Professional Documents
Culture Documents
The attached
copy is furnished to the author for internal non-commercial research
and education use, including for instruction at the authors institution
and sharing with colleagues.
Other uses, including reproduction and distribution, or selling or
licensing copies, or posting to personal, institutional or third party
websites are prohibited.
In most cases authors are permitted to post their version of the
article (e.g. in Word or Tex form) to their personal website or
institutional repository. Authors requiring further information
regarding Elsevier’s archiving and manuscript policies are
encouraged to visit:
http://www.elsevier.com/copyright
Author's personal copy
Short report
Abstract
In a computer simulation, a neural network first received a simultaneous procedure, where the interstimulus interval (ISI) was 0 time-steps
(ts). Output activations were near zero under this procedure. The network then received a forward-delay procedure where the ISI was 8 ts.
Output activations increased to the near-maximum level faster than those of a control network that first received an explicitly unpaired procedure.
Comparable results were obtained with rats that first received trials where a retractable lever was presented for 3 s concurrently with access to
water. Low-lever pressing was observed under this procedure. The rats then received trials where the lever was followed 15 s after by water. Lever
pressing appeared faster than a control group that received the 15-s ISI after an explicitly unpaired procedure. The model used in the simulation
explains these results as connection–weight increments that promote little output activations in a simultaneous procedure, but facilitate acquisition
in an optimal ISI.
© 2008 Elsevier B.V. All rights reserved.
Keywords: Simultaneous conditioning; Neural networks; Rats; ISI; Learning; Performance; Pavlovian conditioning
0376-6357/$ – see front matter © 2008 Elsevier B.V. All rights reserved.
doi:10.1016/j.beproc.2008.02.018
Author's personal copy
(S-FD) received in Phase 1 a protocol that simulated a simul- the same forward-delay procedure that was given to S-FD in
taneous procedure (S). The network received 200 trials where Phase 2. All free parameters of the activation and learning func-
the CS and US elements were concurrently and maximally acti- tions were the same as in previous simulations (see Appendix
vated (activation level of 1.0) for the last 3 ts of a 60-ts cycle. A).
During the entire cycle, the CTX unit was maximally activated Fig. 2 shows the results. The upper left panel depicts the
to simulate a context, for a fixed intertrial interval (ITI) of 57 ts. CR/UR activation of S-FD at ts 59 for the 10 test trials after
S-FD was then given 10 CS-alone test trials where the learning Phase 1 (the simultaneous procedure). It can be observed that the
function was disabled, in order to preserve the final connection CR/UR activation was close to zero, which represents a severe
weights from Phase 1 and assess the effect of the procedure with- performance deficit. The upper right panel depicts the CR/UR
out any learning. In Phase 2, S-FD received a training protocol activation of S-FD during Phase 2 (the forward-delay procedure
that simulated a forward-delay procedure (FD) with an optimal with an optimal ISI). It can be observed that the number of rein-
ISI. The network was given 200 CS–US pairings where the CS forced trials before the CR/UR activation reached the maximum
element was maximally activated for the last 9 ts and the US level was well below 25. This number contrasts with the 150
element was maximally activated at the last ts of the 60-ts cycle, trials that took the CR/UR element of EU-FD to reach the maxi-
for an interstimulus interval (ISI) of 8 ts. The context was again mum activation in Phase 2, as shown in the right lower panel of
simulated by activating CTX maximally throughout the entire the figure. Initial exposure to the simultaneous procedure thus
cycle of 60 ts, for an ITI of 51 ts. facilitated CR acquisition under an optimal ISI.
Another network with the same architecture (EU-FD) The results are consistent with those reported in experiments
received in Phase 1 a training protocol that simulated the sin- with animals where second-order conditioning tests have been
gle alternation variant of the explicitly unpaired procedure used (e.g., Barnet et al., 1991; Matzel et al., 1988). However, our
(EU). The US element was activated once in the middle of simulation used an acquisition test, so it is procedurally more
the 60-ts cycle (at ts 30). The CS and US were thus sepa- comparable to animal studies like those reported by Ross and
rated by a fixed interval of 29 ts. In Phase 2, EU-FD received Scavio (1983), and Salafia et al. (1980). However, neither study
Fig. 2. CR/UR activations at ts 59 for networks S-FD (upper panels) and EU-FD (lower panel). The upper left panel shows the activations during the test trials given
to S-FD immediately after Phase 1 (simultaneous procedure) and before Phase 2 (forward-delay procedure). The lower left panel shows the activations during the
test trials given to EU-FD immediately after Phase 1 (explicitly unpaired procedure) and before Phase 2 (forward-delay procedure). The right panels show the output
activations during Phase 2.
Author's personal copy
on the last two bars of the figure showed that the difference was
significant (t(8) = −12.78, p < .0001).
3. Animal experiment
Fig. 3. Mean initial (leftmost bar) and final weights for CS–sa and sa–ma con- 3.1. Method
nections of S-FD and EU-FD on at the end of Phases 1 and 2. The phase numbers
are indicated in parentheses. CS–sa: connections from the CS input to the sa
element. sa–ma: connections from the sa to the ma elements (see Fig. 1). The 3.1.1. Subjects
error bars represent standard errors. Sixteen Wistar female rats (Rattus norvegicus) of 3 months
of age with no prior history of responding on any experimen-
reported a positive transfer from a short to an optimal ISI. This tal procedure served as subjects. Between sessions, rats were
result is in disagreement not only with the present results, but individually housed with free access to food in a temperature-
also with the evidence from experiments where second-order controlled colony under 12:12 h light/dark cycle. They were
conditioning tests have been used. One possible explanation of maintained at regime of pre-session water deprivation of 23.5 h
this discrepancy is that the preparations used in the experiments and 30 min of post-session access to water.
are different. Second-order conditioning tests have been typi-
cally used with conditioned suppression in rats. Ross and Scavio, 3.1.2. Apparatus
as well as Salafia et al., in contrast, used the preparation of the Experimental sessions were conducted in five identical MED
nictitating-membrane response (NMR) in rabbits. It remains to Associates modular test chambers (305 mm long, 241 mm wide,
be seen whether this discrepancy remains with a second-order and 210 mm high), each enclosed in a sound- and light-
conditioning test using the NMR preparation. attenuating box equipped with a ventilating fan. The front, rear
The observed positive transfer in the networks was due to walls, and ceiling of the chambers were made of clear plastic.
an increment in the CS-connection weights under the simulta- The front wall was hinged and served as a door to the cham-
neous procedure for S-FD. Fig. 3 shows the mean initial and ber. The two side panels were made of aluminum, and the floor
final weights for certain connections in Phases 1 and 2 for both consisted of a stainless steel grid floor positioned above a stain-
networks. At the end of the simultaneous procedure, S-FD’s CS- less steel waste pan. One of the panels featured two retractable
sa weights increased to near-maximum levels (second bar from response levers flanking a liquid dipper. Only the lever to the
left to right). This increment is in sharp contrast to the near-zero right of the dipper was used as a CS and operandum. The other
CS-sa weights observed at the end of the explicitly unpaired lever remained retracted throughout the entire experiment. The
procedure for EU-FD (middle bar). The procedure thus induced lever protruded 19 mm and was located at 70 mm above the grid
a weight loss in this connection, due to the presentation of the floor. The lever had a time cycle of 700 ms and required a mini-
CS alone (the same mechanism accounts for latent inhibition in mal tension of 25 grams to be activated. The dipper was centered
this model, as reported in Burgos, 2003). Exposure to the simul- horizontally on the panel and placed at 25 mm above the grid
taneous procedure thus gave S-FD a sensory learning advantage floor. The dipper featured a receptacle opening (51 mm wide,
over EU-FD. 51 mm high) through which a motor driven dipper arm could be
Such advantage, however, was insufficient for S-FD to raised to deliver 0.01 cm3 of water. A houselight was mounted
respond under the simultaneous procedure, because the pro- 12 mm from the ceiling on the sidewall opposite the intelligence
cedure did not promote a sufficient increment in the sa–ma panel, and was on during the entire experiment. The ventilation
connection weights. There was some increment in these weights fan mounted on the rear wall of the sound-attenuating cham-
at the end of the simultaneous procedure for S-FD (fourth bar; ber provided masking noise of 60 dB. Experimental events were
a comparable increment was observed in EU-FD at the end of arranged using a Med interface connected to a PC controlled by
the explicitly unpaired procedure). However, this increment was Med-PC IV software.
not sufficient for S-FD to respond substantially under the simul-
taneous procedure. A more substantial responding required the 3.1.3. Procedure
increment observed in these weights at the end of the forward- The rats were randomly assigned to two groups (S-FD and
delay procedure (Phase 2) for S-FD (fifth bar). A paired t-test EU-FD) of eight subjects each, and were directly exposed to the
Author's personal copy
experimental conditions, without any adaptation to the chamber, sion. The left panels show the individual data for each group in
or pretraining with the liquid dispenser or the lever. The groups Phase 1. For S-FD (upper left panel), the percentage tended to be
were exposed to procedures that were analogous to those used below 25% throughout the phase, with the exception of rats R18
in the simulation; hence the same labels used for the networks (31.67%) and R17 (35%) in Sessions 1 and 2, respectively. Sim-
in the simulations were used here for the animal groups. ilar results can be seen for EU-FD, under the explicitly unpaired
In Phase 1, S-FD was given seven sessions of 60 trials where procedure (lower left panel).
the lever protruded into the chamber for 3 s (CS), during which The right panels depict the median percentages for each
there was a 3-s access to the cup of the liquid dispenser (US). group and session in Phase 2 (solid thick lines with large closed
This procedure was strictly simultaneous (S) in that the CS and circles). The light dashed lines with small various symbols
US occurred concurrently. In Phase 2, S-FD was shifted to a represent individual data. The median percentage increased to
forward-delay (FD) procedure with an ISI of 15 s for six more 69.17 for S-FD in the first session under the forward-delay
sessions of 60 trails. Group EU-FD, in contrast, was first exposed procedure with an ISI of 15 s (upper right panel). For EU-
to seven sessions of 60 trials of the single alternation variant of FD, the median percentage in the same session and procedure
the explicitly unpaired (EU) procedure, where the CS and US was 5.83. A Mann–Whitney test revealed a significant differ-
occurred alternately, separated by a fixed interval of 30 s. In ence between the two medians (U = 12.5, z = −2.06, p < .05,
Phase 2, EU-FD received the forward-delay procedure for six two-tailed). The differences in the remaining sessions were not
more sessions of 60 trials each. The ITI was held constant at 60 s significant.
for both groups in both phases. Lever presses had no scheduled The number of trials to acquisition was also measured. Acqui-
consequences. sition was defined as the occurrence of at least one response in
three out of four consecutive trials (Gibbon and Balsam, 1981).
3.1.4. Results The median numbers of trials to acquisition were 10.5 for S-FD
The results are shown in Fig. 4, which depicts the median and 75.5 for EU-FD, and the difference was significant (U = 13.0,
percentage of trials with at least one lever press for each ses- z = −2.0, p < .05, two-tailed).
Fig. 4. Percentage of trials with at least one response, for each session and phase of the experiment. (Left panels) Phase 1, where Group S-FD received the simultaneous
procedure (S, upper left panel) and Group EU-FD received the explicitly unpaired procedure (EU, lower left panel). The lines represent individual data. (Right panels)
Phase 2, where both groups received the forward-delay procedure (FD) with an ISI of 15 s. The solid lines with large closed circles represent the percentage medians
per group per session. The light dashed lines with various small symbols represent individual data.
Author's personal copy
− +
Fig. A1. Generic neural processing element. The element consists of a number of inputs that are connected and send excitatory (ai,t ) or inhibitory (ai,t ) activations
to a summing junction j. Each connection has a strength represented by a weight (w− i,j,t for excitatory connections, w+
i,j,t for inhibitory connections). The junction
computes inner products between an activation vector (a) and a weight vector (w) separately for excitatory and inhibitory inputs. Each product (excj,t for excitatory
elements, inhj,t for inhibitory elements) is passed as an argument to a separate logistic function L. The results of both logistic functions are used to compute the
element’s activation state, according to the activation function (see Appendix A). Subindex t represents a time-step (ts).
Author's personal copy
activations is not made in terms of positive versus negative acti- remains to be done. However, it seems reasonable to expect
vations. Rather, it is made in terms of different types of units that such manipulations will make important differences in the
that have differential effects on the activation of those units to model’s capability to simulate certain phenomena.
which they connect:
The activation a of a hidden or output unit j at t is given by
⎧
⎨ L(excj,t ) + τj L(excj,t−1 )[1 − L(excj,t )] − L(inhj,t ) if L(excj,t ) > L(inhj,t ) and L(excj,t ) ≥ θj
⎪
aj,t = aj,t−1 − κj aj,t−1 (1 − aj,t−1 ) − L(inhj,t ) if L(excj,t ) > L(inhj,t ) and L(excj,t ) < θj
⎪
⎩
0 if L(excj,t ) ≤ L(inhj,t )
where
1
s References
L(x) = ; x= ai,t wi,j,t
1+e (−(x−μ)/σ)
i=1 Baker, A.G., 1977. Conditioned inhibition arising from a between-sessions neg-
ative correlation. J. Exp. Psychol.: Anim. Behav. Process. 3, 144–155.
θ j is a random threshold generated according to a Gaussian dis- Barnet, R.C., Arnold, H.M., Miller, R.R., 1991. Simultaneous conditioning
tribution with a mean of 0.2 and a standard deviation of 0.15; demonstrated in second-order conditioning: evidence for similar associa-
τ j = 0.1 is a temporal summation parameter, κj = 0.1 is a decay tive structure in forward and simultaneous conditioning. Learn. Motiv. 22,
parameter; μ = 0.5 and σ = 0.1 are the mean and standard devi- 253–268.
ation (temperature) of the logistic function L. All activations Bevins, R.A., Ayres, J.J.B., 1995. One-trial context fear conditioning as a func-
tion of the interstimulus interval. Anim. Learn. Behav. 23, 400–410.
range from 0.0 to 1.0. Input units are not activated by this func- Burgos, J., 1997. Evolving artificial neural networks in Pavlovian environments.
tion, but their activations are manually assigned, according to In: Donahoe, J.W., Packard-Dorsel, V. (Eds.), Neural Network Models of
some training protocol. In the case of vta and CR/UR units, the Cognition. Elsevier, Amsterdam, pp. 58–79.
above function is used only in the absence of a US or primary Burgos, J.E., 2003. Theoretical note: simulating latent inhibition with selection
reinforcer (whenever the activation of I4 is 0.0). Otherwise, the neural networks. Behav. Process. 62, 183–192.
Burgos, J.E., 2005. Theoretical note: the C/T ratio in artificial neural networks.
unconditional connections (thick arrows in Fig. 1) are assumed Behav. Process. 69, 249–256.
to take precedence, in which case the activations are equal to the Burgos, J.E., 2007. Autoshaping and automaintenance: a neural-network
US magnitude (the level of activation of I4 ). approach. J. Exp. Anal. Behav. 88, 115–130.
Learning is defined as a change in one or more connection Burgos, J.E., Donahoe, J.W., 2000. Structure and function in selectionism: impli-
weights, according to the following learning function: cations for complex behavior. In: Leslie, J., Blackman, D. (Eds.), Issues in
Experimental and Applied Analyses of Human Behavior. Context Press,
αj aj,t dt pi,t rj,t if dt ≥ 0.001 Reno, pp. 39–57.
wi,j,t = Burgos, J.E., Murillo-Rodrı́guez, E., 2007. Neural-network simulations of two
−βj wi,j,t−1 ai,t aj,t otherwise context-dependence phenomena. Behav. Process. 75, 242–249.
Donahoe, J.W., Burgos, J.E., Palmer, D.C., 1993. A selectionist approach to
where αj = 0.5; βj = 0.1 are free parameters; ai,t denotes the reinforcement. J. Exp. Anal. Behav. 60, 17–40.
presynaptic activation; aj,t denotes the postsynaptic activa- Donahoe, J.W., Burgos, J.E., 1999. Timing without a timer. J. Exp. Anal. Behav.
tion; dt = ds,t = ϕt + υt (1 − υt ), if j is an sa or ca1 unit; 71, 257–263.
Donahoe, J.W., Burgos, J.E., 2000. Behavior analysis and revaluation. J. Exp.
dt = dm,t = υt , j is an ma, vta, or output unit; υt = Mean(vt − vt−1 );
Anal. Behav. 74, 331–346.
a w
s
Donahoe, J.W., Palmer, D.C., 1994. Learning and Complex Behavior. Allyn &
ϕt = |Mean(ht − ht−1 )|; pi,t = i,t Ni,j,t−1 ; rj,t = 1 − wi,j,t ;
Bacon, Boston.
i=1 Donahoe, J.W., Palmer, D.C., Burgos, J.E., 1997a. The S-R issue: Clarification of
N = excj,t or N = inhj,t . its status in Donahoe and Palmer’s (1994), Learning and Complex Behavior.
All weights are between 0.0 and 1.0. The term dt desig- J. Exp. Anal. Behav. 67, 193–211.
nates a discrepancy signal that determines whether weights are Donahoe, J.W., Palmer, D.C., Burgos, J.E., 1997b. The units of selection: what
increased or decreased (0.001 is an arbitrary criterion for mak- is reinforced? J. Exp. Anal. Behav. 67, 259–273.
Donahoe, J.W., Vegas, R., 2004. Pavlovian conditioning: the CS–UR relation.
ing this decision), and if increased, by how much. The signal ds,t
J. Exp. Psychol.: Anim. Behav. Process. 30, 17–33.
influences weight changes in the input-sa and sa–ca1 connec- Droungas, A., LoLordo, V.M., 1995. The explicitly unpaired procedure yields
tions (see Fig. 1). The signal dm,t influences weight changes in conditioned inhibition whether the CS and the US alternate singly or ran-
the sa–ma, ma–vta, and ma-output connections. h designates a domly. Learn. Motiv. 26, 278–299.
vector of ca1 activations and v a vector of vta activations. The Friedman, B.X., Blaisdell, A.P., Escobar, M., Miller, R.R., 1998. Comparator
mechanisms and conditioned inhibition: conditioned stimulus preexposure
architecture used in the present study had one of instance of each
disrupts Pavlovian conditioned inhibition but not explicitly unpaired inhibi-
type of unit, so each vector consisted of one value. Factors p and tion. J. Exp. Psychol.: Anim. Behav. Process. 24, 453–466.
r make the function a competitive one, meaning that connections Gallistel, C.R., Gibbon, J., 2000. Time, rate and conditioning. Psychol. Rev.
impinging on NPE compete for a limited amount of weight (1.0). 107, 289–344.
The function also implements a mechanism in which the larger Gibbon, J., Baldock, M.D., Locurto, C., Gold, L., Terrace, H.S., 1977. Trial and
intertrial durations in autoshaping. J. Exp. Psychol.: Anim. Behav. Process.
the weight, the larger the change in that weight, everything else
3, 264–284.
being equal. Gibbon, J., Balsam, P., 1981. Spreading association in time. In: Locurto, C.M.,
Thus far, simulations with this model have not relied on para- Terrace, H.S., Gibbon, J. (Eds.), Autoshaping and Conditioning Theory.
metric manipulations, and a parametric analysis of the model Academic Press, New York, pp. 219–253.
Author's personal copy
Guthrie, E.R., 1935. The Psychology of Learning. Harper, New York. Ross, R.T., Scavio Jr., M.J., 1983. Perseveration of associative strength in rab-
Hawkins, R.D., Carew, T.J., Kandel, E.R., 1986. Effects of interstimulus interval bit nictitating membrane response conditioning following ISI shifts. Anim.
and contingency on classical conditioning of the Aplysia siphon withdrawal Learn. Behav. 11, 435–438.
reflex. Journal of Neuroscience 6, 1695–1701. Salafia, W.R., Host, K.C., Lambert, R.W., Chiaia, N.L., Ramirez, J.J., 1980. Rab-
Kleiman, M.C., Fowler, H., 1984. Variations in explicitly unpaired training are bit nictitating membrane conditioning: lower limit of the effect interstimulus
differentially effective in producing conditioned inhibition. Learn. Motiv. interval. Anim. Learn. Behav. 8, 85–91.
15, 127–155. Schneiderman, N., 1966. Interstimulus interval function of the nictitating mem-
Matzel, L.D., Held, E.E., Miller, R.R., 1988. Information and expression of brane response of the rabbit under delay versus trace conditioning. J. Comp.
simultaneous and backward associations: Implications for contiguity theory. Phys. Psychol. 62, 397–402.
Learn. Motiv. 19, 317–344. Schneiderman, N., Gormezano, I., 1964. Conditioning of the nictitating mem-
McAllister, W.R., 1953. Eyelid conditioning as a function of the CS–US interval. brane of the rabbit as a function of the CS–US interval. J. Comp. Phys.
J. Exp. Psychol. 45, 417–422. Psychol. 57, 188–195.
Miller, R.R., Matzel, L.D., 1988. The comparator hypothesis: A response rule Smith, M.C., Coleman, S.R., Gormezano, I., 1969. Classical conditioning of
for the expression of associations. In: Bower, G.H. (Ed.), The Psychology the rabbit’s nictitating membrane response at backward, simultaneous, and
of Learning and Motivation, vol. 22. Academic Press, San Diego, CA, pp. forward CS–US intervals. J. Comp. Phys. Psychol. 69, 226–231.
51–92. Stout, S.C., Miller, R.R., 2007. Sometimes-competing retrieval (SOCR): a
Miller, R.R., Schachtman, T.R., 1985. The several roles of context at the time formalization of the comparator hypothesis. Psychol. Rev. 114, 759–
of retrieval. In: Balsam, P.D., Tomie, A. (Eds.), Context and Learning. 783.
Lawrence Erlbaum, Hillsday, NJ, pp. 167–194. Sutton, R.S., Barto, A.G., 1981. Toward a modern theory of adaptive networks:
Rescorla, R.A., 1980. Pavlovian Second-Order Conditioning: Studies in Asso- expectation and prediction. Psychol. Rev. 88, 135–170.
ciative Learning. Lawrence Erlbaum, Hillsday, NJ. Wasserman, E.A., Molina, E.J., 1975. Explicitly unpaired key light and food pre-
Rescorla, R.A., Wagner, A.R., 1972. A theory of Pavlovian conditioning: varia- sentations: interference with subsequent auto-shaped key pecking in pigeons.
tions in the effectiveness of reinforcement and nonreinforcement. In: Black, J. Exp. Psychol.: Anim. Behav. Process. 1, 30–38.
A.H., Prokasy, W.F. (Eds.), Classical Conditioning II: Current Research and Yeo, A.G., 1974. The acquisition of conditioned suppression as a function of
Theory. Appleton-Century-Crofts, New York, pp. 64–99. interstimulus interval duration. Quar. J. Exp. Psychol. 26, 405–416.