burgos (2003) - latent inhibition
TRANSCRIPT
-
8/9/2019 Burgos (2003) - Latent Inhibition
1/10
Behavioural Processes 62 (2003) 183192
Theoretical note: simulating latent inhibitionwith selection neural networks
Jos E. Burgos
Centro de Estudios e Investigaciones en Comportamiento, University of Guadalajara, 12 de Diciembre 204,
Col. Chapalita, CP 45030-Guadalajara, Jalisco, Mexico
Received 1 September 2002; accepted 8 December 2002
Abstract
The selection neural-network model proposed by Donahoe et al. [J. Exp. Anal. Behav. 60 (1993) 17] was used to simulate
latent inhibition (LI). The model can simulate increases of LI by the number, intensity, and duration of preexposed conditioned
stimulus (CS). It can also simulate dependence on total CS preexposure time, CS specificity, and attenuation by preexposure
to a compound that includes the to-be-trained CS. It also predicts a potentially new phenomenon: acquisition facilitation by
preexposure to a stimulus that is orthogonal to and synaptically competitive with the to-be-trained CS. The basic mechanism
is the same through which the model simulates extinction, namely, weight decrement. The realization of this mechanism in
the present simulations required two conditions. First, networks had to come to the experimental situation with substantial
initial connection weights in the sensory-association subnetwork (0.15, compared to the 0.01 value we have used in all previous
simulations). Second, the discrepancy threshold for deciding whether to increase or decrease weights had to be larger than zero
(the value we have used in all published simulations). A value of 0.001 was sufficient to produce all the effects.
2003 Elsevier Science B.V. All rights reserved.
Keywords: Latent inhibition; Neural networks
1. Introduction
Latent inhibition (LI) is a significant retardation in
the acquisition of a conditioned response (CR) after a
preexposure to the putative conditioned stimulus (CS)
(Lubow and Moore, 1959).All accounts of this phe-nomenon have relied on the learningperformance dis-
tinction, where learning typically refers to a covert
process that causally affects performance, which de-
fined as overt behavior. Accordingly, Schmajuk et al.
(1996, Table 1) organize such accounts into two cat-
Present address: 413 Interamericana Blvd., WH1, PMB 30-189,
Laredo, TX 78045-7926, USA.
E-mail addresses: [email protected], [email protected]
(J.E. Burgos).
egories. In one category, LI is viewed as a learning
failure, consisting in a disruption of the formation of
associations between a CS and an unconditioned stim-
ulus (US), after preexposure. In the other category,
LI is viewed as a performance failure, consisting in
a disruption of the retrieval of such associations. Athird category arises from the neural-network model
advanced by the authors, where LI is viewed both as
a learning and a performance failure.
My aim in the present note is two-fold. First, I
want to demonstrate that the selection neural-network
model proposed byDonahoe et al. (1993)can simu-
late at least the basic LI phenomenon and some of its
behavioral properties. Second, I want to show how the
model accounts for those effects in terms of a kind
of learning failure that differs from other proposed
0376-6357/03/$ see front matter 2003 Elsevier Science B.V. All rights reserved.
doi:10.1016/S0376-6357(03)00025-1
-
8/9/2019 Burgos (2003) - Latent Inhibition
2/10
184 J.E. Burgos / Behavioural Processes 62 (2003) 183192
accounts. In the next section, I describe the features
that are relevant to the way in which the model simu-
lates LI. Then, I describe four simulation experiments
that show how the model simulates the basic LI effect,some of its behavioral properties, and even predicts
a potentially new phenomenon. I end with a general
discussion.
2. The selection neural-network (SNN) model
The SNN model is a discrete-time model consist-
ing of two sub-models, namely, neuro-computational
and network. The neurocomputational sub-model con-
sists of an activation rule, which determines the ac-
tivation state of a neural processing element (NPE)
at a time-step (ts) t, and a learning rule, which de-
scribes how connection weights change. Activations
and weights take values within the [0, 1] interval. The
network model is a taxonomy of NPEs plus rules on
how they are to be connected. A typical SNN architec-
ture, shown inFig. 1,is a feedforward (cf.Donahoe
and Burgos, 2000)fully-connected (cf.Burgos, 1997,
2000)network consisting of one input, two or more
hidden (labeled as sa and ma; see below), and one
output layer. Thin lines represent variable connections
Fig. 1. A typical DBP neural network. The small empty circles
represent CS input elements and the small filled circle represents
the US input element. Activations of input elements represent
stimuli. The large open circles represent neural processing ele-
ments (NPEs). The thin lines represent variable connections and
the thick arrows innate, fixed, maximally-strong connections. Ac-
tivations are transmitted unidirectionally, from input elements to
output NPEs; sa: sensory association; ca1: Cornu Ammon 1; ma:
motor association;vta: ventral tegmental area; CR/UR: conditioned
response/unconditioned response.
and thick arrows fixed maximally-strong connections.
All activations are conveyed from input to hidden to
output elements.
Larger-than-zero activations of the input elements(small circles labeled as CS1, CS2, CS3, US) rep-
resent stimuli (CSs, represented by the activation of
the small open circles, and USs, represented by the
activation of the small filled circle). Hidden layers
consist of at least one sensory-association (sa) and
one motor-association (ma) layer, and are constituted
by NPEs (large empty circles) that mediate between
input and output activations. Activations of the output
NPE represent either conditioned responses (CRs), if
caused by the activation of one or more CS elements
through the sa and ma NPEs, or unconditioned re-
sponses (URs), if caused by the activation of the US
input element (hence the label CR/UR). A selection
neural network can also have output NPEs that do
not receive connections from the US element (e.g.
Donahoe et al., 1993; Burgos and Donahoe, 2000;
Donahoe and Burgos, 1999), but they are relevant
for operant conditioning. Since my emphasis here is
on Pavlovian conditioning, I shall ignore that kind
of NPE. Other NPEs are the ca1 (for Cornu Am-
mon 1, a hippocampal field) and vta (for ventral
tegmental area, a dopaminergic nucleus) NPEs. Their
activations generate a discrepancy reinforcement sig-nal that modulates changes in connection weights
(see below).
The part that is relevant to the present study is the
learning rule (for a description of the activation rule,
seeDonahoe et al., 1997,Appendix), which is given
by:
wi,j,t=
jaj,tpi,trj,tdt, ifdt
jwi,j,t1ai,taj,t, ifdt < (1)
As in any neural-network model, connection weightsgate the efficacy with which presynaptic input signals
activate postsnaptic NPEs. The rule consists of two
mutually exclusive modes, namely, incremental and
decremental. In the incremental mode, the weightwi,j,tfrom presynaptic process i (which can be either
an input element or another NPE) to postsynaptic
process j at ts t is increased according to the first
conditional. The rate of increase depends on a free
parameter (),js activation (aj),pi (the proportion of
total weight allotted toi),rj(the remaining portion of
-
8/9/2019 Burgos (2003) - Latent Inhibition
3/10
J.E. Burgos / Behavioural Processes 62 (2003) 183192 185
weight left on j that is acquirable by any presynaptic
process), and the reinforcement signal (dt), where
pi,t=ai,twi,j,t1
N(2)
whereNdenotes the inner product of the activation and
the weight vectors that include iand impinge onj, and
rj,t= 1
si=1
wi,j,t (3)
wheres denotes the total number of connections that
impinge onj.Eqs. (2) and (3)make weight increment
a competitive process. ThroughEq. (2), more strongly
activated stronger connections gain weight faster than
less strongly activated weaker connections. ThroughEq. (3), the sum of all connection weights on j is
bounded to 1.0 and an increase in the weight of any
connection leaves less free weight available for other
connections to gain.
In the decremental mode,w i,jdecreases according
to the second conditional. This mode typically func-
tions whenever a CS is presented without reinforce-
ment, which is the critical feature for simulating LI
with this model. The rate of decrease depends on a
free parameter (), the amount of weight gained thus
far byi, and the activations ofi and j.
The two modes are not symmetric because themodel assumes that presynaptic competition and re-
inforcement discrepancy play a role only in the incre-
ment of synaptic efficacies. Whether the rule is in the
incremental or the decremental mode depends on dt,
which is given by
dt=
t+ t(1 t), ifjisan sa or a ca1NPE
totherwise
(4)
wheret= |aca1,t aca1,t1| (5)
t= avta,t avta,t1 (6)
whereaca 1is the activation of the ca1 NPE andavt ais
the activation of thevtaNPE.Eq. (4)assumes that the
reinforcement signal that modulates weight increases
in sa and ca1 NPEs arises from the latter and is am-
plified by the one arising from the vta NPE. This as-
sumption is based on the fact that the functioning of
the hippocampus is amplified by dopamine (e.g. Stein
et al., 1993).Eq. (5) assumes that the ca1 reinforce-
ment signal is always positive, based on the hypoth-
esis that the hippocampus is a change detector (i.e.it affects learning equally, whether the activity of its
constituting cells increases or decreases across time).
In contrast,Eq. (6)assumes that thevtareinforcement
signal can be positive or negative. This assumption
is based on the hypothesis that dopaminergic nuclei
affect learning differentially, depending on whether
the activity of their constituting cells increases or de-
creases across time.
The ca1 and vta reinforcement signals are diffuse
in that they reach all of their associated connection
weights (CS1-sa and sa-ca1, in the case of the ca1
signal, and sa-ma, ma-vta, and ma-CR/UR weights,
in the case of the vta signal), in a way analogous to
dopaminergic modulation of synaptic plasticity. Pa-
rameterinEq. (1)is a threshold for deciding whether
the learning rule is in the incremental or decremen-
tal mode. In all previous simulations, = 0, but this
had to be changed in order to obtain LI and its basic
behavioral properties (see below).
3. Simulation experiments
Tokens of the architecture shown in Fig. 1 were
used. The basic procedure consisted in preexpos-
ing some networks (the PE networks) to some LI
treatment, which involved non-reinforced CSi trials
(CSi). Other networks (the SIT networks) were
not pretreated in any way. Then, all networks were
trained in a forward-delay Pavlovian procedure con-
sisting of 300 reinforced CS1 trials (CS1+), where
CS1 was the activation of the CS1 input element
(see Fig. 1) with a magnitude of 1.0 for 8 ts, and
the US was the activation of the US input elementalso with a magnitude of 1.0 at ts 8. The interstim-
ulus interval during CS1+, thus, was 7 ts. As in all
previous simulations, the intertrial interval (ITI) was
not directly simulated. Instead, it was assumed to be
sufficiently long to make activations to decrease to
a spontaneous near-zero level, which was given by
a logistic function with a mean of 0.5, a standard
deviation of 0.1, and an argument of zero (the re-
sulting spontaneous activation is approximately equal
to 0.066928; for further discussion on the role of
-
8/9/2019 Burgos (2003) - Latent Inhibition
4/10
-
8/9/2019 Burgos (2003) - Latent Inhibition
5/10
J.E. Burgos / Behavioural Processes 62 (2003) 183192 187
Fig. 3. Weight changes in the connection from CS1 to the first
sa NPE from top to bottom in the architecture shown in Fig. 1.
Upper panel: a SIT net during CS1+. Lower panel: a PE net
during CS1+, after preexposure.
Fig. 3 shows changes in the weight of the connection
from CS1 to the first sa NPE from top to bottom in
the architecture depicted in Fig. 1 (that NPE is now
labeled as sa1), for a SIT (upper panel) and a PE
(lower panel) network, during CS1+. As can be seen,
the weight increased noticeably more slowly than the
corresponding weight in the SIT network. The weights
did not exceed 0.7 because the other two connections
(from CS2 and CS3; seeFig. 1) controlled 30% of thetotal weight available on that NPE.
Why did the CS1-saweights tend to increase more
slowly in the PE networks? CS1 preexposure caused
a weight loss in the initial CS1-sa connections.3 An
example is shown inFig. 4, which depicts changes in
the same CS1-sa weight of the same PE-200 during
CS1. As can be seen, the initial weight decreased
from 0.15 to almost 0.05. This decrease reduced the ac-
tivation ofj (aj ,t inEq. (1))and increased the remain-
ing portion of weight acquirable by i (rj ,t inEq. (1))
from 0.55 to approximately 0.65.
4
Parameter was
3 The sa-ca1 connections did not change significantly during
CS1, so they played no a role in the observed differences.
Subsequent simulations, not shown here, demonstrated that an
initial value of 0.15 in these connections only accelerated learning
in SIT and PE networks equally, thus leading to comparable results.4 Before preexposure, CS1, in virtue of its initial weights, con-
trolled 0.15 of the maximum weight available on any of the sa
NPEs, while CS2 and CS3 controlled 0.15 2 = 0.30, for a total
used weight of 0.45. This weight subtracted from the maximum
weight possible of 1 results in a total free weight of 0.55. After
the same for all networks, so it did not play a role
in the observed differences. The same applies to pi ,t,
for only one input element (CS1) was active through-
out the entire experiment. Consequently, there was nodynamic competition among the three input elements.
Further analyses showed that dtdid not differ signif-
icantly, so it did not play a role in the present results
either.
In order to assess the true level of activation of
the sa NPEs, in the absence of weight changes, fur-
ther simulations were run where the learning rule
was disabled (seeDonahoe et al., 1997,p. 207). The
results showed that, in effect, the mean activations
of the sa NPEs in the presence of CS1 tended to be
significantly lower in the PE networks after CS1
preexposure than in the SIT networks. For instance,
the firstsaNPE from top to bottom in the architecture
ofFig. 1, had a mean activation across all ts of 0.014
for a randomly chosen SIT network and 0.008 for a
randomly chosen PE network. Multiplying by r, one
obtains 0.014 0.55 = 0.007 for the SIT network
and 0.008 0.65 = 0.005 for the PE network. The
advantage provided by rin the PE networks was thus
overcompensated by the decrease in the sa activation
levels. This decrease, then, was sufficient to retard
weight increments in the CS1-sa connections in the
PE networks, thus causing them to take more trials toachieve acquisition.
3.2. Experiment 2: effect of preexposed-CS intensity
The experimental evidence has shown that LI in-
creases with preexposed CS intensity (e.g.Schnur and
Lubow, 1976,Experiment 2). To simulate this effect,
two new groups of 20 nave networks were used.
Group PE-0.2 received 200 CS1 of input element
CS1 of 0.2. Group PE-0.8 received 200 CS1 trials
with a CS1 intensity of 0.8. CS1 duration was 8 tsfor both groups. Then, both received 300 CS1+ trials.
AsFig. 5shows, in agreement to the evidence, LI was
directly proportional to the intensity of the preexposed
CS.
preexposure, the CS1-sa weights in the PE networks decreased to
almost 0.05 (seeFig. 3), which increased the free available weight
on any sa NPE to approximately 0.65. Consequently, after CS1
preexposure and before CS1+ training, the value of r for any sa
tended to be larger in the PE than in the SIT networks.
-
8/9/2019 Burgos (2003) - Latent Inhibition
6/10
188 J.E. Burgos / Behavioural Processes 62 (2003) 183192
Fig. 4. Weight changes in the same CS1-sa connection of the same PE net shown in Fig. 3, but during CS1 preexposure.
A two-tail unpaired t-test showed that the dif-
ference between the two group means was signif-
icant (t = 4.295, with 38 degrees of freedom and
P = 0.0001). The difference between the two stan-
dard deviations was not significant (F= 1.683, P=
0.1326) and both groups passed the normality test
according to the KS method [KS(PE-0.2) = 0.1325
and KS(PE-0.8) = 0.1322, with P > 0.1 for both
groups].The basic mechanism was the same one described
in Experiment 1. Preexposure to a more intense
Fig. 5. Results of Experiment 2. LI was increased by preexposed-CS intensity. Error bars represent standard deviations.
CS1 caused a more substantial decrease in the ini-
tial CS1-sa weights. Such a decrease was due to the
fact that weight decrements are as much activation-
dependent as weight increments, so the larger the ac-
tivation of the pre- and/or postsynaptic elements (ai ,tandaj ,tin the decremental mode ofEq. (1), respec-
tively), the more substantial the weight loss. This loss,
in turn, caused sa activations in the PE-0.8 networks
to be smaller, and hence, their CS1-sa weights to in-crease more slowly, thus requiring more CS1+ trials
to achieve acquisition.
-
8/9/2019 Burgos (2003) - Latent Inhibition
7/10
J.E. Burgos / Behavioural Processes 62 (2003) 183192 189
3.3. Experiment 3: dependence on total
preexposure time
The evidence has shown that LI is also increasedby the number of preexposed CSs (Lantz, 1973, Ex-
periment 1) and CS duration (Lubow, 1989, p. 63;
Westbrook et al., 1981, Experiment 3). Further re-
search indicated that the critical variable here is total
CS preexposure time, regardless of how it is dis-
tributed across trials (e.g.Ayres et al., 1992).Specif-
ically, LI increases with total preexposure and groups
with the same total exposure time, but different CS
durations, demonstrate comparable amounts of LI.
To simulate these effects, four new groups of
20 nave networks were used. Groups PE-100/5,
PE-50/10, PE-200/10, and PE-400/5 were preexposed
to 100 5-ts, 50 10-ts, 200 10-ts, and 400 5-ts CS1
trials, respectively. The total preexposure times were
500 ts for the first two groups and 2000 ts for the
last two. CS1 intensity was 1.0 for all groups. After
preexposure, all groups received 300 CS1+ trials. As
Fig. 6shows, in agreement with the evidence, LI was
directly proportional to total preexposure time.
A one-way four-level ANOVA showed significant
differences among the group means (F= 12.155, with
a total of 79 degrees of freedom and P
-
8/9/2019 Burgos (2003) - Latent Inhibition
8/10
190 J.E. Burgos / Behavioural Processes 62 (2003) 183192
To simulate these effects, two new groups of 20 nave
networks were used and compared to Groups SIT and
PE-200 (the latter now called PE-CS1, for ease of
comparison) from Experiment 1. Group PE-CS1/CS2was preexposed to 200 CS1/CS2 trials consisting
in the non-reinforced concurrent activation of input
elements CS1 and CS2. Group PE-CS2/CS3 was
preexposed to 200 non-reinforced CS2/CS3 trials
consisting in the non-reinforced simultaneous activa-
tion of input elements CS2 and CS3. All CSs during
preexposure had a duration of 8 ts and a magnitude
of 1.0. Then, both groups received 300 CS1+ trials.
As Fig. 7 shows, in agreement with the evi-
dence, preexposure to CS1/CS2 produced less LI
than preexposure to CS1 (compare PE-CS1 to
PE-CS1/CS2). CS specificity was also obtained (com-
Fig. 7. Results of Experiment 4. The effect of preexposure was
specific to the preexposed CS (compare PE-CS1 to PE-CS2/CS3),
LI to CS1 was attenuated by non-reinforced preexposure to the
CS1/CS2 compound (compare PE-CS1 to PE-CS1/CS2), and ac-
quisition under CS1+ was facilitated by non-reinforced preexpo-
sure to the CS2/CS3 compound (compare SIT to PE-CS2/CS3).
Error bars represent standard deviations.
pare PE-CS1 to PE-CS2/CS3). Furthermore, preexpo-
sure to CS2/CS3 substantially facilitated acquisition
(compare SIT to PE-CS2/CS3).
A one-way four-level ANOVA showed that themean differences were significant (F= 189.15, with
a total of 79 degrees of freedom and P < 0.0001).
A TukeyKramer multiple comparisons test showed
that, except for SIT and PE-CS1/CS2 (q = 3.08,P >
0.05), the differences among all the relevant means
were significant (for PE-CS1 and PE-CS1/CS2, q =
15.122, P < 0.001; for SIT and PE-CS2/CS3, q =
21.226, P < 0.001; for PE-CS1 and PE-CS2/CS3,
q = 33.268, P < 0.001; for PE-CS1/CS2 and
PE-CS2/CS3, q = 18.146, P < 0.001). The two
new groups passed the normality test, according
to the KS method [KS(PE-CS1/CS2) = 0.15 and
KS(PE-CS2/CS3) = 0.1728, with P > 0.1 for
both groups]. However, the differences between the
standard deviations of PE-CS1 and PE-CS1/CS2
(F= 5.196, P= 0.0004), PE-CS1 and PE-CS2/CS3
(F= 7.661, P
-
8/9/2019 Burgos (2003) - Latent Inhibition
9/10
J.E. Burgos / Behavioural Processes 62 (2003) 183192 191
a consequence of CS preexposure. Except for the
PE-CS2/CS3 networks in Experiment 4, CS1 pre-
exposure caused a reduction in CS1-sainitial weights
in all the PE networks, which, in turn, caused sa ac-tivations to be smaller early in CS1+ training, thus
decelerating CS1-saweight increments and retarding
acquisition. The amount of weight reduction during
CS1 (and, to that extent, of sa activations early in
CS1+) was directly proportional to the CS1 in-
tensity (Experiment 2) and total preexposure time
(Experiment 3).
In the first three experiments, the critical factor in
Eq. (1)was a reduction the postsynaptic (in this case,
thesa) activations. This reduction decelerated weight
increases in the CS1-sa weights during CS1+, and was
caused by a decrement in the initial CS1-sa weights,
due to CS1 preexposure. Under the SNN model,
then, LI arises as a learning failure. The way in which
this failure occurs, however, differs from the way it
does in other accounts, such as the one found in the
SchmajukLamGray (SLG) model, for instance.
Like the SNN model, decelerated weight increase
in the SLG model is due to a reduction in an ac-
tivation parameter (called novelty in the latter
model) that affects the rate of weight increments.
In contrast to the SLG model, the reduction of that
activation in the SNN model is due to a reductionin initial weights. Furthermore, in the SLG model,
the reduction in novelty takes place in a separate
subnetwork (the so-called novelty system), which
consists only of fixed connections. In the SNN model,
in contrast, the relevant reduction occurs in the same
subnetwork where subsequent learning occurs (viz.
thesa subnetwork).
Additionally, the minimal architecture necessary to
obtain the present effects is simpler in the SNN than
the SLG model (compare Fig. 1 of the present pa-
per with Fig. 1, p. 324, of the SLG paper). Also,the SLG model has 11 free parameters, whereas the
SNN model has only five (the two of the learning
rule plus three in the activation rule). The SNN ac-
count thus seems to be more parsimonious than the
SLG one, at least regarding minimal architecture and
number of free parameters. To be sure, there are many
other LI features that await simulation by the SNN
model (e.g. LI of conditioned inhibition; dependence
on ITI duration, context, and other compound-CS ma-
nipulations; etc.), so they may well require larger and
perhaps even non-fully connected selection networks.
This possibility raises the issue of the role of the archi-
tecture in the SNN model. This issue has been previ-
ously discussed in relation to handcrafted architectures(Burgos, 2000; Burgos and Donahoe, 2000; Donahoe
and Burgos, 2000)and implications for the evolution
of learning, behavior, and nervous systems, as ex-
plored through the use of genetic algorithms (Burgos,
1997).
The facilitation of acquisition that was observed in
the PE-CS2/CS3 networks in Experiment 4 is a po-
tentially new effect that, to the best of my knowledge,
is not predicted by any other conditioning model. It
represents a sort of negative LI, for which it can be
called latent disinhibition, or LD. I am not aware
of research with living organisms that is directly rel-
evant to LD, so the prediction awaits verification,
which raises the issue of how to design a conditioning
experiment to test for it. In the present simulations,
LD arose through preexposure to a stimulus that was
orthogonal to andsynaptically competitive with the
trained pattern. Orthogonality seems to be relatively
unproblematic to obtain in a conditioning experiment.
For example, a light and a tone presumably qualify
as orthogonal stimuli, in that they represent differ-
ent sensory modalities. Orthogonality, however, does
not ensure the absence of synaptic competition. TheSNN model asserts that there can be synaptic com-
petition among orthogonal input patterns and that it
occurs in sa cortex. However, it is unclear exactly of
which stimulus parameters synaptic competition is a
function. One possibility is frequency. For example,
a higher-pitch tone may be synaptically more com-
petitive with a shorter- than with a longer-wavelength
light.
Preliminary simulations demonstrated that if the
initial CS1-saweights were set to 0.01 and/or = 0,
none of the explored properties of LI was obtained.Substantial initial CS1-sa weights and > 0 thus
were both necessary to obtain the present results,
which has implications for the neural substrates of
learning. If one assumes that the present networks
were experimentally nave, the first condition implies
that experimentally nave need not mean tabula
rasa. Experimentally nave organisms may well come
to the experimental situation with substantial sen-
sory learning. Conditioning thus could involve larger
synaptic changes in motor than in sensory brain areas,
-
8/9/2019 Burgos (2003) - Latent Inhibition
10/10
192 J.E. Burgos / Behavioural Processes 62 (2003) 183192
at least in vertebrates (see Donahoe and Palmer,
1994, pp. 177210, for a discussion of this possibility
within a broader approach to learning and behavior).
The second condition suggests that the physiologicalthreshold for changes in dopaminergically-modulated
synaptic efficacies is larger than zero, although its pre-
cise cellular and molecular nature is unclear. Further
experimental and theoretical analyses are needed.
Acknowledgements
I thank John W. Donahoe, Peter R. Killeen, and an
anonymous reviewer for useful comments.
References
Ayres, J.J.B., Philbin, D., Cassidy, S., Bellino, L., Redling, E.,
1992. Some parameters of latent inhibition. Learn. Motiv. 23,
269287.
Burgos, J., 1997. Evolving artificial neural networks in Pavlovian
environments. In: Donahoe, J.W., Packard-Dorsel, V. (Eds.),
Neural Network Models of Cognition. Elsevier, Amsterdam,
pp. 5879.
Burgos, J.E., 2000. Superstition in artificial neural networks: a
case study for selectionist approaches to reinforcement. Mex.
J. Behav. Anal. 26, 159189.
Burgos, J.E., Donahoe, J.W., 2000. Structure and function inselectionism: implications for complex behavior. In: Leslie,
J., Blackman, D. (Eds.), Issues in Experimental and Applied
Analyses of Human Behavior. Context Press, Reno, pp. 3957.
Donahoe, J.W., Burgos, J.E., 1999. Timing without a timer. J.
Exp. Anal. Behav. 71, 257263.
Donahoe, J.W., Burgos, J.E., 2000. Behavior analysis and
revaluation. J. Exp. Anal. Behav. 74, 331346.
Donahoe, J.W., Palmer, D.C., 1994. Learning and Complex
Behavior. Allyn & Bacon, Boston.
Donahoe, J.W., Burgos, J.E., Palmer, D.C., 1993. A selectionistapproach to reinforcement. J. Exp. Anal. Behav. 60, 1740.
Donahoe, J.W., Palmer, D.C., Burgos, J.E., 1997. The S-R issue:
its status in behavior analysis and in Donahoe and Palmers
learning and complex behavior. J. Exp. Anal. Behav. 67, 193
211.
Lantz, A.E., 1973. Effect of number of trials, interstimulus
interval, and dishabituation on subsequent conditioning in a
CER paradigm. Anim. Learn. Behav. 1, 273277.
Lubow, R.E., 1989. Latent Inhibition and Conditioned Attention
Theory. Cambridge University Press.
Lubow, R.E., Moore, A.U., 1959. Latent inhibition: the effect
of non-reinforced preexposure to the conditional stimulus. J.
Comp. Phys. Psychol. 52, 415419.
Mackintosh, N.J., 1973. Stimulus selection: learning to ignore
stimuli that predict no change in reinforcement. In: Hinde, R.A.,
Hinde, J.S. (Eds.), Constraints on Learning. Academic Press,
London, pp. 7596.
Reiss, S., Wagner, A.R., 1972. CS habituation produces a latent
inhibition effect but no active conditioned inhibition. Learn.
Motiv. 3, 237245.
Schmajuk, N.A., Lam, Y.-W., Gray, J.A., 1996. Latent inhibition: a
neural network approach. J. Exp. Psychol. Anim. Behav. Proc.
22, 321349.
Schnur, P., Lubow, R.E., 1976. Latent inhibition: the effects of
ITI and CS intensity during preexposure. Learn. Motiv. 7, 540
550.
Stein, L., Bao, G.X., Belluzzi, J.D., 1993. A cellular analogue ofoperant conditioning. J. Exp. Anal. Behav. 60, 4153.
Westbrook, R.F., Bond, N.W., Feyer, A.-M., 1981. Short- and long-
term decrements in toxicosis-induced odor-aversion learning:
the role of duration of exposure to an odor. J. Exp. Psychol.
Anim. Behav. Proecss. 7, 362381.