. Human Arain MappirS 2t234411994) .

Neural Network Interactions Related to Auditory

L.earning Analyzed With Structural Equation Modeling

F. Gonzalez-Lima and A.R. Mclntosh

ltistitute Jor Neurascience and Depart'nent oJ Psycholog!, Ulritercity ol T.Jtas, Austin, Texas (F.C.-L.),
a d Labantory ol Newosciellces, Natio al Instit te an Aging, Natianal lnslitules oJ Health,
Belhesda, Matltland (A.R McL)

it it

Abstra.r lf learning is an emergent property of inteEcting brain regions, unde6tanding it requires a

networkanalysis ofthe patterns of interaction between brain regions. Brain mapping techniques have the
potential of providing information about tunctional interactions within entire neural systems, but
compulational methods aie needed to make sense oI the complex interactions thattake place in the brain-
This leview paper illustraies, with exanpies flom auditory learning studies, the apptcation of a
computatjonal m€thod that quantifies the interactions anong neuraL regions in terms of netwo.ks, and
whatis gained lrom this approach. The method ls Shuctural Equation ModelinS, and it computes network
interactions by combining anatomical circuilry wiih the covariation in ihe activity between auditory
structures. Activiiy was assessed by fluorodeoxyglucose upiake, and the fmctional strengths ofauditory
pathways we.e quaniified by computjng paih coefficients representinS the strenSth of the functional
influence through each anatomical path. Changes in ihese values were used as indices of how leamed
infolmation was proce$ed and modified within the audiiory sysiem. Structunl hodels of the audito.y
system .evealed the patterns of network inieractions related io habituation to a sound, as well as the
patterns.elated to the opposite learned associative properties ol the same sound. Theyillustlated how the
functional interactions among parallel auditory neural paihways were nodified as a .esuli of audiiory
learning. This analysis supports the hypothesis ihai audito.y learning is an emergent network property,
dishibutedamongconneciedbrainregionsshowingspecificpatteftsofinteraction..reeawikr-Lj"",h".

Keywords: auditory learning, sbuctural equaiion modeling, path analysis, Pavlovian conditioning,
habituation, 2 deoxygldcose, fluorodeoxyglucose, neuroimaging, network interachons, covan-
ance analysis, neural pa thway, .a t

INTRODUCTION tions in which neural activiiychanges are subtle. With

the introdu.rion of neb funLhonal neuroimaSinS
Advances in analytic techniques are especially im- techniques for data acquisihon, new data analysis
portant for neuroimatinS studies of learning func lechniques h.ve .lco been developFd Ifor re! iew., -ee
Horwitz et al., 1992, Mclntosh and Gonzalez-Lima,
1992b1. For €xample, averat€ image subtraction [Fox et
Received for plblication May 2Z 1994; r€vision accepted September
al., 19881. statisiical parametric maps lFdston et al.,
26, t994.
19911, cluster analysis lPoline and Mazoyer, 1993, this
Addr€ss reprnrt requests to Dr. F. Conzalez Lima, Unnersity of
TexasatAustin, Mezes 330, Austin, TX 78712. issue; Roland et a1., 19931, and landscape param€tdc
A.R. Mclntosh h now at the Rotnan Research lnstitut€, Baycrest
profiles [Barreto and Gonzalez-Lima, 1994] have been
C€nrre lor Geriahic Care, 3s60 Bathust Sheet, Toronlo, Ontano, developed to reveal changes in r€donal neural activ-
ity. With modern metabolic mapping techniques it is
o 1994 Wiley-Lissj ltrc.
 . Gonzalez-Lima and Mclnlosh
possible to obtain functional data from mostregions in
a sinde brain lconzal€z-Lima et al., 1992J. This is a
Ereai advance over tradiiional neuroscience meihods
of studying learninS because the op€rations of many
neural structures can be visualized simulianeously in
d \inBIe brdin T)picdll) mort dndlyses of i,,.rpprng
data are restricted to the comparison of average
regional activity between experimental groups or
conditions. This is like treating each brain re$on as if
ii were separate from the system, but the neural
functions that mediate associative learning are prob-
ably dependent on the inieraciions that take place
within and between these systems. By restricting the
analysis to one iegion at a iime, brain mapping studies
of learninS may iSnore important sources of informa-
iion about the interaction beiween reFons. Therefore,
in order to make better use of brain imaging data io
study leaming functions, computational iechniques
need to account for the in teractions between elements
in a neural network.
The use of computational approaches to measure
network inieractions in brain siudies may be iraced
back to the Cerstein ei al. [1978] analysis of nultiunit
electrical recordings and ihe Cevins et a]. 119851
analysis of electroencephalographic waves. The first
applications of covariance analyses to brain metabolic
data were reported in 198,1by Clark et al., Horwitz et
a1., and Metter et al. They demonstrated the impor-
tance of functional interactions in neural systems by
the use of interredonal Pearson produclmoment
correldlion L oeth, renis to quanlrf) meraboli. n'aPPin8
data. This method ofcovariance analysis has provided
a major advance in the effort to determine which brain
regions are functionally associated with one another
during a particular experimental condiiion. However,
when one wants io examine interactions in a multi-
siruciure system, inierpretations based on Pairwise
correlations can become complicated. Other ap-
proaches that can provide information about sysiem
interactions are the vafous multivariate techniques
such as discriminant analysis, pincipal components
analysis. facior analysis, and canonical correlation.
Some of these techniques have been applied success-
tully to positron emission tomoSraphy (PET) meia-
bolic data from humans lclark et al., 1991; Frjston et
al., 1993, Frision, this issuel and optical recordings
from cortex lshoham et al., 19911. One commonality
between these techniques is the assumption thatbrain
regions that function together have corelated aciivi-
ties lHorwitz et a].,19921. These iechniques determine
patterns of regional interrelations independent of
infl uences mediated by specifi c neuroanaiomical path-
wcy-. However, brd'n fun!lron- relalFd lo ledrning rre
.
Likely to be highly dependent on the specific connectiv-
itybetweenbrain regions lconzalez-Lima, 1989, 1992;
Mclniosh and Gonzalez-Lima, 19941.
The reasoning that brain connectivity influences
brain funciioning gave the impetus for the develop-
meri of our siructural modeling approach io quantify
functional interaciions mediated by specific neural
connectiorls IMclntosh and Conzalez-Lima, 1991,
Ioo2db louj, looa. ll-ir ir-uel ln de\elopinB our
approach/ we have kept in mirld a basic principle thai
we should like to state explicitly- This is the principle
or neural rnter.r, hon Tl \ldle. lhd I il npurdl retion5 drF
synaPtically connected. the disturbance in the postsyn-
aptic action potentials of a region is passed on to
anoiher. In other words, brain regions do not merely
aci locally, ihey interact with one another in complex
neural networks. Therefore, brain activiiy is interde-
pendent on the aciions and reactions of ihe compo-
nents that form the neural neiworks. We can now
monitor ihe operations of many ne ral regions simul-
taneously using brain mapping meihods. The model-
ing technique presented here and similar analyses of
network interactions may help us to understand the
combined actions of these neuralregons and how this
activity is related to behavior.
Mclniosh and Conzalez-Lima 11991, 1992ah, 1993,
19941 demonsirated how €ovariance structural equa-
tion modeling [Bentler, 1985; Berry, 1984, Bollen,1989,
Davis, 1985, Hayduk, 1987; ldreskog and Scirbom,
1979, 1984, 1989; Loehlin, 1987, Long, 19831 (also
known as path analysis) could be used to quantify
simultaneously the interactions between iniercon
nected brain regions. When applied io neural systems,
sfructural equaiion modeling uses information about
the anatomical pathways and ihe corelation coeffi-
cienis beiween brain regions io determine the func-
tional paihways in a given experiment. Struciural
modeling uses algorithms thai attempt to account tor
an observed pattern of correlations based on the
causal structure of a system defirled by ihe neuroana-
tomical conneciions. The models use inierregional
correlaiions ofactivity to quaniify the pairwise pattern
of interaciions ihai take place between brain regiorls.
The application of struciural equaiion modelinS io
neural data uses the basic assumption ihai ihe ob-
servedpatternof corlelationsbetweenbrain regionsis
due, at least in part, io either common influences to
both regjons or direct anatomical connechons be-
tween them. By combining ihe correlations between
brain regions and the anatomical connections (causal
order), parameters are estimated for all inlluences in
the system. This structural equation modeiint apptica-
iion is a logical extension of ihe ini€rregional correla-
 . P.rh Analsis ofAudilory LearninC.
tional approach since the solutions for the models are
derived from correlaiion coefticienis. li adds to the
analysis of mapping data in that instead of looking at
reSions that are functionally associated h a pairwise
manner/ the models can demonstrate how specific
pathways in entire sysiemsfitcract in different experi,
mental paradigms. Our structural modeling approach
has also been recently applied to the study of human
brain function [McIniosh et al., 1994] and dysflulciiorl
lcrafton ei a1., this issuel.
In the companion paper lMclniosh and Gonzalez-
Lina, ihis issuel we explain the general application of
structural equation modeling io brain inaging,provid-
ing the methodological background nceded for a
better understanding of the present paper. In this
paper we review specific applications of structural
equaiion modeling to funciional brain .iata lrom
mapping expernnents of audiiory learninS lunctions.
The paper is both a review of our previously pub-
lished results as well as an exposition of how siruc
tural equation modehlg may be apptied in ihe study
of complex neuroanaiomical networks. The examples
used are from animal experimenis with rais, bui the
Suidin8 pfinciples and the spccific approach pre-
sented also apply to the siudy of neural network
interactions in the central nervous systems of hunans
and othermammals. The focus ofthis paperis on what
is ganled from application of ihe strlrctural modeling
approach to the understanding of auditory learning.
The theoretical and techrlical issues ofstructural equa-
tion modeling have been extensively reported ir1 its
Seneral use [e.9., Bentler, 1985; Berry, 1985r Bollen,
1989; Davis, 1985; Hayduk, 1987; loreskog and SoF
bom, i979, 1984, 1989; Loehlin, 1987, Long, 19831, and
in its applicaiion to brain nraging lMclniosh and
Conzalez Lima, 1991, 1992a,b, 1993, ihis issrei.
Obiectiv€s
The basic strategy of our research proSram is to
cxamine how the same auditory stimulus afiecis brain
metabolic activity when the siimulus serves differcnt
bph"vio a' r, e- r(qurred thr.ud I ledrnifB F\peri
ences IConzalez-Lnna, 1989, 1992; Conzalez-Lima and
Agudo, 1990, Conzalez Lima and Scheich r98h-c,
1985, i986a,b; Conzalez Lima et al., 1989a,b, 1993,
Mclntosh and Corrzalez-Lima, 1991, 1992a,b, 19931.
Data from two learning paradigms will bc used as
examples to illustratc what kind of new information
may be derived from structural equaiion modeling.
The first example invoived the study of shori-ierm
and lon8 iern habiiuation of the acoustic srarile r€rlex
lConzalez-Lima et al., 1989a,b, Mclntosh and conzalez
.25.
Lima, 19911. Habiiuatiol is regarded as the simplest
formoflearninS, and it mav b€ defined as the decrease
in behavioral responding to a stimulus afier rcpeated
presentation. Habituation of the acoustic starile rc-
sponse is widely used as a model for siudying the
( ur.rl bJ.r- of be ,.r\ ror.r. lr.,brludlron lror rcvi-q . -ee
Davis and Fi1c, 19811. The second example comes from
a siudy of associative learning comparing metabolic
activity produced by the same ionc irained as erther a
Pavlovian conditioned excitor or inhibitor lMclntosh
and Conzalez-Lima, 1993, 199,11. Conditioncd inhibi-
tion is a kel' notion in ihinkint about associative
learniDg [Rcscorla, 1988]. Miller and Sped I19851
review how it has beco,nc clcar that subjects leam to
associate negative correla iions beiwecn thcjoinioccur
rence of two stimuli, and that this constituies more
ihan simpl), ihe failure to learn a positive relation.
These irvo examples illusiraie nonassociauvc and
associaiive forns oi audiiory learning depending on
presentation of single siimuli wiihout any conse
quences (habituation paradigm) or paired stimuli in
which ihe sounds signal the presence (excitor) or the
absence (inhibitor) of an aversive stimulus (condiiion-
ing paradigDr).
The objective for ihc structural equatjon modeling
analysis was to use ihese exisiinB meiabolic clata trom
audiiory learning experiments usint fluorodeoxyglu-
cose (FDC) autoradiography Iexplained by Conzalez-
Lima, 19921 to develop structural models with ihe
methods discusscd here. The daia on interregional
correlations in FDC uptake were used io consttuct the
structural models based on ihe anatomical interconnec
tions between auditory regions. The audiiory sysiem
models explained in the habituation expcriments
IMclntosh and Conzalez-Lima, 19911 demonstrate the
use of structural equaiiorl modeling io provide a
dcs.riffirc index of the functional interactions within
the auditory sysiem. This approach permitted models
to be consirucied for each experimental group sepa-
raiely, but dley could not be compared statisiically.
Therefore, it was not possible to make nrllcrttdl
staiements regarding the observed dif{erences in ihe
models. Instead, the nodels could be used only h a
Our nost recent modeling efforts have used a
muliiple group or stacked rnodel appLoach to compare
functional models staiistically, allowing for hferential
siaiements retarding dif{erences between experimen,
tal models. We hav€ used this approach successfully
with models of the visual system, as explained in
Mclntosh and Gonzalez,Lima [1992a]. The auditory
{.lem nodcl. r\plained in lhe,ondrlroninS <\peri-
ments lMclniosh and Conzalez Lima, 19931 demon-
 . Gonzalez-Lina and ticlntosh .

strale the use of an "stacked" models .rpproach and autoradioSraphy as explained before IConzalez-Lima
how the models can be evaluated infereniially. The et al., i989a,bl. With this expcrimental design we
obiective was to use the siacked models to compare tested the hypothe$is that althouSh subjects in ihe
how opposite learned associative properties oI the long-te.m and short-term Sroups received the same
same auditory stimulus are represenied by the pattern auditory stimulation during ihe postinjection period,
of network interactions bet*,een audibry system re- their auditorv systems would reflect different patterns
gions. The circuitry of the audibry system model was of aciivation related to their differeni kinds of behav-
modified by the addition of ihre€ anatomical paihs io ioral habiiuaiion to the saDre stimulus. Therefore, any
our previous habituation model to make it more group differenct's in the evoked patterns of neural
anatomically accurate. They include reciprocal paths responses io the same sound by the auditory sysiem
between the two cochlear nuclei lLorente de No, 1981, u'ould be presumed to rellect learning-related difJer-
Osen, 19871 and a descending path from the external ences in behavioral habituation, and not difterences
nucleus of the in ferior colliculus io the dorsal cochlear due b ihe sensory stimulation, which was identical
nucleus [Osen, 1987; Shore etal., 19911. durinS the testins of activation in the shorl
'reural
term and long-term grolrps.
METHODS The experimental desisn and meihods for the asso-
ciativ€ learning experiments have been explained in
Experimental designs detail by Mclniosh and Conzalez-Lima 119931. Briefly,
iwo groups of animals lvere irained with the samc
Theexperim€ntaldesitnandnlethodsforthchabitu- sound serving as a Pavlovian conditioned excitor or a
ation experimenis have been explaired in detail by Pavlovian conditk)l1ed inhibitof lLolordo and Fairiess,
Conzalez-Lima et al. [1989a,b]. Briefly,ratswere housed 19851. Rats were trained with reinforced trials of thc
in a iesiing apparatus like ihe one of Cassella and condilion(d e\citor (A J rnternri\p.t \rth nonrein-
Davis I19861. The acoustic startle respons forced trials of a compound Drade bv ihe tone and a
sured after bursts of white noise of 100 dB. The size of litht (AX ). The bne was the e\citor (A-) for th!'
this response (jump ampliiude) did not enier into the Conditioned Excitor group and ihe light was ihe
brain activity correlations. Each sessidl was made of inhibitor (X ), whereas the tone $'as thi' inhibitor (X )
270 noise bursts presented at irregular in tervals (mean, for the Conditioned Inhibitor group and the light was
10 sec; ran8e, 4-16 sec) during ,15 min. IrreSular the excitor (A+). ln this experinent, the excitor acted
intervals were needed to eliminate the possibility of as an "alarm" to signal the subsequent presentation of
temporal conditioninS. Rats were injeci€d i.p. with l8 an aversive event, whereas the inhibitor acted as a
ACi/100 8 body weight of [']C(U)]2-deoxyglucose or "safety" signal to predici absence of the aversrve
['rC(U)12-0uoro-2-deoxyglucose. The first Sroup of event. Animals show a conditioned tear response to
subjects was long-term habituated for 5 days (2 ses- the cxcitor that is suppressed in the presence of the
sions/day separated by t hr) and injected with the inhibitor- Thus the same physical tone can serve two
tracer immediately piior to their ninth session. The opposite behavioral roles dependinS on th€ associa-
second Broup of subjects was not long{erm habiiu- tive leaminS iraininS. After training, all rats were
dled lo rhc 5ound. fhe( .rnimals had no e\perience iniected with [!C(U)]FDC and exposed to the same
wiih the stimulus butwere handled like the firsi group tone. FDC autoradiography was used for imaging
in every other respect. Then they were iniected with reSional neural activity lConzalez-Lima, 1992, Gonza-
the tracer prior to their first and only session of lez-Lima et al., 19931 and to generate interregional
shorFterm habituatiorl. ln this paradigm, shorlterm correlations of activity resultinS from the presentaiion
and long-term habituation werc defined operationally ofthe ione lM€lntosh and Gonzalez-Lima, 1991, 19931.
as the habiiuaiion wiihin-session (short term) and
iha i between sessions (lon8-term). Short-term habiiua- Application of structural modeling to the
tion is a more transient modification in behavior, with auditory system
spontaneous recovery a fter stinlulus remova l, w hereas
lonS{erm habituation is a more permaneni decrement The analysis began by constructin8 an fltntotxical
in behavior that takes place across daily sessions of model of the auditory system linking th€ regions
stimulus repetition [Leaton et al., 1985]. A ihird group
through their known anatomical conneciions. Then,
of subjects was iniected like the other groups but t/'r.liorai models for ihe various groups were con-
received no stimulaiion. At 45 min postinjection, rats structed using structural equation modelinS. In the
were killed and theirbrains processed for quantiiatjve modeis, activiiy correlations were decompos€d k)
 . Path Analysh ofAuditory
calculate numerical weiShts, or path coefficienis, for
each anatomical path. These path coefficients were
th€n used to compare the network interactions for ihe
different group models. Therefore, ihe application of
structural modeling to analyze functional interactions
in the auditory system involves five major steps: 1)
construciion oI the anatomical model or path diagram
of the auditory system,2) computation oI the conela-
iion mdtri\ of d(trvrtv between ll'e auditor) region.
that form the model; 3) creation of the structuml
equaiions for the system to express mathemaiically
the vdridnce rn drlivily in each dudrroD region J- d
function of the variance of othef brain regions and
some residual influencei 4) solving the equations in
the model to find optimal soluiions foi path coetri-
cients and residuals in aSreement with the observed
coffelation matrix of activity, and 5) statistical evalua-
tion of the vaious functional models to determine
significant diffeiences due to $orp treatments. These
steps are explained below.
Anotomi.ol model of the cen''ol oudi.ory tfstem
Th€ brain of exp€rimental animals such as the rat is
well known in ierms ofanatomical connections. How-
ever, like most neurobioloSical da ia, these connections
repiesent the best available approximation of the
actual circuitry of the central audiiory system based
on the currently available experimental techniques.
Most review a.counts of neuroanaiomy are inad-
equate for building an accurate connectivity model of
the auditory system from a particular species. Pimary
source journals wete consulted io document each
path in the rat model. The anaiomical model by
necessity is a simpler approximation of a more com-
plex biological reality. Parameter identification is
geatly dependent on how the connections beiween
structures in a model are determined. This relates to
the accuracy and interpretability of ihe model. When
decidingwhich connections toinclude in amodel, one
must rely on the current anatomical knowledge of the
sysiem. Connections that are part of a well defined
path should definitely be included. For instance, in
modeling the audiiory system, the geniculocortical
connection would constitute a well defined paih. There
are, however, cases in which the connections are less
well documented or are minimal. ln these instances,
one must rely on a parsimonious judgment. IncludinS
questionable connections that "may exisi" can lead to
the problem whereby the number of connecfions
within a system exceeds the number of correlations
between the regions, so that no single solution to the
path coefficients can be calculated. In the final model,
LeaninC.
a compromise is usually reached between anatomical
accuracy and inierpretabiliiy. A model that includes
all potential connections wili be more accurate but
may not be interpretable, or worse, may not be
solvable. On ihe other hand, a model that does not
include all well'established anatomical connections
willmost likelybe solvable, but may have less theoreti-
cal utility.
The anatomical model for ihe rat central auditory
system is shown in Figure 1. The directional arrows in
Figures 2 and 4 indicate the anatomical paths between
auditoiy regions shown in Figure I All the paths
linkinS the regions were based on well esiablished
anatomical connections llor€nte de No, 1981j Osen,
1982 Coleman and Cle ci,1982 Webster, 1985, Shore
et a1.,19911. The central auditory sysiem ofihe rat,ljke
that ofother mammals, is made up of six main regions
located along the neuroaxis ftom the medulla to the
neocofi ex lwebster, 198s].
1. The cochlear nucleus, located in the laerar me-
dulla, could be subdivided into dorsal (DCN) and
ventral (VCN) cochlearnuclei in our autoradiographic
'ma8es.
2. The superior olivary nucleus in the ventral pons
could alsobe subdivided into lateral (LSO) and medial
(MSO) olivary regions. The lateral region was made
up of the lateral superior olivary nucleus, whereas the
medial region comprised both the small medial olivar\
nucleu. in lhe rdl ind ll-e ar8er <uperio-pardoh\dr\
nucleus ddjdcenl lo :r. The'e ."f'airl.on- or .1,
cochlear and olivary nuclei are important because
they showed clearly distinct patterns of activity and
could be connected in parallel io differenibrain regions.
3. The nucleus of the lateial lemniscus (LL),located
at rostral pontine and caudal mesencephalic levels,
showed a well-defin€d ventrai nucleus in our autora-
diographs.
4. The lar8e inferior colliculus in the lower midbrain
could be clearly subdivided into a central nucleus
(lCC) and an externalcortex (lCE).
5. The medial teniculate (MG)body is located in the
caudal thalamus.
6. Finally, the primary audiiory cortex (AC) is
located in the temporal lobe at approximately the
same coronat planes where the medial g€niculaie
body is found.
The \i\ main .eBion< dnd their subdivi\ion- men-
tioned were included in the anatomical model. Ii is
expected that with hiSher resolution metabolic and
neuroimaging techniques, such as cytochrome oxidase
histochemistry lconzalez-Lima and Jones, 1994] and
 . Gonzalez-Limd and Mclnlosl.

,y'\
VCN

LSO -----::- MSO

FiSure l.
Anatomical model of the.enral alditory systefr ot the rat. neural nerwork formed bI resons VCN, LSO. and MSO aid
A!dno./system re8tons ares) were placed on
(tabe ed white three their anatom ca paths is indicated by p,, Pr, and Pr labels. StructuE
schematctransvercesecrionsofthe brain (hiidbra n. midbrain, and eqoar on model ng uses iniormat on about this connectl! ty and the
torebEin evek frod lett ro righO Anarom cal pathways berween corelat ons of act v ty between r€tions to determine the strength
res ons were iidicated by schemat c 'neurons with cell bodes o{ influences throuSh each Path (es path coefflcients for p, Pr,
(cncbt within reSion and then axons (iinet p.ojecrinS ro anothef
a and pr explained in the l'lethods sectior). DCN, do6al cochlear
region ofsynapr c.ontact. These uond p.ojections corespond to nucleus; VCN, ventE <och ear nucleusr LSO, lateralslPer or o ivary
the paths (arrows) used n rhe $ru.tural modes in subseqlent nu.le!s; lvlso. medal superior olivar/ and suPerior Pamolivary
fisLfesForthesakeofsimPicrv.iPslateraandcoiralateraPalhs nuclei: LL. venra ntceus of the lateral lemnis'tsi lcE external
are showi on the same side n the fitufes. No left riSht side nucleusoftheinferior.olliculls;lCC,<entralnuceusofthe nferior
commissuE paths were n<llded in the model because no left riSht col culls; l'lc, media Senculate nuceusi AC, Primarv auditor/
side differen.es in rea ona act v ty were determined. The reduced cortex lAdaPted from YcLntosh and Gonalez Lima l993l

landscape parametric profiles LBarreto an.l Conzalez-
Lima, 19941, these regions wiLl be flrrther subdivided
in the fuiure, pcrnitting the claboration of mo.e
As indicated above, the sclcction criteria follorved to
bur.d r\r- Inrrorhrl lI'.1. 1 rr\ol\pd J pr\noniou.
judgnent about the number of regions that could be
disthSuished n1 thc nrages and in their connecti')nq
v€rified in the liierature. Specific details of how these
selection criteria were applie.l are Siven below with
examples ofinclusion and cxclusion ofsPccific fegions
and paths. The main limiting factor in the number of
rctions included in the model was tiven by the
regional spaiial rcsolution oi the autoradiograPhic
meihod uscd io collect thc aciivity data For eranPle,
small perioUvary nuclei like the ventral nuclets of the
irapezojd body could not be clearly differentiated
from adjacent reliions in the rat. Thcrefore, activity m
such small nlrclei could noi be measured and they
were excllrded fro,n thc model. Limitations of sPatial
resolution are a decisive factor in any brain imaging
study. In the case of PET studies, for instance, most ol
ihe subcortical regions included in our auditory sys'
tem model cannot be investiSaied at this time.
Eramples of ercludcd paihs were the olivocochlear
paths lwarr, 19921, i'hich in the rat oriSinare trom
neurcns in the lateral superior dive (lateral olivoco-
chlear path) ancl ventral nuclcus of the trapezoid body
(medial olivocochle.rr pdth). Boih of these paths iermi-
naie in the co.hlen. iltich is not in our autoradLo-
graphs of the Lilnin, and thus these paths had to be
excludcd from th€ mod€I. However, it could be ar-
gued that some oli\!.ochlear fiberssendaxon collater-
als frorn ihc lilcril superior olive io the cochlear
nuclei granlrlar la\e. The anatomical status ot these
collatcrals na\ Lie iLlmmarizecl in the following quota

.2E.
 . PathAnahri.,f
tion as an example of what constitutes a piih i::r: r:
less well documented or mnrimal and thus i: r\. -.-..-.
from our model: ". . . the existence of collater:1i ::.::1
LOCS (lateral olivocochlear paths) io the....:1.::
nucleus has been somewhat inconsisienth rr,::...
-u88e\rint ,\.r , r(ir rurnber \ p .r . (.\
vadable across species. . . . The existence !,i.,,i1.::.r,rli
from LOC axons io the cochlear nLr.Fu. ]1:i L.een
deriedb\ JllIr\e.lrSdror.h.rl rdrr -. ,. f -.
lwarr, 1992, p.4291. As a Beneral rule...ll;Ier:rls or
axons of questionable anatomical e\isi!..e:r the r.1t
and other rodents were cxcluded fr.nr ih. nr.l€l ]n
addition, paihways that have not Lieen \'€fifiE!t in th€
rat using modern axonal transPort te.hni!upi te.g.,
reirogradc tracing ofhorseradish Fero\r!taie or dniero,
Srade tracing of this or similar conlpoundsl rlere
excluded. For example, connections d€iennin€d \!,ith
lesion degeneration techniques roLri.t no t .l irer€ntiate
lrun d;m.rge ro cpll bodie. i. , t.I ^-,.i .o p,*rnB
axons through a region, which resulted n misleading
evidence abort the location of thc neurons whose
axons deSenerated. Thercfore, proiections described
in such lesion studies were cxcluded lrom the model
unless they were contumed rvith modern axonal
trarlsport method s in ihe rai.
Coftelarion mattices of duditory rcgionol octivities
After construction of the anatomical modcl, ihe nexi
step is io compute interregional correlations ior the
data file. lt should be emphasizcd that ihe umis used
to quantify netabdic data (e.9., isoiope incorporaiion,
optical density, glucose uiilization) are irretevant for
computation of interregional correlations bccause the
covariance relationships are the same regardless of
which linear transformation of the data is applied.
This is a major advantage of structural €quation
modeling for a wide varieiy of mapping studics
because activity units are immaterial in the assessmeni
of funciional relaiionships between reSions ot the
Correlations can reflect a bias becanse the data are
all derived from the same brain [Ho {iiz, 1989; Hor
witz et al., 1984; Soncrant et al., 19861. To correct for
this bias, a relative measure of activity such as u'hole-
brain ratios ls used. These fatios are calculated by
dividingtheactivityvalue (e.g.,isotopeincorporation)
from each brain region by the avera$ vahe for the
entire brain of that subject. This gives a better indica-
lron ot rhe tun, rio'1il d-\ocrdlron\ br'heen regr.n-
beyond any association due to a common source of
variance. Pariial correlations of FDG uptake werc
\ omPured s ,rhin cr.h EroLp .r.rns \uojecl..
AuditoryLearninS.
Siructural equation modeling is ai{ecied by sample
iize (e.t.,less than s0 [Boomsma, 198s, Loehlin,1987]).
.\ sample size of five or fewcr subjecis per group is
.ommon for brain imaging experimenis. To compen
sate for this, one can obiain multiple measures of the
sane brain region in,nany brain sections. In the
analysis of FDC autoradiographic images, more than
one rcading ofeach brain regiorl is iaken, uslrally from
more than one brain seciion. To compute correlaiiols,
the multiple sections per brain are treaied as a source
of variance in a siatistical design, as explaine.t in
Mclntosh and Conzalez-Lima [1992b]. Briefly, when
ihc within subjeci source of variance is partialled oui,
the resllltinS partial correlations are corrected for this
source of variance, which in ei{ect renders within-
subiect observaiions independent. This pariialling tech
nique can be done by rnany statisiical software pack
agcs, such as SPSS, SAS, and BMDP.
ln ihe siudies reviewed here we used threc adjacent
autoradiographic sections from each brain reSion for
each subjeci to caiculaie inierreSional correlations.
Since ihe toial number of measuremenis included
multiple lneasures from each subjeci (i.e., 3 sections/
Nbje.t), the group correlations were inflated due to
the within-subject source ofvadance. However, as the
source of this vadance was known, it could be re,
moved using a partial correlation procedure LPedha
zur, 19821. This partialling technique is comparable to
ihc variancc partiiioninB donc h ANOVA, in which
there are four known sour.es of variancc: oerween-
group, between-Letion, between-section (the latter
two are wiihin-subiect), and error variance. ANOVA
separates the iotal variance into between-snbject,
within-subjeci, and error variances. Since these sources
of variance are independent, ANOVA can assess the
statistical significance of each source rclative to error.
The objective of our partialling iechrliquc is analogous,
thai is, we assessed ihe si$ificance of the bctween-
group and beiween'region variances, with the oiher
variance sources held consiant or pariialled out- When
between-section variance is pariialled out, ihe measure
ments are rendered mathematically independent for
the pLrrpose of calculaiinS interregional correlahons.
In ihismanner the calculaied partial correlations were
conected for the extra source of beh,{een section
vadance, and the effeciivc sample size was increased
lMclniosh and Conzalez-Lima, 19941. AlthouSh the
actual value of a corelation calculaied using the
pariialUng techniqlre will differ from that calculated
with one section per region per subjeci, ihe pattern of
correlaiions will be the same. Although the patterns oi
correlaiions and €stimates of path coefficients ob
tainecl with 3 seciionslsubject are the same as ihose
 . GoDzalez-Lima and Mclntosh
obtained with 1 section/subjeci, the inferential implica-
tions of this approach are unclear [Mclntosh and
Conzalez-Lima, 1992b1. Therefore, in the statistical
evaluation of the functional models explain€d below,
the probability values should be interpreied as an
indication of the relative size of an effeci rather than
the exact likelihood ofa particulai x':value.
Stru.turol equotions for rhe auditory sysaem model
Correlahon. between acti\ ilies o{ iudilory re8ion.
are decomposed based on the causal order of the
system lBoll€n, 1989, Loehlin, 1987]. For example,
consider only th€ reduced "neural network" in FiSxre
1 made of three brain regions VCN, LSO, MSO and
their inierconneciions. These anatomical connections
provide the causal order for this network. The direct
paths show that region VCN projects to LSO and
MSO, region LSO projects also to MSO. Using path-
tracing rules outlined by Wrighi las cited in Cohen
and Cohen,1983, Loehlin, 19871 for paih analysis, this
network can be represe \ted by a set of path equntiolls,
in which the corelations between activities in VCN,
LSO, and MSO are expressed as the sum of the
compound paths conneciinS the threeregions. For the
correlation between activities in VCN and LSO (called
rr) th€h direct path coefficient (called pr) is ihe same
value, i.e., rr : p . The €onela tion between activiiies in
|
VCN and MSO (rt, on the other hand, equals the sum
of their direct path coefficient (pt plus the cornbined
indirect influence mediat€d ihrough ihe VCN to LSO
path (pr) via the LSO to MSO paih (pr, i.e., r: = pr
(dircct influence) + prpr (indirect influence). Finally,
ihe correlation between LSO and MSO (r3) equals the
sum of thei direct path coefficient (p3) plus the
indirect paths ftom VCN (pr and pr, i.e., 13 = p3 +
prpr. There are three known parameteis (the corela-
iions beiween regions) and three unknown param-
eters (the paths between reSions). These data, to-
gether with the other correlations in the entire model,
are j11st adequate to solve for the unknowns throuSh
altebmic substitution. The values obtained (pr, pr, p3)
represeni the pdfh coerfi.ierfs foi the influences among
the auditory reSions in the reduced network. Unlike
the example presented here, for most struciural equa-
tion models the number of known quantities is geater
than the number of unknowns. This limits thepossible
mnte of values a paih coefficieni can take and pro-
vides a somewhat more reliable soluiion than the
"jusi-determined" (i.e., three unknowns and three
knowns)solution presenied in the above example.
Path eqmtions help to understand what structuml
equation modelingdoes and how the path coelhcienis
.30.
.
may be computed. They represeni ihe basic method
for structural equation modeling, although the equa-
tions and procedures are somewhat more complex
lfor specific details see Beny, 1985, Jdreskog and
Sorbom, 1979, 1984, 1989, McDonald, 19801. In most
structural equation modeling programs, ihe causal
order ofihe system is represented by a set of st/r/cirlnl
eqratiors rather than path equations. ThrouSh dedva-
tion, siructural equations can be related to path equa-
tions, by expressing the correlations in raw form, i.e.,
converting the correlaiions back to variances and
covadances lfor the mathematical derivation, see
Loehlin, 19871. While path equations specify the com-
ponen ts of each correlation coefficieni, structural equa-
tions specify the components of the variance for each
auditory region in the system. An example of the
structural equation for the ieduced networkis: LSO =
VCN pr + VVCNand isread as:"The vafanc€ oILSO
is equal to the variance of VCN times a weighting or
path coeaficieni (pr) plus avariance residual source (V)
unique to VCN that is external to the model." For this
€quation LSO and VCN rcpresent known parameten,
and pr and {. represent unknown parameters that are
to be solved. The solutions for ihe siruciural equahons
dre obtained in a ha) simrlar to multiple reSre\qion
but instead of solving for each eqmtion indepen-
dently, they are solved simultaneously using matrix
computations [dreskogandS6rbom, 1979].
Solving the model for poth.oefricien s and residuols
The maihematical goal of structural equationmodel-
ing is to obtain an optimal solution to a set of
structural equations that minimizes the differences
between the observed covariance relationships (from
the correlation matrix of activit),) and those that ar€
implied by the solution. The calculation be$ns by
8ivin8 each unknown parameter a stating value.
These values are ihen used to stari a series of succes-
sive iierations as to whai the parameters must be,
given the correlaiions behleen yariables. As with all
iterative fitting algorlthms, there is no Buarantee that a
soluiion may not be at a locaL minimum rather ihan a
Slobal minimum. Holi€\er, the LISREL software
{Jcireskog and SorL'on. 19891 used for solving the
model uses a non iter.ii\ e least-squares estimation to
provide starting vaLLrei. increasing the probability that
the iterative sohrtron tinds a global minmum. As
indicated abole hrth paih equaiions, the causal oidei
$ven by the.lna(rmical connections in the path
diagram (Fig. 1r iussests sone of th€ factors that
influence the correlations between re-
gions. Using ih€ re::il.rn., dructural equation model
 . Gonzal€z-Lima and
evaluation of different models. Second, it improves
the fitness of ihe models by increasin8 the degrees of
fr€edom, due to the larSer number of obse ed values
(correlaiions) relative to ihe number of unknowns.
Each Broup's correlation mat x operates as a set of
obseryed values when ihe models are estimaied simul,
taneously. Therefore, when certain parameiers are
constrained to be equal between groups, the ratio of
knowns io unknowns is increased, and this makes th€
parameter identification of the models more likely
fHayduk, 19871. The stacked models approach will be
presented with the auditoiy turnctional models of ihe
conditioninB groups.
After the final solution for the struciural model is
obtained, the next step is io inierpret the findings in
terms of neural mechanisms. While therc is a tempta-
tion to interpret the direciion of a path coefficient
(positive or negative) as a reflection of excitatory and
inhibitory influences in the postsynaptic electrophysi-
ological sense, the interpretation of path coefficients is
besi likened to correlaiions beiween variables. When
two variables increase in size in a proportional way,
they are positively coirelated. On the contmry, when
one variable increases while the other decreases, they
are negatively correlaied. ln the case of neural struc-
tural models, the path coefficients measuring the
strentth of interconnected siructures can be defined
more specifically. A positive paih coef{icient means
that a unit increase in the activity of one structure
leads to a direct increase in the activity ofstructures to
which it is linked proporiional to the size of the
coef{icient. Conversely, a negative path coefficient
means ihat an increase in activity in one structure
leads to a direct proportional decrease in the activity
of structures to which it is connected. Thus positive
and negative path coefficients measure the direction
ofthe covariance relationship between components of
the network. A change in a paih co€fficient would
represent a change in the funciional interaction. The
de-i8n.rrion of d chdn8e in a pdrh d- dn rncrei\e i1
postsynaptic excitation or inhibition will depend on
what is known about the physiology and pharmacol-
ogy of the pathway in question. lnteryreiations are
best done considering the researcher's theoretical
stance and th€ current knowledge of ihe system.
ln the case ofstructural equation modelingbas€d on
2-deoxyglucose (2-DC) data, the models do noi explic-
itly indicate the temporal flow o{information. Instead,
structural equation modeling indicaies the patterns of
interaciions between structures within a neural sys
tem based on the functional activity measured by
2-DC and relat€d glucose analogs. This would also be
the case with other metabolic mappinS techniques
.32.
Mclntosh.
based on measurements taken over a fixed period of
time, e.8., cytochrome oxidase acrivity Iconzalez-
Lima and Garrosa, 1997; Gonzalez-Lim4 79921, c-t'os
mappinS ISharp and Sagar, 19921, and cerebral blood
flow IFox et a1.,19881.
AUDITORY SYSTEM MODELS OF SHORT.TERM
AND LONG.TERM HABITUATION
Co[elations for the auditory functional modets
were based on data repoited by conzalez-Lima et al.
11989a,b1. Three values for each auditory region from
each subjeci were obtained. Pdor to calculatinS the
cofielation matrix used aE data for the models, ihe
within-subjects variation due to multiple observations
was pariialled out as previously described. Activity
measures were obtained from the DCN and VCN, the
LSO, iheMSO, the nucteus of the LL, the ICE and ICC,
the MC, dnd prrmdry AC re8ion\ show.l in FiSrre L
Analyses were condrcted using LISREL version 6.6
lJdreskog and Sdrbom, 19841. The residual inflrences
repr€senied by V were fixed at 0-4 for all regions. To
ass€ss the possibility that these estimates were unrea-
sonable, functional models were computed with lower
values of q., and it was deteimined thai the overall fii
and the estimated path coefficients for each model
were not drastically chanted. In the 6nal models the
partial denvatives for q,s wer€ also very near zero,
suggesting that the value of 0.4 for Vs was a ieason-
able estimate. The intemction between DCN and VCN
was depicted with a bidirectional arrow egual to the
correlation between the two regions, representint the
influence of ihe cochlea. Estimates of the parameiers
were produced using unweighted least squares. For
the models, path coef{cients thai did not exceed 0.3
(absolute value) were not considered subsianiive.
Dire€t effe€ts
Both the direct and iotal effects for these models are
presented in graphic format only. Speciftc values for
each effect may be found in Mclntosh and Gonzatez-
Lima [1991]. Presenting the path models glaphically
rather ihan as a series of tables containing ihe coeffi-
cients seems to aicl in observing the pattern of effects_
The functional models are represented graphically in
Fiture 2. The focus was on how ihe lemniscal (VCN ro
ICC) and extralemnrscal (VCN io LSO, DCN to ICE,
et€.) pathwavq ilere operating in the ihre€ models
IWienberBer and Diamond, 19871.
Direct effects are grven by path coefficients ihat
quantify ihe cliect oi a given reSion on anorher to
which it is monos\ naptically connected. The tunc
 . Palh Analysis otAuditorJ LearDins.

Unstlmulated Contiol

Short-term Habituauon Long-term Habituation

Figure2.
Graphi< representations ot the functiona models of rhe aoditorydnectl/ connected. The magnitude ofrhe direcr effecr is propor-
sFtem for unstimulated conro and short- and lonS term habitua- rional to arow wldrh lor €ch path. valles for the width sradient
tion 8rcoPs- The dnectional arros indicate the anatomical paths are given n the legend at the bottom ritht of the llgure. Posirive
betwen auditory system s$!ctu.es. The bidirectional afow going path <@fllcients de shown 4 solid arows, whered neaative ones
to DCN and VCN represents reidual influences common to borh &e shown d setmented arrcws. [Adapred from l'lclntosh dd
the
reSions and includes the cochlea. vaiues represented here are Gonalez,L ma, 199l.l
dire<t effects of each sru.ture upon all regions to whi.h it is

tional model for the unstimulated control group (top.
Fig.2) can be thouSht to rcpresent thebaseline activity
of the auditory system in the absence of the acoustic
stimulus producing the startle respons€- The patterns
of interactions observed in ihis model w€re that the
influence of the "lemniscal system" (from VCN to
ICC) was apparenily minimal and ihat the extralemnis-
cal paths dominated. The unstimulated control group's
model could be characterized as a balance of positive
and negative path co€fficients- DCN and VCN had
both positive and negative direct effects. Mosi influ-
ences on AC were negative path coefiicients, while
relationships among brainstem and midbrain regions
were mostly positive.

.33.
 . Gonzalez-Lima and Mclotosh
A fundamentally different pattein of inieractions
was observed {or the shorterm habituanon runc-
tional model {bottom left of Fig. 2). All path coefficients
from VCN were positive. The positive influence ap,
peared to continue throuth ICC and ICE, and to AC
through MG, sng8esting that l€mniscal influence was
much stionger in this model- The descending path
coefficients from AC were both negative, a reversal of
thepaiiem in the unstimulaied conirol group.
The bottom right of Fiture 2 shows the model for
the long-term Broup based on data ftom habituated
subj€cts that had previous daity exposures to the
acoustic stimulus. The pattern of interactions in this
model demonstrated characteiistics of the two prior
models. While the hitherlemniscal effects appeared to
be pdmarily positive, the direct effect from VCN to
ICC was negative. The posiiive influence appeared to
be mediated through the supe or olivary nuclei,
whichbelong to the extralemniscal system. Boih DCN
and VCN had neSative and positive path coeficients,
with th€ inflnence of DCN sironger than in the
short-term model. The unique effect of the long-term
habituation model was a stront posiiiv€ d€scending
influence ofAC on ICE.
Effects decomposition
Effects decomposition provided an evaluation of
the total and indirect effects in the functional models.
Total ef{ects represent the influence of a region throu8h
both dir€ct effects and indirect effects. These effects
are ihe alfbraic sum of the influence ofa brain region
on all others throughboth direct and indirect anaiomr-
cal pathways. Examination oI the total effecis showed
that the greatesi dif{erences in the three models werc
inihe ef{ects from threenuclei:the VCN, the LSO,and
the relation oflCE withMG andAC. Therefore, Fi8uie
3 shows ihe iotal effects of VCN, LSO, and ihe circuit
amongICE, MC, and AC for the three groups.
The total ef{ects for the unstimulaied control group
(iop of Fi8. 3) showed a balance of positive and
negaiive influences. This was consistent with the
pattern ofinieractions observed {or the direct effects.
For the short-ierm group &ottom left oI Fig. 3), the
large positive effects originating at VCN were domi
nant. Totaleffects ofVCN were relatively sirong for all
possibl€ direct and indirect routes. In contrast, no
oiher brainstem region seemed to have substaniial
influence on upstream ar€as. Most o{ the total effects
for other regions were close to zero or failed io exceed
0.3. Attenuation of iotal effects existed in all three
Srorps but was most pronounced in the short term
model. This is especially €vident in the effects of ICE
.
on MC. The total ef{ect was 0.481, while rhe direct
effectwas 1.286. This leaves an indirecteffectof -0.805
(total minus direct). The only rouie by which ihis
could be transmitied is throughAC.
For ihe lonS{erm habiiuated Sroup the modifica
tion of ascendins effects is obsewed by comparison of
the toid effectsofVCN on otherregions (bortomnght,
Fig.3). The total effect of VCN decreased in hii.:l
brain re8ions (ICC, MG, and AC) The totd eiee oi
VCN and LSO was 0.794 and it was highest at ICE
(0.902) but at ICC it was rcduced to 0.282 and ai AC it
was down to 0.20. However, th€re were also large
indirect effects ofVCN. While the total effects of VCN
were ali positive, the direct effect from VCN to ICC
was negative (Fig.2), suggestintthat there was a large
indirect ef{eci, presumably via the supeior olivary
nuclei (LSO and MSO).
LSO had larger total effects in the long{erm group
relative io the short-term model (except for LSO effect
on MSO). This suggests that, in ihe long-term group,
pari of ihe influence from VCN was passed to and
poientially modified by the LSO. It may be ihat
another site of iranslormation involved the ICE since
it received a large influence ftom VCN, but did not
appear to iranslate that to the MG. Cortical influences
were also modified since the long-term goup showed
the largest descendinS influence from AC io ICE.
Discussion of auditory system models of habituation
The functional models of ihe auditory system dur
ing shorlterm and lontterm habituation showed
how the inieractions of parallel pathways chanted as
a result of repeated experience with a single strmulus.
It seems that with long-term habiiuaiion functional
influences are shifted from the lemniscal paths domi-
nant in short term habituation to the extralemniscal
pdlhwdv\. wilh .lron8 invol\pmenl ot lhe \uperior
olivary nuctei and the lCE. It also appears that the
auditory sysiem in a baseline state (control gloup)
maintains some kind of balance among positive and
negative influences.
The patterns of interaciions in the functional mod-
els can be related to the regronal paitems of 2-DC
uptake obsewed in the auditor\'siruciures lconzalez-
Lima et al., 1989a1. For example, the largest difference
in 2-DC uptake beiNeen th€ shori, and lonS-t€rm
Sroups was in th€ LSO. lt appears ihai part of this
effect may be relaied to ihe shiii oi inierachons trom
the direct route (\'C\ to lCC) to the indireci rouie
(VCN to LSO to ICC] ai rlell as the direct effect from
LSO alone. This aciilit| chante, thereiore, may be
partly a refleciion oi the shift of inieraciions related to
 . Palh Analysis orAudilory L.arniDg.

Unstimulated Control

Short-Term Habituauon Irng-Term Habituatton

/gN
Fiaure3.
Gmphic reprcsentatlon ofthe totaleffects frcm VCN and LSO and tnnsmitted through direct and indirecr anaromical routes. The
the toralinterre ation betlveen lcEwith f1G and AC forrhecontrol magnitude of rhe rota eftect is proponional to drow width for
troup (top) and shon-term (bottom eft) ad lonS-term (bottom tradient are d in Fiture 2. lAdapted
each path. Values for the width
risht) habituation groups. Eflecrs are rhe sum of all innlences from l'lclntosh ad Gonzale-Lima, l992b.l

processint and modification of the ascendint sitnal sho\rn in the functional model. For example, the LSO
lMclntosh and Conzalez-Lima, 19911. The increase rn regional activaiion may be related to a pedpheral
activity in LSO may also rcsult from changes not ieedback to th€ cochlea, not measured in our model,as
related to shifts in central connectivity interaciions as in ihe case of activation of the descending olivoco-
 't)
. Gonzalez-Lim! add Mclnlosh .

chlear pathways lconzalez-Lima et al., 1989a1. It is aptic inhibition of the input pathway, as suggested by
worth meniioling that our auditory sysiem model is intracellular unit recordinS studies in invertebrates
designed to quantify interactions in ccntral pathwa)'s, lKandel and Schwartz, 1982]. Instead, the models
and not changes mean reSional activity due io rcvcaled that behavioral habiiuation in response to a
activation of peripheral outpuis. It is obvious ihai to solrnd involves ar active modification of complex
a ssess central peripheral interaciions, activity measure- interactions between different auditory pathways.
ments o{ peripheral targets are required. When such
data become available they can be incorporated in ihe AUDITORY SYSTEM MODELS OF ASSOCIATIVE
model. However, addition of a peripheral ouiPut CONDITIONED EFFECTS
paihway to the present anatomical rnodel iviU have no
effeci on the solved path coefJicients of the central The findin8s from Mclniosh and Corlzalez Lina
paihs. Thai is because ihe residual influences used in li993l arc summarized here with more emphasis on
the calculations of ihe siructural equations already neurobiological inierpreiations than computaiional
account for any influence not measured in the model. aspects. The puryose was to show that struciural
Experimenis with simulated data showed thai any equaiion modelinB permiis much nlore information to
effect of adding such an output path like the oiivoco- be obtahed irom any neuroimaging experiment in
chlear bundle, for example, will be on the residual which functional daia arc measured from a complete
influence on LSO in the model rather than on the neural system. We used structural equaiion modeling
ceniral a ditory path coeFficients Isee Mclniosh ancl to provide €vidence that network interactions in the
Gonzalez-Lima, this issuc, for ciiscussion with sjmu auditory s],siem are parily related to the associative
lated datal. significance of auditory stinuli- The interregional
Together with modifications at ihe level of the co elations of aciivity for the functional models were
superior olivary complex, difterent Patterns of inierac- obtained usinS auditory system FDC uptake patterns
tions beiween paths acting on ICC and ICE and ihe in response to the same ione trained as eiiher a
influences of these areas on MG and AC may suggest Pavlovian conditioned excitor or inhibitor. Condi
further modificaiion of the auditory signal ai the mid- tioned exciiaiion consists of presentinS two stimuli
brain level of ICC and lCE. This is consistent with re- with one stimulus (A ) always paired with another
gional differences in 2-DC upiake in short- and long- stimulus (an aversive foot shock). The A then be-
term groups in ihe ICC found by Gonzalez-Lima ei al comes an "alarm" signal ihai predicts the occurrence
[1989a]. The total effects from VCN to output regions of the aversive stimulus, orUS. Conditioned inhibitiorl
appeared to be relatively stronS in the short term dif{ers in that the stimulus (X ) predicts the abscncc of
group- However, in the long-term group, this Pattern ihe US. We used a procedure in whi.h X is trained by
of influences appeared to be modified substantially at compound presentation with A. (Pavlovian condi-
various sieps in the pathway. The descendint influ- iioned inhibition, A-/AX ). This associative learninS
€nce of AC io ICE may have acted to attenuate the involves contrasting A trials with trials in which
€ffects. This influence may rcduce the arousal co'nPo stimuli A and X are presented bgetheras a compound
nent associated with the signal since ihe ICE is Pari of that signals the absence of ihe US. This results rn
ihe cxtraiemnis.al system [Weinberger and Diamond, siimulus X acquiring the behavioral properiies of a
19871, and this system is proposed io relay nonsPecific conditioned inhibitor (a "safety" siBnal). This para-
qlralities of the audjtory stimulus related to stimulus digm was Lrsed to train the same physical tone as a
siSnificance lParham and Willot, 1990]. ln the short conditioned excitor (A.) for one Erorp and a condi-
ierm group, the total effect of ICE was larger ihan the tioned inhibitor (X ) ior anoiher. It was used io iesi
long-term 8roup, suBgesting that the aroLrsing comPo- ihe hypoihcsis ihai nctllork interactions between
nent of the siSnal was transmitted io hjgher alrditorv aLrdibry svstem regi'rns rvouLd reflect thc differentiaL
centers in the short-term 8roup, but noi in the long' associative prop€rtie! (i.e., alarn vs. safety siSnalinS)
term habiiuaied group. From these models it was clear of the same phYsical ione.
lhdl habrrudtior tn o r.peared r.und i. nor r p.r'.i\e Followirlg conrpleiiol of ihe FDG uptake per'od, rat
processlinked to a simPle decrement in brain llrnctjon- brains wcrc reno\ ed arld processed for autoradiogra
in8, as suggested by early elecirophysiologjcal stuclies phy, as eaplained in .letail elset{here [Conzalez-Lima,
lHernandez-Peon, 1960; Condon and Weinberger, 19921. FDC uptake in the auditory system was mea-
i9911. lt is also unlikely that the nultiPle changes sured Lrsing qu.lnhtaiiye imagc analysis of isoiope
obse ed in ihe ceniral auditor], system in lonS-term I 'c,In..rl. r. U ,' | '. .,udrr.'ry -y-r"m re8i,', r Ind
habituation can be accounted for merel]' by a prcsyn- lyzed Ncrlr the DCN and VCN, the LSO, a more

.16.
 . Pnrh Anah5is ofAuditory
medial region coniaining ihe MSO, ihe ventral nuclcus
ol .hF I L, the ICF ,r d ICC rr,L \,1C .,nd il,( p, inr. '
AC. To hclp reduce interanimal variaiiol rn the
,,r(c, ,,r) of l-DC ldbtslint rn,,l".Fd lo,hF c\pFnn,
r,il m.n,puldlron rC vdlur. tron- ed.h b,din reSi' I
were divicled by the average rC value for the ('hole
brain of each annral (whole brain ratio). Thi\ proce-
dure was validated by determining ihai thcre were no
sysienatic dif{erences in the r{hole-brain means be-
tween the groups (i(13):1.sa, P > 0.10). Within
group interregional correlaiion matrices were based
upon ihree obscrvations for each audiiory region for
each slrbject. Irartial correlation nairiccs were com-
puted as before [Mc]ntosh and Gonzalez Lima, 19911:
the variance due to multipLe observaiiols was re-
moved, thereby increasinS the number of observa
tions per group by a facior of 3.
LISREL version 7, fronr Scientific Software lnc.
Ildrcskog and Sorbom, 19891, was used b compute
naximum-Likelihood estimates for the path coeffi-
cients representing the influences through the anatomi
cal pathways. Statistical comparison of the modelswas
done using ihc multiple Sroup or siacked models
approach whcreby path coc{ficienis were constrainecl
to bc cqualbetween condittuns (nullhypothesis model)
with ihos€ in which the coefficients were allow€d io
differ (alternative model). Each model was assessed
for its ability io reproduce ih€ original correlahon
matrix ihrolrgh the X2 goodness-of-fit statistic. If the
differencc in the X: value for the null and altcrnaiive
models was significant (X],ft test), then the coefficienis
that were allowed io vary beiween conclitions were
statistically differeni ILoehlin, 1987].
Direct effe€ts
Figure ,l graphi.ally represents the flrnctional nci
works for ihe Conditioned Excitor nrodcl and the
Conditioned lnhibitor model. The values for the direct
effects irl thc figure are based on the standardized
-^ urior rn whirh .h, bcrween-tsroup
equated to a common scale ranging from zero to onc
in absolute magniiude. Tables wiih the speci6c values
for each effect may be found in ihe oi€linal report
lMclntosh and Conzalez-Lima, i993J.
Differences betwecn the exciior and inhibitor mod-
els were in ihe extralemniscal ouipui of VCN. The
direci lemniscal paih from VCN to ICC was ihe sa e
betweenmodels. Exiralemniscal paths fronr VCN were
manlly positive in thc Condiiioned Excitor, with the
largest positive paih coefticieni to LSO. li€laiive io
this, the influences of VCN in th€ Conc:titioned Inhibr-
ior showed differences in both nagnitucle and sign.
ar
learning.
lhF rriluen.e uI \C\ .'n LcO .howed " lrrgp de-
creasein lnagnitlrde to near zero in the Conditioned
Iniribitor. Coefficients from VCN to LL and ICE wcre
negaiive in the Conditioned lnhibitor, but the abso-
lute magnitude of ihese influences was similar to the
Conditioned Exciior.
Both associative models showed strong descending
paths from AC b lCE, and frorn ICE to DCN. For the
Conditioned Excitor, these descending paths helped
to form a positive path coeffrcients loop involving the
direct paths from DCN to VCN to LSO to ICC to MG
and then AC, and completed with the descendinS
paths (AC b ICE to DCN). This loop's influences were
discontinlrous for ihe Condiiioned Inhibiior, wiih the
path from VCN to LSO reduced io zero.
Efects d€comPosition
For each auditory region, iotal and direct effccts as
$el d. v"lup\ tor ,1, (dl]^sed .hdn8F :1 i\r-
'o
application) are used for effects decomposition. Total
effects represent thc impact of a region ihrough both
direci eftects and indirect effects. Total effects are the
algebraic sum of the influences of a region on all
othcrs throuth both direct and indireci anatomical
paths. From the Xr dif{erence tesi ihe output from
VCN and ihe residual influences (V) on DCN were
found io vary significanil)' between ihe Excitor and
Inhibiior. The differences in the direct effecis from
VCN were significant upon comparison (x:1"(3) =
12.68; P < 0.0s). The value of V for DCN was
significantly larger in ihe Condiiioncd Inhibitor (0.90)
compared wiih the Conditioned Excitor (0.292; X3ir
(1) = 5.02; P < 0.05), so less of the variance in DCN
$'as accounted for by the alrditory paths in the
Conditioncd Inhibitor model.
Differences in between-group total eftects were
evaluated descriptivell, since there is currently no way
to evaluate directly ihe statistical siSnificance of such
differences. However, sincc thc direct effects were
staiistically differcnt, it is likel], that the total effects
lvere as well. Our criterion for a substantive difference
in toial ef{ects was an absoluie value of 0.3 or greaier.
The greater differences between the iwo functional
models were in the effects from VCN and DCN. This
comparison displayed an interesting paiiern of re-
sulis, illustrated in Figure 5, in which the iotal effects
are prescnicd graphically. These effects represent the
sun of direct and indirect influences and are not
meant to denote direci anaiomical pathways.
VCN mainly exhibited sirong positive total effecis
for the Conditioned Exciior, while thcse were nega-
ii\c ior ihe Conditioned lnhibitor (Fi8. 5). These sign

. Gonzalez-Lina and
Condrtbncd Exclto. Model
Path
iff","
Left: cEphic repr$ertation of the direcr effects ror the Condi-
tionedExcitormodel.|lagnitudeofrhedirectefiectisproponional
to arow width for sch path. Vaues for the width gEdient are
given in rhe egend ar the bottom of the 1lgure. Positive path
coefficients are shown d solid afows, whers netative ones are
chantes were not accompanied by larte differences in
absolute magnitude, with the excepiions of the ICE
and LSO. The direct effects from VCN to ICC were
negative for both models, bui ihe total effe.ts were
positive in the Conditioned Excitor (0.404) and nega-
iive in ihe Conditioned Inhibitor ( 0.487). The posi
iive total effect in the Conditioned Excitoi model
rcsulted from the lar8e Posiiiv€ indirect effecr rra n qm ir
ted through LSO and LL (indirect e[fects = iotal
direct, 0.647). The direct path from VCN to LSO was
smaller inihe Conditioned Inhibitor (cf. Fig.4) and the
indircct effect was small (-0.14), so the total effect did
not differ much from the direct ef{eci. The siSn
reversal ofthe totaleffects ofVCN on ICC maysuffiest
that the ascending influences were modified in paral-
lel pathways throrSh the Lso and LL.
In the Conditioned Excitor, the DCN showed mod-
erate total effects on most areas excepi ICE (e-8., total
eff€ct on ICC = 0.303, on ICE : 0). However, for ihe
Conditioned lnhibitor DCN had zero total effecis on
9
I'lclniosh.
Condltlon€d LrHbltor Modcl
Ccn lents
__:Y","
shown 6 seSmented arows. Righe GEphic representation or the
direct efiects of the Conditioned Inhibitor model. Direct effects
that were sisnifi.antly different from the Conditioned Excitor
hodelwere the Paths from VCN to LL, lCE, and LSO. lAdapted
from I'lclntoshand Gonalez Lima, 1993.l
most regions except for ICE { 0.239) and VCN. Th€
direct effects from DCN did not dif{erbetween groups,
suggesting that the DCN pattern of toial effeci differ-
enceswas a function of theinfluences through parallel
pathways, prcbably mediated via ihe indirect effects
throuth VCN. The differences in ihe total effects of
VCN between conditions were generally more in sisn
ihan in magnitude. The differences in the total ef{ects
of the DCN were more in ma8nitude and not sign.
Discussion ofauditory system models ol conditioning
While DCN and ICE i'ere ihe only r€dons that
show€d mean ditlerences in FDC upiake between
Bioups/ structural equaiion modeling demonstrated
that differenc€s beille€n ihe audiiory system interac-
tions oiginated at the le\el of the iwo cochlearnuclei.
ln the Conditioned Inhibitor model, paihs from VCN
to LL and ICE !ho\ ed direct effects opposite in siSn
and a reduciion in th€ eilect on LSO.
.38.
 . Path Analysh ofAudito.y LearninC.
Condiuoned Excttor Modei
Fitur€ 5.
left: Gdphic representarion of the total effe.n from VCN and
DCN in the Conditloned Excitor model. Totalefiecrs are the sum
ofallinlluences t€n5min:ed through direct and indirect anatomicl
routes. The matnitude ofthe total efiect is proponional to arow
The residual effects on DCN were also higher in the
Conditioned lnhibitof. The larger residual value for
the DCN in the Conditioned Inhibitor model may
have occurred by a combination of internal processes
within DCN and the influence of functional interac-
tions of the region with extra-auditory re8ions not
included in the model. Thereis anatomical evidence of
a cuneocochlear pathway in the rai brain [Weinberg
and Rushoni. Lc87l hom the cunedre nucleus. a pri
mary somatosensory nucleus, to the DCN. Since a
relevant source of extra-auditory influences du nt
conditioning was the footshock (US) effect, this direct
connection with ihe DCN provides an opportuniiyfor
ione-shock interactions at this level. Extra-auditory
interactions would be expected to be stron8er in the
Conditioned lnhibitor model because the action of the
tone condiiioned inhibitor would be on systems out-
side the audiiory system.
Network interaction effects within ihese models
maybe interpreied based on direction and magnitude
diff€rences in th€ two auditory subsystems, lemniscal
and extralemniscal lcraybiel, 1973]. The lemniscal
subsystem (VCN to ICC to MG) is associated with
Processin8 of the "pure" sensory components of
stimuli, while the extralemniscal paihways relay the
"nonspecific" components of a stimulus, which may
Tot2l Dfects
Conditloned Inhibitor Model
width for ea.h path. Values for the width tEdient are d in Fiture 4.
Righe Gmphic repr$entation ol the total effe.G from VCN and
DCN inthe Conditioned Inhibilormode. [AdaPted from l'1.litosh
and Gonzaiez-Lima, 1993.1
involve stimulus siBnificance and arousal [1\'einberg€r
and Diamond, 19871. The direct effects did noi difi€r
between Broups in the lemniscal paihwa|s. Since ih€
physical qualities of the ione iLere ihe same tor ihe
two groups, ii is expected that ihe djreci iniluences
wiihin this subsystem were similar for the tlLo SiouPs.
However, the absoluie magnitude of ihe iotal €ffects
differed in sign, in aSreemeni *'ith the opposite
associative properties of the ione for the two Sroups.
Thus, the direct lemniscal paths from VCN appeared
to act to relay invaiant information about the physical
qualities of ihe stimulus, while the opposite associa-
tive information may have b€en relayed throuth the
parallel paths.
Th€ network intera€tions revealed that the associa-
tive value of the tone influenced the maSritudes of
both direct and total effects along the parallel extrale-
mniscal pathways. These associative netwoik differ-
ences were reflected in the interactions from the two
cochlear nuclei, and these findinSs could not have
been seen by simple calculation of mean r€gional
effects. Th€ dbect effectsofVCN on the extralemniscal
regions LSO and ICE were smaller (LSO) or differeni
in sign (ICE) in the Conditioned lnhibitoi model,
while the tolal effecl: of VC\ on lhe\e region: sere
zero. The total effects of DCN were smaller in the
 . Gonzale2-Liha and Mclnlosh
Conditioned Inhibitor modelbecause ofthe change in
the indirect effecis of DCN ihrough VCN (i.e., the
direct effects of VCN). An exception to th$ was the
total effect of DCN on ICE, which was negative in the
Conditioned Inhibitor model compared with zero in
ihe Conditioned Excitor model. Both the DCN and
ICE showed relatively less FDG incorporation in the
Conditioned lnhibitor 8roup. For the DCN, this differ-
ence in mean FDG rptake was consistent wiih the
relatively lower iotal functional influence of DCN on
most oiher auditory system reFons. For the ]CE, the
difference in means was consisient with sitn chan8es
in the directeffect ofVCN and the total ef{ect ofDCN.
The combination of both lower mean activity and
negative covariance influences in the Inhibitor model
may be taken as an indication of relatively weaker
functional couplint in the extralemniscal pathway for
ihe Condiiioned Inhibitor model. They furiher sng-
gest a dominant role for ihe cochlear nuclei and the
ICE relaied to functional processing of the associative
significance ofth€ sornd after conditioning
GENERAL DISCUSSION AND CONCLUSIONS
We envision confirming the results obtained with
FDC dutorddiotrdphy u\int olher melrbohc mdppinS
techniques such as .ytochrome oxidase histochemrs-
try lconzalez-Lima and Garrosa. 199']; Conzalez-
Lima, 19921 in adjaceni sections of the same subjects
us€d for these studies. However, using multiunit
recording or electroencephalo$aphy, for instance, is
likely to rcsult in different data sets difficult to com-
Pare with data obtained from metabolic mappint of
b.ain energy metabolism. This difference is noi related
to the meihod of struciural equation modeling but to
the kind of activity measurement technique used to
gen€rate the data used for the models. For example,
meiabolic mappint iechniques such as FDC and cyto-
chrome oxidase mapping refleci cellular energy meta-
bolic demand, a functional index that is berrer corre-
lated with overall cell membrane ionic pumpinS than
axonal action potentials Isee Gonzalez-Lima et al-,
1992, for discussions on this issu€1. tf this is the case,
membrane el€ctrical activity can be highly correlated
between a brain re$on "A" sendint action potentials
thai stimulate a target region "B" as revealed by
coupling between metabolic activity (action poten
tials) in "A" and me tabolic aciivity (perisynaptic poten-
tials) in "8." This tunctional coupling can be visual-
ized by metabolic mappint of FDG using correlative
analyses because the FDC uptake does not isolate the
different cellular events (i.e., a€tion vs. perisynaptic
potentials) contributing to ener8y utilization. On the
.
other hand, if a unit recording techniqu€ that isolates
aciion potentials is us€d to monitor aciiviiies in re-
gons "A" and "B." it may fail to deted any functional
activity in "B," ifno action potentials are t€n€ratect in
"B" as a resuli of the stimulation from "A." Conse-
quentlt a different correlative index of aciivity canbe
obtained between regions "A and B" depending on
the kind of activity measured. Cl€arly, metabolic and
unit iecording techniques monitor different cellular
events and therefore may not always be expected to
reflect the same correlative relationships when ana-
lyzed wiih covariance methods as proposed here-
Metabolic mapping techniques that map electrical
activity changes throughout cell membranes obtain a
comprehensive functional index of activity, but extra-
cellular unit recording techniques, for instance, isolaie
pariicular electrical events such as action potentials.
Caution must be exeicised in comparing data sets
ieflectinS difjerent phenomena. Validation of the ana-
lyiical method of structural equation modeling is
betier pursued using computer-simulated data sets
with known r€lationships, as explained in the compan-
ion paper lMclntosh and Gonzalez,Lima, this issue].
For example, the issue of accuiacy of the anatomical
model has been of great conc€rn to us, and w€ have
recenily conducted daia simulations to evaluate the
impact of incomplete anatomical models with omitted
regions or paths. ln general, it appears that the
exclusion of re$ons or paths does not affect the
estimates of path coefficients themselves, brt rather
the estimate of the residual influences lMclntosh and
Conzalez-Lima, this issuel.
The studies reviewed showed that not only do
learning-related effects occur in the auditory system,
but also ihai the nature ofthese effects depends on ihe
behavioral significance of ihe auditory stimulus (i.e.,
habituated, excitor, inhibiior). These and previous
studies lconzalez-Lima, 1989 and 1992 for review]
indicate that multiple levels within the auditory sys-
tem show learning changes. Furthermore, the use of
structural equation modeling permitied characteriza-
tion of the specific funciional interaciions amont
parall€l neural pathNavs that presumably reflected
the associated behaIioral pfoperties of the tone stimu
lus. It is worth mentioning that the conclusions de-
rived from ihe auclitorl mod€]s reviewed will noi be
affecied by further in\ estigating extra-auditory mod-
€ls of r€dons in|ohed in siartle or conditioned re
sponses, since ihe Soal of the piesent analysis was to
see if auditor\ s\siem rnt€ractions were affected by
stimulus sitniticanie. and not to identify all the brain
regions inYoh€.1 in itarile or conditioned behavior.
We are presefil\ lo.runt at extra-auditory interac-
 . Path AnalJsis ofAudilory Learning.
iions in these paradigms lNlclntosh and Conza]ez
Lnna,1994l.
Even a felatively sinple ionn of Learning such as
long-term habituation involv€d multiple parallel neu-
ral pathways acting together in the iniact brair. From
this it may be concluded that auditorv brain reSions
use a combine.l network strategv io represent specific
changes in stimulus siSnificance. These distribuied
chan8es in auditory pathwav int€ractions are not in
.Sreemenl wrlh lhe older \res. ot lp"rninB d. rF
stricted to a local activity chang€ limiied to "hiSher-
order" reBions or to some specialized neuroanatomi
cal system with a memory-based mechanism. The
funciional models of the auditory system suggest ihat
purely audiiory sensory responses io the sound may
work in parallel with nemory-based mechanisms
taking place in the auditory system [Conza]ez-Lima,
1989, 19921. The findings in the conditioning study
may be compared with the functional neiwork model
for long-term habiiuaiion of the acoustic stariie reflex
lMclntosh and Gonzalez-Lima, 19911. Parts ofboth the
Excitor and Inhibitor models in the conditioninS
experiment resemble the model for long tern h2bihra-
iion, in which we found that the lemniscal path from
VCN to ICC also had a netative path coefficient arld
the posiiive influence was throuSh extralemniscal
paihs, especiaily fiom ihe superior olivary nuclei. This
simihniy may imply that this pattern of interachons,
involving paths parallel to the more direct lemniscal
paihway, occurs whenever an acoustic stimulus has
some acquired mcaning (habituated or conditioned).
Descending cortical paths also had strong effects in
the long ielrn habituation models and in both models
in ihe present study. Funciional changes in the audi-
tory cortex in learning paradigms have suggested ihat
auditory cortex plays a central role in learning related
plasticity in the auditory system [Wenrberger and
Diamond, 19871. The funciional models in the long-
term habituaiion and the present conditioning siridies
provide support for this hypoihesis, as ii appears thai
the descending cortical pathways from cortex become
functionally stronfr in va ors conditions $'hen an
acoustic stimulus has acquired m€anin8.
The various ftrnctional models developed using
structural equation modeling [Mclniosh and Gonzalez
Lima, 1991, 1992a,b1 sugSest that parallel paihways n1
the auditory system interact to process ihe physical
(lemniscal) and behavioral (extralemniscal) compo-
nenis oi acoustic stimuli. Sirong functional rnterac-
tions between parallel sensory pathways have been a
consistent finding in all structural equation models of
brain imaSing data pe ormed thus far in auditory
lMclntosh and Conzalez-Lima, 1991, 1992b1 and visual
systems lMclniosh an.l Conzalez-Lima, 1992a; Mcln-
tosh ei al., 19931. Totether, these studies dcmonsirate
that whilc a certain paihway or area nlay be more or
less critical for a particular function, operations in ihe
inld, I br"rn I.r\.'l!( inlerd.lion. dmonB n'dn) reBionr
and at nany levels. Our findings suPPort the hyPoth
esis thai auditory learning is an emergent property of
distribllted brain activity ancl network interactions
lconzalez-Lima, 1989, 19921, and thai such an emer-
gent property cannot be understood solely by consid-
ering ihe component parts of the brain one at a time
Alargenumberof earlyelectrophysiological studies
have shown the development ofPhysiological change
at many levels of ihe audiiory system during learnhg
Ie.g., Hernandez Peon, 1960, Buchwald et al., 1966,
Olds et a1., 1972; Gabdel et al., 19761. More recent
studies by Weinberger et al. [1990] have accumulated
evidcnce of learning-relaied plasticity of neuronal
rcccptive fields distributed h the mammalian alrcli-
tory system using single unit recordints. For insiance,
the receptive 6elds of auditory MC neurons show
retuning following associative learning in Suinea pigs
lEdeline and Weinberger, 19911, while similar recep
lr\L field pla.l crl\ n.r- bLe.l rF.orded '-on' P-in-ar!
AC neurons in cats and grinea pigs lBakn and
Wpinberger, loo0l. loB,Fll-e' r\e.F Lle.trJLh\-rol r-
ral.tudie. and our n'e,dbol. n,.,pFing - | li' -.,r-L-'
thrcL B(ne,dl conclu.ron-: .'f(,.,',1 .f"1'- '
rrceplivF field proDerriL- i1 r.dirol - r. 1
n.odiiiFd ov . ur.EL, in rl. -1p ,rnj . a .
stimuli acquired through learningr l) liiher€ni r€grr..i
and paihwa_v.'s within the audLior! :\ it.nr sho\ ! er!
specific changes ln acti|ii]. ancl iunctiorlal interactirrs
relatcd to l€arning paradigms, ind 3lelectroph!ridogr
cal as well as metabolic corfelai€s oi th€se l€aLnint
effectscan occur as carly as in ihe first regional statrons
of the auditory system. The view held once that ihe
auditory system was made of nonplastic neural ele-
menis erclusively concerned with deiection of physi-
cal stimulus properties is no longer tenable. lt should
be replaced by a more inclusive view that includes the
involvement of sensory regions in behavioral siimulus
properties acquired by leaming.
The combined use of functional brah mapping and
a na tomically based siruciural equa tion modelinS h old s
great potential io promoie our un.lerstancirng ot
brain-behavior relations at the system level. Stru.tur.il
n,oJelI.rd pen,,il- rr-edn her\ lo ldlp
"d\dr.td-.
funciional data sets from compleie neural s\sieln! t\
quantifyingthe networkirieractions. Presenth..te.
rurdl pgudlron modclrnS cir provide n-o-p.l . I
informaiion aboui functional influences medi:ri.i :.
.pe,r6. n(uroanatomi.dl pdlh!\d\- r\.n " r :
 . Gouale2-Liha
mathematicat technique available to analyze neuml
network interactions. Metabolic maPping techniques
have the potential io provide us with daia ihai were
noi availabl€ in the Past. With new data acquisiiion
iechniques there must also be new data analysis
iechniques that can take advantage of these opPortu-
nities, such as structural models. These new tech-
niques cannoi, however, oPerate alone to give clues
about brain function. The researcher must combine
these new iechniques wiih a solid theoreticalbase and
a good working knowledge of neuroanatomy and
neurophysiology to provide a better understanding of
how the brain op€rates to achieve dif{erent behaviors
ACKNOWLEDGMENTS
This work was supported by NSF grani 1BN9222075
and NIMH grant R01MH433s3.
REFERENCES
Bakin IS, WeinberEer NM (1990): Classical .onditioning induces
the
CS sPecific recePiiv€ 6cld Plasticity in th€ atrditorlr
"der 't
glin€a pig. Brain Rcss36:271 286
Brreto J, Conzalez Lina F (199'l): A landscaPe Paranetric Prohl€
approach for rat brain ihaSe analysis. Neurolnage {in Pre$)
Bentl€r PM (1985): Th€ory and ImPlementation of EQS, A Shuctural
Equations Program. Los Angel.si BMDI Stahstical sortwir..
B€rry WD (198,1)i Nonrecursive Causal Modeh Sag€ Univ€rs'tv
laperseries on Quantitative APPlications in the So'i:l s'ien'€\
Bevcly Hilk: Sage Publications
Bollen (A (1989): Structural Equations with Latent Varl:bler Neu
Boomsma A (1985), Nonconvergence, mProPei soruions, and
startin8values in LISREL naxinun likelihood €stiharion Psv'
clrometnka s0:229 2'12
and
Buch{ald JS, Halas ES, Schramm 5 (1966): ChanEes in corti.al
{rbcortical !nit activitl durinE behavior.l conditioninS. Phvsiol
Ca$€lla JV, Davis M (1986): Th. d€si8n and calibration startle
m€asurement system. Physiol Behav 36:377 383
ClarkCM, Kessl€r R, Buchsbaon M, MarSolin R, Holconb H (1984):
Correlational methods for det.rmining tuuPlin8 ol r€$on.l
gltrcose netabolhmr A pilot stldy. Biol Psychiatry l9663J7ll
Clark CM, Anmann W, Martin WRW, Ty P, Hayden MR (1991) The
FDC,TPET nethodolo8y forearly detection of dise:sP: A statirti
cal nodel.C€icb Blood llow M€tab 11:a96 tl02
J
Cohen l, Cohen P (1983): APPlied MultiPlc ReEression/Coiielation
Analysis lor th€ Behavioral Scien.€s Hillsdalc, NJ: Laltrencc
Erlbaum Asso.Dtes.
ColemanlR,CleftiWJ(1987):Sourc€olProjcctionstosubdivisions
of theinferioicolliculusin therat. JComPNeurol262:215-226.
Condon CD, WeinberScr NM (1991)i Habituation Produces trc
q!€ncy{pecific Plasticity of r€c€Ptive 6clds in the trditorl'
cortex. Behav Neurosci 105:416-430
Davis lA (1985) The Logic ol Causal Order' Sag. Universitv PaPer
Series on Quanrftadve APPlications in th€ Social S.iences.
Bev€rly Hillsi Sage Publi.ations.
.12.
and M.ldtosh.
DavislA, File SE (1981) lntinsic Jnd ertrinsic mechanisms oi habit
uation and scnsitization: lmplications for thc design and analy
sis ot expernnenc In: Peeke Hvs, P.trinovich (eds), Habitualion,
Sensitizrtion, and B€havior. Orlando: Acidemic lre$, PP 287 323.
Ed.llneJ-M, W€nrberger NM (1991): fhalamic shoir tein Plasticrty
in the audiktrysyst.n: Asso.iativ€ reluningofrecePlive 6elds in
theventral medial Seniculat€ body. BehavNeurosci 10s 618 639.
Iox PT, Mintun MA, ReiDan EM, Raichlc ME (1988): Enhnnced
d€te.tion ol t.cal biain responses using intersubject avern8'ng
and change dGtribution analysis ol srbtracted PET inages J
Cereb Blood Flou Mctab 3:642 653.
Friston KJ, Frith CD, Liddle PF, Frt.kowiak RSJ (199r): Conparing
tunctional (PET)im.8.s: The asse$mentol siSnificant chan8c l
cer€b Blood Flow Metab l1:690 699.
frislon KJ,Irith CD, Liddl€ PI, Frackowiak RSJ (1993): Functional
connectivity: The prnrcipal comPon.nt analysis of larS€ (PET)
datascts.J Cer€bBlood FlowM€tab 13:5 14.
Gabri€l M, Miller JD, Salhvick 5E (l976) MultiPle unit ictivny ol the
rabbit n€dial Eeniculate nucleus in ..nditionin& ertin.tion, and
ieversal Physiol Psy.hol,1:124 13.1.
Gersrein CL, Perkcl DH, Subrananian KN (1978): Identificntion ot
lunctionally relat€d ncurrl assenrblies. Brain R€s 140:'13 62
Cevins AS, DoylcJC, Cuhllo BA, S.hafler RE, TannenillRS, Br€$l-"r
SL 11985): Neurocogniti\,€ Patlern analysis ota visuosPahrl task:
Rapidly shifling loci olevoked correlations between electrodes
l\ychophysiology 22:32 -13.
Gonzalez LimaF(1989):Functionrlbra'ncircuitrtrelatedtoatuusal
and learninSin rats. ln: E$ertl-P, Arbib MA (€dt: Visromotor
Coordination: Amphibi.ns, ComparGons, Models and Roboh
NewYork: Pl.nun lress, pF 729 766
Gonzalez l-im, F (1992): Brain ima8in8 of auditoiy learninS Iunc
tions nr rah: 5t!dies with nuorodeo\ySlucose autoradrograPhy
and cytochrone oxidase histo.hemistry. In, Go.zalez Lima r',
Finkensta.dt T, Scheich H (eds): Advances in Metabolic MaP
ping T€chniques for Brain lmaging of B€havioral and Learnin8
Functions. NATO ASt series, Vol D68. Dordrecht: Kllwer Aca
.l€nic lublGh€r, PP 39 r09.
Conzalez Lima F, Agado I (1990): Functionalreorgan'rahon ol neural
auditory maps by difiei€ntial learninS. NerrorePort 1,161 164
Conzalez Linr F, Gaiiosa M (1991): Quantitative histochenistry ot
cytochiome oxidase in rat briin. Nenrosci Lett 123:251 2s3
Conzalez Limi F, iones D (1994): Q!.ntilativ€ n"PPirg ot .yto
.hione oxidas€ actnity in the cenlral auditory systen or the
gerbil: A study with .alibratcd a.hvity standards and metal
intensili€d histo.hcmistry. Brain Res 660:}u9.
Gonzalez Lima F, S.h€ich H (1981a): Functional a.tivation i'r the
audilory sFtem of the ral troduced by arousinS reticular
stimulation:A2 deox)glucosenudv BiainRes229?.o1 2rr
Conzalez-Lima F, Schei.h H {198rb): Neurat subsirates for tone_
conditioned bradr.ardia demonshnt€d with 2 deov8lu.ose. I
Activationoiauditorvnucl.i Beha!BrainResr4:211233.
Gonzale/ Lina F, S.h€i.h H (l98rc): Cla$ical con.litionrng en
auditot l,l!.\rglucose Faiterns in the inleior colli.u_
hances
lus \etrrosci Lett 5l:tq 3a
Conzalez-Lnna F, 5th!,.h r1 (198i) AscendinE relicular activatnrg
systemin th. iit \:ieo\vSlucose strdy. Brain Res 334170-88
Conzalez'Liina F Srh.,.h H (r986a): Neural substrates for rone_
.ondition.d br.d\.. ia demon*r.t€d with 2 deotyglucose Il
Auditur\ ..rF\ fltrlr.rtl Beha! Brain Res20:281 ?e3
Gonzalcr Lini F Sihei.h H (1936b)i Classical conditioning ol
tone{ignaleJ brrl\t:rJ1a modines 2 dcovglucose uPtake Pat_
ternsin.lft!\ if rltrrrui,habenula,cardate_PutanenandhiPPo
campal Iorn:r.n 3-,n R.s 363i239-256.
 . Path AnalJsis ofAuditory L€a.ninC.
Conzal€z-Lima I, r'inkenstaedt T, Ewdt Jl (1989a): L€arninE
related activation in the auditory syst€n ol the rat Produced by
longrern habituation: A 2-deoxyglucose rtudy Brarn RPs 139,67
Conzalez-Lina F, Fink€nstaedt T, Esert JP {1989b): Neural sub
strates lor l.ng tern habituation of the acoustic startle refle\ nl
rats A2 deoxySlucos€ study Ncur.sciL€tt96:151 156.
Gonzalcz Lima F, FnrkensraedtT, Scheich H (eds) (1992): Advances
in Metrbolic MappinS Techniques for Branr ImaginE of Behav'
ioral and LearninS Functions NATO ASI seiies, Vol D68
Dordrecht: Kl!wei Acadcni. lrubl'sh€!s
conzalez Lima F, Helnstetter FJ, A8r.1o J {1993): Iunctional maP
pinE ot th€ ratbrain durinE diinkingbehavior: A fl uor.d.oryElt!
cose study. l'hysiol Behav 54:605 612
Craybicl AM (r9rc) The thalanocorlical proiections of th. so call€d
poslerior nu.lear SrouP: A study sith nnr€rograd€ d€generahon
nethods the.at. Brain Res.19 229 244
Hayduk LA (1987): Skuctural Equation Modellng with LISREI-,
Essentiak and Advances. Baltinore: John HoPkins University
H€mandcz Peon R (1960): Neurophysiologi.al.orrelat.s or habitu_
ation and other nanifestations of Plastic inhibition. EEC Clin
Neurophysiol, Suppl l3:101 r1,1.
HoNitz R (19u9): Functional nelral systens analyzed by usc of
int€rre8ional .orrclations oI glucos€ metabolhm. ln: Ewerl l-P,
Arbib MA (edt VisuoDrotor Coordination: AnPhibians, Con
painons, Models and Roboh. New Yorl: l'lenun Press, PP
873-892.
HorwitzB,DuaraR,RapoportSI(198'l) lnrercorrelationsol glucosc
metaboli. iat.s beiween brain r.Sions APPlic.tion to h€althy
males in a state of reduced s€nsory inPut. J Cereb Blood FlolL
Horwitz B, Soncrant TT, Haxby iV (1992): Covarian.. analysis ot
functional int€ractions in the brain using metaboljc rnJ blood
flow data.ln: Conzalez-Lina I,Iinkenrtn.drT, Schei.h H (eds):
Advances in Metabolic MaPPinS Techniques lor Brain lmaging
ol Behavidal nnd L€arning Functions NATO ASI series, Vol
D68. Dordrecht: Kluwer Academic Publishe6, PP 189 218.
J6reskog K, sdrbon D (1979): Advanccs in lactor Analvsn and
Structural EquationModels CtmbridEe,MA: AbtBooks
Jijr.skog K Sdrbon D (1984): LISREL VI, Analysis or Linear
Stru.tural Relationships by M.ximun Likelihood, lnstruncntal
Variables, and Le.st-Squares Methods Moorcsville, lN: Scien
Joreskog K, Sijrbon D (1989): LISREL 7 l-sers Referenc€ Cuide
Mooresviue, IN 5cienti6c Soft$are lnc.
kndel ER, Schwartz JH (1982): Molecular biolooi leainnrg:
Modulat'on oI transnitrei release Science218:133 113
Leaton RN, Casseua JV, Rorsz.2 GS (198s): Shorl-lerm and long
term habituatnrn of the acousti. startle resPonle in ihroni.
decerebrat. rats. BeharN€urosci 99:901 912.
Loehlh JC i1987) Latent Vaiable Modch, An l.tro.lu.h.. to
Factoi, Path and Structurrl Analysis Hilhdal€, NJ: LalLfen.e
Erlbaun Associate5.
Lolordo VM, Ianl€ss JL (198s): Pavlovian condiiioned inhitnron
The lilerature sinc.1969.In: Miller RR, Sp€arNE (€ds): inn{nt
tion Proc€ssing in Animals: Conditioned Inhibition. Hrllld.le
NJ: Lawrenc. Eilbaum Asso.iates, pp 1 19
Long jS (1983): Covaiiance Structural Mod€ls, An lntrcdu.tior nl
LISREL SaEe University Paper Serics on Qunntitati!. AfFl,.r-
tions in th€ SocialSci.nces. BeverlyHills Sage lublication!
Lorente de No R {19{3.1): The Prnnary Acoustic Nuclei. NeLf \rtr:
McDonJld RP (1980): A iimPle coDrPrehcnsive Drodel forth€ analy_
sis otcovariancc srrnchtres. Drl Math Stal Psychol3l:s9-72
Mclntosh Al{, Gonzalez l-ima F 11991): Skuctural Drcdcling or
iunctional ncuial pathNays naPPed with 2 deoxyglu.ose: Ef
fects of acouslicstaltle habituation on the atrdilory systen. Biain
Rcss47:295-302
M.lntosh AR, Gonzalcz Lima | (1992.): Structural modelin8 ol
functional visual pathways mrPP€d with 2 deoryglucose: Ef
lects ol patterncd li8htand footsh.ck Brain Res573 75 8r,
Mclntosh AR, Conzalez-Lina F (r992b) The aPPlicat'on ol struc
tur,l modeling to metabolic maPPinS oi lunctional nerral
slncms lni Conzalez Lima F, Fink€nsta.dt T, Scheich H (eds):
Advances in Metabolic Mapping Techniqu€s ior Brain lnaging
of B€hJvioral and Lenrning lunctions NATO ASI senes, Vol
D68. Dordre.ht: Klls ei Acadenic Plblishes, PP 219 255
Mclntosh AR, Conzalez'Lina F (1993): Neh{ork analysis ol func'
honal auditory Path$ays maPPed with nuotud.oty8hcor
Asso.iativc cuects ola tone conditioned as a Pavlovran excrior or
inhibitof. Brain R€s 627129 140
Mclntosh AR, Gonzalez Lina f (1994): Netsorkint€ractionsanong
linbi. cortices, basil for€brain and cer€b€llun dilf€rentiJte a
ton. conditioned as a Pt!l.vian excitor oi inhibitor: Fluorod€o\y
glu.ose mapFing nnd covariance sttu.tural nrodehrg. J Ncuro
physiolT2:1717 1733
Mclntosh AR, Crady CL, Una€rleidf LG, Haxby JV, RaPoPort SI,
Horwitz B (1994): Network anrlysis ol.oii.al vislal Pathways
napped with Pl1.JNeurosci l'1,rt56 666
Metter EJ, Ricge WH, Kuhl DE, FhelPs ME (1984): Cerelrral meta
bolic r€lationshiPs lor s€lected brain regrons in hetlthy adult J
Cer€b Rlood llow M€tab 4:1 7.
Miller RR, SpearNE (1985): lnformation Pro.essin8 inAfimals: clon
ditioned lnhlbilion. Hillsdale, NJ: Lawrence Erlbaum A$o.i
Olds l, Disterholt Jr,se€al M, Komblith CL, Hirsh R (1972)l
Learnnrgcent.rs ot rat branr naPPcd by measunng the ht.ncies
ofconditioned unit resPonscs.I NeuroPhysnn 35,212 2r9
Os€n KK (1987): AnatoDry of th. ma'nmalian cochlear nucleii A
review.ln, SykaJ, Masterton B (eds): Auditory Parhway: Struc_
ture and Iunction. New York: Plenum Press, PP 65 75.
Parham K, willoitJF (1990): Eflects of interior .olli.ul!s lesions on
th. rcorstic startle r.sPonse. Behav Neurosci r04:331-840.
P€dhazur EJ (1982): MultiPlc Re8ression in Rchavioral R€s€ar.h
Erplanation and Pr.di.tion 2nd ed. N.{ Yorkr Holl, Reinhai
Poline J'8, Mazoyei BM (1993): Analysis ol individual Positron
enis sion tonograPhy activation maPs by detectbn of hi8h
rrrio pixel clusters. I Cei€b Blood Flos Metab
signal-to-noise
73:425437
Rescorla RA (1988): Behavioral studies of l'avlovian .onditioning
Annu Rc! N€urosci 11329 352
Rolrnd PE, Levin B, Knsashnna R, Akeman S (1993)r Three_
dinicnsional analysis ol .lustered lox€ls in ':O Bltanol brain
actnalion inages. Hum Biain MaPPing l 3 r9.
shnrp r'R, Sagar SM (1992): MaPPlng sensoiNrtor Pathways in rat
biain 6ing 2 deoxyglucose auloradiograPhy nnd crosxnmuno'
.lto.hemistrv.ln: Conzalez'Linra F, Finkenstaedt T, Scheich H
leds): Advancer in Metaboli. Mapping Tcchniqles lor Branr
Im aginS ot B€havionl and L.ninin8 Functions. NATO ASI
i€ries, Vol D.;8. Dordrecht: Kluw.r A.ademic lublishe$, PP
j9:55
:h.h:nr D, Ullman S, Grinlald A (1991): Characterizatlon ot
.]\nanri. patr.rns of cortical activitJr by a small number of
:ri..ipalconrFon€nts AbskSocNeuroscilTrl0S9
 . Gonzalez-Uma and liclnaosh
Shore 5E,Hillert RH, Bledsoe 5C, Altshuler R& Codfrey DA (1991):
Descendingprojections to the dorsal and venr.al divisions of rhe
cochlear nucteus in theguinea pig HeannE Rer 52:255-268.
Soncrant TT, Horwitz B, Holloway HW Rapoport 5t (1986): fte
Patten ol tunciional coupling ol brain reEions in the awake rat.
War WB (192): Organization of olivocochiear elferenr iy5uns ,n
mammals.In: websi€r Dq Popper AN, Iay RR (edr: The
Mamalian Auditory lathway: Neurcanatomy. Bertinr Spnnger,
VerlaS, pp 410--448.
.
Webster WR (1985): Auditory sysrem. ln: Paxinos C (ed): The Rat
Neaous Sysrem, Vol 2. Sydney: Academic Press, pp 153-184.
WeinberE R.J, Rustioni A (1984: A cuneocGhlear pathway in the rat.
Neuroscience 20:209-219.
Weinberger NM, Diahond DM (1984: PhysioloEical plasricity in
au.litory cortex: Rapid induction by learning. Prc6 Neurcbiol
29:1-55.
WeinberSer NM, Ashe Jri, Methe.ate R, McKenna TM, Damond
DM BdUn J rlaao,: Relunrn8 audnory ,orr€\ b) teami^8: A
Prelimindrv modelof receptive 6"ld p dstrcrly Con.epts \FUro
scil:91-132.