rly

depe

nds

on s

hapi

ng f

orce

sim

pose

d by

the

mec

hani

sm th

at e

xtra

cts

the

stru

ctur

e-ex

perie

nces

do

not t

ell

us d

irec

tly h

ow to

gen

eral

ise

in a

nd o

f th

emse

lves

. How

ever

, it i

s no

t at a

ll cl

ear

as y

et ju

st w

hat f

onn

thes

e sh

apin

g fo

rces

rea

lly h

ave

to ta

ke.

It is

cle

ar th

at a

t lea

st s

ome

aspe

cts

of la

ngua

ge s

truc

ture

can

ari

se f

rom

the

shap

ing

forc

es im

pose

d by

rat

her

sim

ple

conn

ectio

nist

net

wor

ks. F

or e

xam

ple,

sim

ple

netw

orks

invo

lvin

g an

inpu

t lay

er, a

hid

den

laye

r an

d an

out

put l

ayer

lear

nre

gula

r m

appi

ngs,

suc

h as

nat

ural

lang

uage

infl

ectio

nal s

yste

ms,

muc

h m

ore

easi

lyth

an th

ey le

arn

arbi

trar

y as

soci

atio

ns (

Plun

kett

& M

arch

man

, 199

1; R

umel

hart

& McClelland

, 198

6). I

f th

e m

achi

nery

of

cogn

ition

is m

ade

up o

f co

nnec

tioni

stne

twor

ks, t

his

alre

ady

sugg

ests

a p

ress

ure

tow

ards

the

regu

lar

map

ping

sch

arac

teri

stic

of

natu

ral l

angu

ages

. Fur

then

nore

, whe

n su

ch n

etw

orks

are

trai

ned

with

wor

d lis

ts th

at in

clud

e m

any

exam

ples

of

a re

gula

r m

appi

ng, b

ut o

nly

a fe

wexceptions, as in the English past tense, they routinely produce over-

regu

lari

satio

n er

rors

, esp

ecia

lly f

or it

ems

of r

elat

ivel

y lo

w f

requ

ency

, thu

ssh

owin

g th

at th

ey, l

ike

youn

g ch

ildre

n le

arni

ng la

ngua

ge, d

o im

pose

str

uctu

reth

emse

lves

as

they

are

lear

ning

. Ind

eed

, Har

e an

d E

lman

(19

92)

have

rec

ently

show

n th

at it

may

be

poss

ible

to a

ccou

nt fo

r th

e st

ruct

ure

of th

e E

nglis

hin

flec

tiona

l sys

tem

, and

inde

ed f

or th

e ev

olut

ion

of th

is s

truc

ture

ove

r tim

e, in

tenn

s of

the

shap

ing

forc

es im

pose

d by

suc

cess

ive

gene

ratio

ns o

f co

nnec

tioni

stne

twor

ks. T

hey

begi

n w

ith a

trai

ning

' cor

pus

refl

ectin

g th

e st

ruct

ure

of th

e sy

stem

at o

ne p

oint

in ti

me,

and

trai

ned

a fi

rst-

gene

ratio

n ne

twor

k w

ith th

is c

orpu

s. T

hey

then

trai

ned

a se

cond

- gen

~ra

tion

netw

ork

with

the

outp

ut o

f th

e fi

rst-

gene

ratio

nne

twor

k. D

ue to

the

regu

lari

satio

n pr

oces

ses

men

tione

d ab

ove,

the

outp

ut o

f th

efi

rst-

gene

ratio

n ne

twor

k w

ill te

nd to

hav

e a

sim

pler

and

less

arb

itrar

y st

ruct

ure

than

its

inpu

t. T

he s

econ

d-ge

nera

tion

netw

ork

furt

her

regu

lari

ses

the

stru

ctur

e,an

d th

en s

erve

s as

the

inpu

t to

a th

ird-

gene

ratio

n ne

twor

k. T

he p

roce

ss c

ontin

ues

over

sev

eral

gen

erat

ions

, and

the

stru

ctur

e of

the

infl

ectio

nal s

yste

m g

radu

ally

evol

ves.

Usi

ng th

is p

roce

dure

, Har

e an

d E

lman

(19

92)

have

bee

n ab

le to

cap

ture

seve

ral a

spec

ts o

f th

e ev

olut

ion

of th

e E

nglis

h pa

st- t

ense

sys

tem

. The

y ha

ve b

een

able

to s

how

that

if th

ey s

tart

the

firs

t net

wor

k of

f w

ith th

e in

flec

tiona

l sys

tem

char

acte

rist

ic o

f E

arly

Mid

dle

Eng

lish,

in w

hich

ther

e w

ere

seve

ral v

aria

nts

ofth

e re

gula

r pa

st te

nse ,

the

corp

us g

radu

ally

evo

lves

ove

r ge

nera

tions

to a

sys

tem

like

our

curr

ent s

yste

m, i

n w

hich

onl

y on

e va

riant

of t

he r

egul

ar fo

nn d

omin

ates

.T

hus

we

can

see

the

stru

ctur

e th

at e

mer

ges

as th

e re

sult

of a

pply

ing

and

reap

plyi

ng a

set

of

cons

iste

nt s

hapi

ng f

orce

s-th

ose

appl

ied

by th

e ne

twor

k-to

the

corp

us.

The

re r

emai

ns a

gre

at d

eal o

f deb

ate

abou

t whe

ther

net

wor

ks th

at le

arn

mor

phol

ogy

capt

ure

all o

f th

e de

tails

bot

h of

the

adul

t sta

te a

nd o

f th

e co

urse

of

acqu

isiti

on (

Pink

er &

Pri

nce,

198

8; P

inke

r, 1

991)

, but

ther

e is

no

disp

utin

g th

efa

ct th

at n

etw

orks

do

impo

se s

truc

ture

. The

ext

ent t

o w

hich

they

do,

inte

rest

ingl

y,is

a m

atte

r of

thei

r pa

ram

eter

s. T

hus,

for

exa

mpl

e , it

is v

ery

wid

ely

know

n in

Stru

ctur

e Im

pose

d by

the

Lea

rner

As

a fi

nal p

oint

, I w

ould

like

to a

ddre

ss o

ne m

ore

kind

of

evid

ence

that

is o

ften

take

n as

sup

port

ive

of a

nat

ivis

t pos

ition

. Thi

s is

evi

denc

e th

at th

e hu

man

min

dis

not

sim

ply

a pa

ssiv

e re

posi

tory

of

expe

rien

ce, b

ut a

ctua

lly im

pose

s st

ruct

ure

on e

xper

ienc

e be

yond

that

dic

tate

d by

the

envi

ronm

ent.

The

exi

sten

ce o

f ov

er-

regu

lari

satio

n er

rors

in c

hild

ren

s us

e of

the

past

tens

e is

one

cas

e in

poi

nt; s

uch

beha

viou

rs h

ave

been

take

n as

indi

catin

g th

at th

e co

gniti

ve s

yste

m is

impo

sing

orde

r on

its

inpu

ts, o

rgan

isin

g th

em in

to a

sys

tem

. In

fact

, the

ver

y fa

ct th

atla

ngua

ge h

as s

ynta

x ha

s of

ten

been

take

n as

evi

denc

e th

at th

ese

orga

nisa

tiona

lpr

inci

ples

are

inna

te. A

fter

all

, lan

guag

e is

a c

reat

ion

of th

e hu

man

min

d, s

o--

the

argu

men

t goe

s-th

e st

ruct

ure

of la

ngua

ge m

ust b

e a

char

acte

rist

ic o

f th

est

ruct

ure

of th

e m

inds

that

use

it. I

f al

l lan

guag

es e

xhib

it ce

rtai

n st

ruct

ural

char

acte

rist

ics,

it w

ould

app

ear

to f

ollo

w th

at th

e sp

ecif

icat

ions

of

wha

t the

sest

ruct

ural

cha

ract

eris

tics

shou

ld b

e m

ust b

e in

nate

as

wel

l.

Page 15: Nurture in Development: A Parallel Distributed …stanford.edu/~jlmcc/papers/McClelland94.pdfThe Interaction of Nature and Nurture in Development: A Parallel Distributed Processing

McC

LE

LL

AN

D4. NATURE AND NURTURE

IN D

EV

EL

OPM

EN

T

the

conn

ectio

nist

res

earc

h co

mm

unity

that

if o

ne w

ants

to p

rodu

ce a

mec

hani

smth

at e

xtra

cts

the

gene

ral s

truc

ture

com

mon

to m

ost o

f th

e ite

ms

in a

trai

ning

set

,ig

nori

ng id

iosy

ncra

tic c

hara

cter

istic

s of

indi

vidu

al e

xem

plar

s, th

en o

ne s

houl

dco

nstr

ain

the

num

ber

of c

onne

ctio

n w

eigh

ts. I

n re

spon

se to

this

, sev

eral

rese

arch

ers

have

use

d th

e m

etho

d of

Wei

gand

, Rum

elha

rt a

nd H

uben

nan

(199

1),

intr

oduc

ing

an e

xtra

tenn

in th

e fu

nctio

n th

at th

e ne

twor

k m

inim

ises

, ove

r an

dab

ove

the

disc

repa

ncy

betw

een

its im

plic

it pr

edic

tions

and

obs

erve

d in

puts

,re

flec

ting

the

num

ber

of n

on-z

ero

conn

ectio

n w

eigh

ts. W

hat h

appe

ns in

lear

ning

depe

nds

on th

e w

eigh

t giv

en to

this

ext

ra te

nn. W

hen

this

cos

t ten

n is

wei

ghte

dhe

avily

, net

wor

ks w

ill b

ecom

e ve

ry le

an, imposing a severe tendency to

regu

lari

se th

eir

inpu

ts; o

n th

e ot

her

hand

, if

the

pres

sure

is m

ild, n

etw

orks

can

be fr

ee to

mem

oris

e th

e de

tails

of t

he tr

aini

ng c

orpu

s, g

reat

ly r

educ

ing

the

syst

emat

icity

of

thei

r ha

ndlin

g of

nov

el f

onns

. The

wei

ght a

ssig

ned

to th

e co

stte

nn is

an

exam

ple

of a

net

wor

k pa

ram

eter

that

mig

ht v

ary

acro

ss d

omai

ns,

perh

aps

set d

iffe

rent

ly in

dif

fere

nt p

arts

of

the

syst

em b

y ge

netic

reg

ulat

ion.

Of

cour

se, i

t will

be

clea

r th

at th

e w

ork

desc

ribe

d to

dat

e fa

lls f

ar s

hort

of

dem

onst

ratin

g th

at th

e fu

ll st

ruct

ure

of n

atur

al la

ngua

ge c

an b

e un

ders

tood

intenns of the in

tera

ctio

ns o

f ne

twor

k le

arni

ng p

roce

sses

with

exp

erie

nce.

Nev

erth

eles

s, it

doe

s se

em c

lear

that

the

POP

appr

oach

off

ers

a ne

w w

ay to

thin

kab

out j

ust w

hat k

ind

of b

uilt-

in c

onst

rain

ts m

ight

be

invo

lved

in m

aint

aini

ng a

ndin

deed

impo

sing

str

uctu

re in

a d

omai

n. R

athe

r th

an th

ink

abou

t dom

ain-

spec

ific

prin

cipl

es, s

peci

fied

in te

nns

of th

e ac

tual

con

tent

of

the

dom

ain

, we

are

enco

urag

ed to

thin

k in

tenn

s of

the

setti

ngs

of th

e pa

ram

eter

s of

ver

y ge

nera

lm

echa

nism

s. O

f co

urse

, nat

ure

does

set

thes

e pa

ram

eter

s, a

nd p

resu

mab

ly s

hedo

es s

o fo

r a

purp

ose.

Ind

eed,

we

can

imag

ine

that

it is

one

of

the

task

s of

nat

ural

sele

ctio

n to

spe

cify

just

suc

h pa

ram

eter

s, p

erha

ps s

epar

atel

y fo

r di

ffer

ent p

arts

of th

e co

gniti

ve s

yste

m, s

o th

at s

ome

part

s w

ill b

e in

clin

ed to

str

uctu

re a

nd e

ven

rest

ruct

ure

thei

r in

puts

and

oth

ers

to m

emor

ise

indi

vidu

al c

ases

.

abou

t lea

rnab

ility

bas

ed o

n th

e dr

awba

cks

of c

lass

ical

ass

ocia

tioni

sm n

eed

to b

ere

-eva

luat

ed. I

hav

e ar

gued

that

the

char

acte

rist

ics

of th

ese

mec

hani

sms

offe

rus

way

s to

und

erst

and

the

fitf

ul c

ours

e of

dev

elop

men

t; th

at th

ey a

llow

us

tofr

ame

expl

icit

alte

rnat

ives

to n

ativ

ist a

ccou

nts

of e

arly

com

pete

nce;

and

that

they

impo

se s

truc

ture

in w

ays

that

cor

resp

ond

to th

e st

ruct

ure

that

lear

ners

impo

se.

The

se d

emon

stra

tions

sug

gest

that

the

prin

cipl

es th

at g

over

n th

e tim

e-co

urse

and

the

outc

ome

of d

evel

opm

ent m

ay b

e bo

th s

omew

hat b

road

er a

nd a

t the

sam

etim

e so

mew

hat m

ore

spec

ific

than

they

ofte

n ap

pear

to b

e in

nat

ivis

t app

roac

hes.

On

the

one

hand

, if

inde

ed g

ener

al p

rinc

iple

s of

pro

cess

ing

and

lear

ning

app

lyac

ross

dom

ains

, the

n th

e ul

timat

e ex

plan

atio

n of

cog

nitiv

e st

ruct

ure

and

itsem

erge

nce

will

be

base

d la

rgel

y on

thes

e ge

nera

l pri

ncip

les.

On

the

othe

r ha

nd,

the

way

s in

whi

ch th

ese

gene

ral p

rinc

iple

s pl

ay th

emse

lves

out

in d

etai

l in

spec

ific

dom

ains

are

obv

ious

ly h

ighl

y va

ried

, and

an

effo

rt to

und

erst

and

wha

tgi

ves

rise

to th

e st

ruct

ure

that

em

erge

s in

any

giv

en d

omai

n w

ill f

orce

a f

urth

ercl

arifi

catio

n of

the

natu

re o

f the

dom

ain

itsel

f, an

d of

the

way

s in

whi

ch it

sst

ruct

ure

inte

ract

s w

ith th

e ge

nera

l pri

ncip

les

to g

ive

rise

to th

e ph

enom

ena

that

we

actu

ally

obs

erve

in b

ehav

iour

. In

this

sen

se, w

e m

ay o

ne d

ay a

chie

ve a

dua

lun

ders

tand

ing,

at o

nce

of th

e ge

nera

l and

of

the

spec

ific

, and

of

the

way

in w

hich

the

spec

ific

ref

lect

s an

d m

odul

ates

the

char

acte

rist

ics

of th

e ge

nera

l pri

ncip

les.

The

ulti

mat

e ev

alua

tion

of th

e ut

ility

of

the

conn

ectio

nist

app

roac

h w

ill, o

fco

urse

, dep

end

on h

ow m

uch

insi

ght t

hey

ultim

atel

y yi

eld

into

dev

elop

men

t. A

tpr

esen

t it i

s no

t at a

ll cl

ear

how

far

con

nect

ioni

st id

eas

abou

t lea

rnin

g an

dre

pres

enta

tion

will

take

us.

For

now

, it s

eem

s cl

ear

that

thes

e id

eas

are

poin

ting

us in

inte

rest

ing

and

usef

ul d

irec

tions

. Whe

ther

, as

rese

arch

con

tinue

s, to

day

conn

ectio

nist

idea

s th

emse

lves

sur

vive

, or

whe

ther

they

lead

inst

ead

to e

ven

mor

epo

wer

ful i

deas

, is

a qu

estio

n w

e m

ust l

eave

to th

e fu

ture

.

The

re is

a g

reat

dea

l mor

e to

s~y

abo

ut th

e po

ssib

le im

plic

atio

ns o

f con

nect

ioni

st

mod

els

for

the

stud

y of

cog

nitiv

e de

velo

pmen

t, an

d th

e w

ork

is ju

st b

egin

ning

.I

hope

that

this

cha

pter

beg

ins

to il

lust

rate

the

reas

ons

why

I e

xpec

t the

app

roac

hto

bec

ome

a m

ajor

foc

us o

f re

sear

ch. T

he m

ain

poin

t of

this

cha

pter

has

sim

ply

been

to s

how

that

con

nect

ioni

st m

odel

s al

low

us

to e

xplo

re in

exp

licit

com

puta

tiona

l mod

els

the

vari

ous

fact

ors

and

forc

es th

at m

ight

be

thou

ght t

osh

ape

deve

lopm

ent,

The

se in

clud

e th

e en

viro

nmen

t, th

e ba

sic

prin

cipl

es o

fpr

oces

sing

and

lear

ning

, and

dom

ain-

spec

ific

arch

itect

ural

and

par

amet

ricva

riat

ions

that

may

sha

pe th

e tim

e-co

urse

and

the

outc

ome

of d

evel

opm

ent.

Ase

cond

them

e ha

s be

en to

arg

ue th

at m

any

of th

e ar

gum

ents

that

hav

e le

d m

any

prom

inen

t the

oris

ts to

fav

our

nativ

ist a

rgum

ents

des

erve

to b

e re

-eva

luat

ed. T

ore

cap

brie

fly,

I h

ave

argu

ed th

at c

onne

ctio

nist

mod

els

of le

arni

ng o

ffer

mor

epo

wer

ful t

ools

than

the

clas

sica

l ass

ocia

tioni

st a

ppro

ach,

so

that

arg

umen

ts

RE

FER

EN

CE

S

AK

NO

WL

ED

GE

ME

NT

S

SUM

MA

RY

AN

D C

ON

CL

USI

ON

ST

he a

rgum

ent p

rese

nted

in th

is c

hapt

er h

as b

een

deve

lope

d ov

er s

ever

al y

ears

, and

sev

eral

colle

ag~e

s an

~ co

llab~

rato

rs h

ave

cont

ribu

ted

in s

ubst

antia

l way

s to

its

deve

lopm

ent.

Iw

ould

m p

artic

ular

hke

to a

ckno

wle

dge

cont

ribut

ions

by

Jeff

Elm

an, G

eoff

Hin

ton,

Jam

es H

oeffn

er, A

nnet

te K

arm

iloff-

Smith

, Yuk

o M

unak

ata,

Ran

dy O

' Rei

lly, K

imPl

unke

tt, D

ave

Rum

elha

n, M

ark

Seid

enbe

rg a

nd R

oben

Sie

gler

. The

mem

bers

of

the

PDP

Gro

up a

t CM

U h

ave

prov

ided

use

ful c

omm

ents

and

fee

dbac

k. T

he w

ork

has

been

supp

oned

by

a C

aree

r D

evel

opm

ent A

war

d (M

H- O

O38

5) a

nd a

Pro

gram

Pro

ject

Gra

nt(M

H-4

7566

) fr

om N

IMH

,

Bai

llarg

eon,

R. (

1987

). O

bjec

i per

man

ence

in 3

. 5-

and

4.5-month-old infants,

Del

'ldop

men

tal

Psyc

holo

gy,

23, 6

55--

664.

Bai

llarg

eon

, R. (

1993

). T

he o

bjec

t con

cept

revi

site

d: N

ew d

irect

ions

in th

e in

vest

igat

ion

of in

fant

sphysical knowledge. In C. Granrud (Ed),

Vis

ual p

erce

ptio

n an

d ('O

~nition in infancy, C

arne

gie

Mel

lon

Sym

posi

a on

Cog

nitio

n. H

illsd

ale,

NJ:

Law

renc

e E

rlba

um A

ssoc

iate

s In

c.

Page 16: Nurture in Development: A Parallel Distributed …stanford.edu/~jlmcc/papers/McClelland94.pdfThe Interaction of Nature and Nurture in Development: A Parallel Distributed Processing

McC

LE

LL

AN

DN

AT

UR

E A

ND

NU

RT

UR

E I

N D

EV

EL

OPM

EN

T

Bai

llarg

eon,

R.,

Gra

ber,

M.,

DeV

os, J

" &

Bla

ck, J

. (19

90).

Why

do

youn

g in

fant

s fa

il to

sea

rch

for hidden objects?

Cog

nitio

n,

36, 2

55-2

84.

Baillargeon, R.

, Spe

lke,

E,S

., &

Was

serm

an, S

. (19

85).

Obj

ect p

erm

anen

ce in

five

-mon

th-o

ldin

fant

s.

Cognition, 20,

191-

208.

Bat

es, E

., &

Elm

an,

L. (1993). Connectionism and the st

udy

of c

hang

e, In

M.J

ohns

on

(Ed.

Bra

in

development and cognition:

read

erpp

. 623

--64

2. O

xfor

d: B

lack

wel

l.C

ase,

R. (

1978

), In

telle

ctua

l dev

elop

men

t fro

m b

irth

to a

dulth

ood:

A n

eo-P

iage

tian

inte

rpre

tatio

n.

In R. Siegler (Ed.

),

Chi

ldre

ns thinking: What develops?,

pp. 3

7-71

. Hill

sdal

e, N

J: L

awre

nce

Erlb

aum

Ass

ocia

tes

Inc.

Dia

mon

d, A

. (19

91).

Neu

rops

ycho

logi

cal i

nsig

hts

into

the

mea

ning

of o

bjec

t con

cept

dev

elop

men

t.

In S, Carey & R. Gelman (Eds),

The epigenesis of mind: Essays on biology and cognition.

pp.

67-1

10. H

illsd

ale,

NJ:

Law

renc

e E

rlba

um A

ssoc

iate

s In

c.E

lman

, J,L. (1990). Finding structure in time, Cognitive Science, i4,

179-

211.

Flav

ell,

J.H, (1963),

The developmental psychology of Jean Piaget.

New

Yor

k: V

an N

ostr

and,

Har

e, M

., &

Elm

an, J

.L. (1992). A connectionist account of English in

flec

tiona

l mor

phol

ogy:

Evidence from language change, In

Pro

ceed

ings

of t

he i4

th A

nnua

l Con

fere

nce

of th

e C

ogni

tive

Science Society,

pp, 2

65-2

70. H

illsd

ale,

NJ:

Law

renc

e E

rlba

um A

ssoc

iate

s In

c.H

into

n, G

.E, (

1986

). L

earn

ing

dist

ribut

ed representations of concepts, In P

roce

edin

gs o

f the

Eig

hth

Annual Conference of the Cognitive Science Society,

pp. 1

-12.

Hill

sdal

e, N

J: L

awre

nce

Erl

baum

Ass

ocia

tes

Inc.

Inhelder, B., & Piaget, J. (1958).

The

gro

wth

of

logi

cal t

hink

ing

from

chi

ldho

od to

ado

lesc

ence

.N

ew Y

ork:

Bas

ic B

ooks

,K

anni

loff

-Smith, A, (1986), From meta-

proc

esse

s to

con

scio

us a

cces

s: E

vide

nce

from

chi

ldre

nmetalinguistic and repair data.

Cog

nitio

n.

23, 9

5-14

7.

Kan

nilo

ff-S

mith

, A. (

1991

). B

eyon

d m

odul

arity

: Inn

ate

cons

trai

nts

and

deve

lopm

enta

l cha

nge.

In S. Carey & R. Gelman (Eds),

The

epi

gene

sis

of m

ind:

Ess

ays

on b

iolo

gy a

nd c

ogni

tion,

pp, 1

71-1

97. H

illsd

ale,

NJ:

Law

renc

e E

rlba

um A

ssoc

iate

s In

c,K

arm

iloff

-Smith, A. (I 992a).

Beyond modularity:

deve

lopm

enta

l per

spec

tive

on c

ogni

tive

scie

nce.

C

ambr

idge

, MA

: MIT

Pre

ss.

Kan

nilo

ff-S

mith

, A. (

I 99

2b).

Nat

ure,

nur

ture

, and

PD

P: P

repo

ster

ous

deve

lopm

enta

l pos

tula

tes?

Connection Science.

4, 2

53-2

69.

Kei

l, F,

C. (1981). Constraints on knowledge and cognitive development.

Psychological Review,

88, 1

97-2

27.

Kei

l, F,C. (1987), Conceptual development and category st

ruct

ure,

In

U, N

eiss

er (

Ed.

),

Con

cept

s

and conceptual development: Ecological and intellectualfactors in categorization,

pp. 1

75-2

00.

Cam

brid

ge: C

ambr

idge

Uni

vers

ity P

ress

,K

eil,

F,C. (l99Ia). Invited address to the 1

3th

Ann

ual M

eetin

g of

the

Cog

nitiv

e Sc

ienc

e So

ciet

y,C

hica

go, I

L, A

ugus

t.K

eil,

F.C

, (19

91 b

). T

he e

mer

genc

e of

theo

retic

al b

elie

fs a

s co

nstr

aint

s on

con

cept

s, I

n S,

Car

ey &

R. Gelman (Eds),

The epigenesis of mind: Essays on biology and cognition,

pp, 2

37-2

56, H

illsd

ale,

NJ:

Law

renc

e E

rlbau

m A

ssoc

iate

s In

c,K

now

lton,

B.1

., R

amus

, S.1

., &

Squ

ire,

L.R

. (19

92).

Inta

ct a

rtifi

cial

gra

mm

ar le

arni

ng in

am

nesi

a:

Dissociation of classification learning and explicit memory for specific instances.

Psyc

holo

gica

l

Scie

nce.

3.

172

-179

.

Kut

as. M

., &

Hill

yard

S.A

. (19

80).

Rea

ding

sen

sele

ss s

ente

nces

: Bra

in p

oten

tials

ref

lect

sem

antic

inco

ngru

ity,

Science, 207,

203-

205.

McC

lella

nd, J

.L. (

1989

). P

aral

lel d

istr

ibut

ed p

roce

ssin

g: I

mpl

icat

ions

for

cog

nitio

n an

dde

velo

pmen

t. In

R,

M. Morris (Ed),

Para

llel d

istr

ibut

ed p

roce

ssin

g: im

plic

atio

nsfo

r ps

ycho

logy

and neurobiology,

pp. 9

-45.

New

Yor

k: O

xfor

d U

nive

rsity

Pre

ss.

McC

lella

nd, J

.L. (

in p

ress

). A

con

nect

ioni

st p

ersp

ectiv

e on

the

proc

ess

of d

evel

opm

enta

l cha

nge,

In T

. Sim

on &

G.S. Halford (Eds), D

evel

opin

g co

gniti

,'e c

ompe

tenc

e: N

ew a

ppro

ache

s to

pro

cess

mod

elin

g.

Hill

sdal

e, N

J: L

awre

nce

Erl

baum

Ass

ocia

tes

Inc.

McC

lella

nd, J

.L., & Jenkins. E, (1991). Nature, nurture. a

nd c

onne

ctio

ns: I

mpl

icat

ions

of

conn

ectio

nist

mod

els

for

cogn

itive

dev

elop

men

t. In

K. V

an L

ehn

(Ed.

),

Architectures for

inte

llige

nce,

pp

. 41-

73. H

ilisd

ale,

NJ:

Law

renc

e E

rlba

um A

ssoc

iate

s In

c,M

cCle

lland

, J.L

" St

Joh

n. M

" &

Tar

aban

, R. (

1989

), S

ente

nce

com

preh

ensi

on: A

par

alle

ldistributed processing approach.

Language and Cognitive Processes,

, SI 2

87-3

35.

Mun

akat

a, Y

. (19

92).

Litt

le b

y lil

lle: T

owar

d a

grad

ualis

tic a

ccou

nt o

f ob

ject

per

man

ence

. Pos

ter

presentation at the i992 McDonnell Summer institute of Cognitive Neuroscience,

Dar

tmou

th, M

A.

Rei

lly, R

,C, (1992), The self-organization of translation invariant representations, T

ech.

Rep

ort

PDP.

CN

S.92

.5,

Dep

artm

ent o

f Ps

ycho

logy

, Car

negi

e M

ello

n U

nive

rsity

, Pitt

sbur

gh, P

A.

Pasc

ual-Leone, J, (1987), Organismic processes for neo-Pi

aget

ian

theo

ries

: A d

iale

ctic

al c

ausa

laccount of cognitive development.

international Journal of Psychology,

, 531

-570

.Pinker, S. (1991). Rules of language.

Scie

nce,

25

3, 5

30-5

35,

Pink

er, S

., &

Pri

nce,

A, (

1988

). O

n la

ngua

ge a

nd c

onne

ctio

nism

: Ana

lysi

s of

a p

aral

lel d

istr

ibut

edprocessing model of language acquisition.

Cog

nitio

n,

28, 7

3-93

.Pl

unke

tt, K

., &

Mar

chm

an, V

. (19

91).

Ush

aped

lear

ning

and

fre

quen

cy e

ffec

ts in

a m

ultil

ayer

ed

perceptron: Implications for child language acquisition.

Cog

nitio

n,

, 43-

102,

Plunkett, K" & Sinha, C. (1991). Connectionism and developmental theory.

Psy

kolo

gisk

Skr

iftse

reAarhus. i6,

34.

Reb

er, A

, S. (1976). Implicit learning of synthetic languages: T

he r

ole

of th

e in

stru

ctio

nal s

et.

Journal of Experimental Psychology: Human Learning and Memory.

2, 8

8-94

.R

eber

, A.S. (1989). Implicit learning and tacit knowledge.

Jour

nal o

f Exp

erim

enta

l Psy

chol

ogy:

Gen

eral

, //8

, 219

-235

.Rumelhart, D.

E.,

& M

cCle

lland

, J.L

. (19

86).

On

lear

ning

the

past

tens

e of

Eng

lish

verb

s. I

n J.

McClelland, D.E. Rumelhart, & the PDP Research Group (Eds),

Para

llel d

istr

ibut

ed p

roce

ssin

g:Explorations in the microstructure of cognition,

Vol

. II,

pp,

21

~27

1. C

ambr

idge

, MA

: MIT

Pre

ss.

Rumelhart, D.

E.,

Hin

ton,

G.E

., &

McC

lella

nd, J

.L. (

I 98

6a).

A g

ener

al f

ram

ewor

k fo

r pa

ralle

ldi

stri

bute

d pr

oces

sing

. In

D,E

. Rum

elha

rt, J

. L. M

cCle

lland

, & th

e PD

P R

esea

rch

Gro

up (

Eds

),Parallel distributed processing: Explorations in the microstructure of cognition,

Vol

. I, p

p.45

-76,

Cam

brid

ge, M

A: M

IT P

ress

.Rumelhart, D.

, Hin

ton,

G,

. & W

illia

ms,

R.I

. (19

86b)

. Lea

rnin

g in

tern

al r

epre

sent

atio

ns b

y er

ror

propagation. In D.

E, R

umel

hart

, J.L. McClelland, & the PDP Research Group (Eds),

Para

llel

dist

ribut

ed p

roce

ssin

g: E

xplo

ratio

ns in

the

mic

rost

ruct

ure

of c

ogni

tion

Vol

. I, p

p, 3

18-3

64.

Cam

brid

ge, M

A: M

IT P

ress

.Rumelhart, D,

. McC

lella

nd, J

.L.,

and

the

PDP

Res

earc

h G

roup

(E

ds)

(198

&),

P

aral

lel d

istr

ibut

edprocessing: Explorations in the microstructure of cognition,

Vol

.!. C

ambr

idge

, MA

: MIT

Pre

ss,

St. J

ohn,

M,F

., &

McC

lella

nd. J

.L. (1990). Learning and applying contextual co

nstr

aint

s in

sen

tenc

eco

mpr

ehen

sion

. Artificial Intelligence,

, 217

-257

.S

chac

ter,

D. L

. (19

87).

Im

plic

it memory: History and current status.

Journal of Experimental

Psychology: Learning, Memory and Cognition. i3,

501-

518.

Sch

mid

t. W

. C.,

& S

chul

tz, T

.R. (

1992

). A

n in

vest

igat

ion

of balance scale success. In Pr

ocee

ding

sof

The

i4th

Ann

ual C

onference of The Cognitive Science Society,

pp. 7

2-77

. Hili

sdal

e, N

J: L

awre

nce

Erlb

aum

Ass

ocia

tes

Inc,

Sie

gler

, R. S

. (19

76).

Thr

ee a

spec

ts o

f cognitive development.

Cogniti,'e Psychology,

, 481

- 520

,O

rlan

do, F

L: A

cade

mic

Pre

ss.

Sie

gler

, R, S. (1981). Developmental sequences between and within

conc

epts

. M

onog

raph

s of

the

SocieTy for Research in Child Development,

46 (

Who

le N

o. 1

89).

Spelke, E.S. (1991). Physical kn

owle

dge

in in

fanc

y: R

efle

ctio

ns o

n Pi

aget

's th

eory

. In

S. C

arey

&R. Gelman (Eds), The epigenesis of mind: STudies on biology and cognition,

pp. 1

33-1

70. H

ilisd

ale,

NJ:

Law

renc

e E

rlbau

m A

ssoc

iate

s In

c.Spelke, E.S.

, Bre

inlin

ger,

K.,

Mac

ombe

r, J

., &

Jac

obso

n, K

. (19

92).

Ori

gins

of

know

ledg

e.

Page 17: Nurture in Development: A Parallel Distributed …stanford.edu/~jlmcc/papers/McClelland94.pdfThe Interaction of Nature and Nurture in Development: A Parallel Distributed Processing

McC

LE

LL

AN

D

Psy(

'hol

nRit'

al

Rr,

'irM

', 4

, 605

~32,

Weigand, A,

S.,

Rum

elha

n. D

,E.,

&: H

uber

man

, B.A. (1991). Generalization by weight-elimination

with

app

licat

ion

to f

orec

astin

g. I

n R

,P. L

ippm

an, J

.E, M

oody

, '"

D,S

. Touretzky (Eds),

Ad,

'ana

sin nrural in/ormation procrssinR systrms.

Vol

. 3. p

p. 8

75-8

82. S

an M

ateo

. CA

: Mor

gan

Kau

fman

n.W

ilken

ing.

F.,

'" A

nder

son.

N,H

. (1991), Representation and diagnosis of kn

owle

dge

stru

clU

res

in d

evel

opm

enta

l psy

chol

ogy.

In N

.H. A

nder

son

(Ed.

),

Con

trib

utio

ns to

in!n

rmal

ion

intr

Rra

tion

throry. Vol. III: Drllr/opmrnlal, pp

. 45-

80. H

illsd

ale,

NJ:

Law

renc

e E

rlba

um A

ssoc

iate

s In

c,