Scalar Timing in Memory

J O H N GIBBON
New York State Psychiatric Institute
New York,New York 10032;and
Department of Psychology
Columbia University
New York.New York 10027

RUSSELL M. C H U R C H AND WARREN H. MECK

Department of Psychology
Brown University
Providence, Rhode Island 0291 2

INTRODUCTION

A recent report of ours’ proposed an information-processing account of temporal

generalization. The account posited a clock process, which was the basic time
measurement device, and working and reference memory for storing the output of the
clock either temporarily or relatively permanently. Records of time intervals in
working and reference memory were then compared using a binary decision process,
which dictated responding or not responding. The analysis concentrated on a relativis-
tic Weber’s law property of the data from temporal generalization, and the constraints
this property imposed on sources of variance in the information-processing stages. Our
purpose here is to summarize that work and generalize the model in two ways: First we
consider several sources of variance operating simultaneously. The original analysis
demonstrated that if only one source of variance is present, it must be a scalar source,
that is, it must result in a variable memory for which variance increases with the square
of the mean.’ In the generalized account proposed here, we will develop the conclusion
that scalar sources dominate in some time ranges, while other sources may dominate in
others. These ideas are then applied to two additional timing tasks with different
characteristics.

TEMPORAL GENERALIZATION

The reference experiment for temporal generalization is a straightforward task with

rats that was reported earlier.3 The procedure is simple: The houselight in an operant
chamber is turned off for a duration of time, T. Then a retractable lever is inserted into
the chamber. After a 5-sec opportunity to respond, the lever is withdrawn and a 30-sec
intertrial interval begins. If the light-off stimulus lasts for the “correct” duration, say,
4 sec, a response is reinforced with a pellet of food. If it is not 4 sec, no reinforcement is
forthcoming.
The typical result is that rats come to respond with high probability when the
duration is 4 sec, and with decreasing probability for stimuli either longer or shorter
than 4 sec.
FIGURE 1 shows how this decline changes with changes in the size of the reinforced
duration (S+). The probability of responding on the lever for three groups is shown as
a function of signal duration when the rewarded duration was either 2 sec (top panel),
52
GIBBON et al.: SCALAR TIMING IN MEMORY 53

4 sec (second panel), or 8 sec (third panel). In each case the maximum response
probability is near S+, and the spread of responsiveness around S + increases
considerably with increases in the S+ value.
The way in which the spread is related to the size of S + is shown in the bottom
panel, which plots these data on a relative time scale in which signal durations are
taken as proportions of the S + duration. The data from the different groups roughly
superpose in this metric. This property, which we have called the scalar property, is
ubiquitous in animal timing work in the seconds-to-minutes range.* The scalar
property we will see exerts strong constraints on admissible sources of variance and on
admissible comparison rules for the decision whether to respond or not.

.s

FIGURE
groups 1. Response
of rats studied with
probability
three different
gradients
placements
for three
of kE
U
m
w
0
.s

i--'
the positive signal. Darapoints represent median response
probability and are replotted in the bottom panel on a
relative time scale. (After Church and Gibbom2)
g .s
Q

p;
ki
v)
W
Q?

,.6 '

1 2 3
- T/S+ +
Our information-processing model for this situation is shown in FIGURE 2. The top
row shows the clock process, which includes a pacemaker and a switch for gating pulses
to an accumulator in working memory. The pacemaker generates pulses a t a mean rate
(A) that we assume is high relative to the time values (seconds to minutes) that we use
in these experiments. The switch, after appropriate training (instructions), gates pulses
for a mean duration ( D T )to an accumulator in working memory (second row) when the
timing signal is present. The accumulator records and stores the number of pulses
(mean of M T ) . When, at the end of a given trial, a response is made and reinforced
( T = S + ) , the time value recorded in working memory on that trial is stored in a more
54 ANNALS NEW YORK ACADEMY OF SCIENCES

permanent reference memory for reinforced values (mean of M$+). The third row
shows the decision process. A response occurs when a comparator yields a judgment
that the current record in working memory for this trial is “close enough” to the
reference memory for the reinforced duration to warrant a response.
In FIGURE3 we show the clock process in more detail. Imagine a pacemaker
generating pulses with interpulse intervals, 7, with mean rate A = A. In this figure, the
pulses are evenly spaced to indicate no variance. In later analyses we consider variance
in these parameters.“
The pulses are switched into the accumulator by the switch indicated in the middle
box. The switch (SW) is assumed to have some latency to close (tl) after the signal goes
on, and some latency to open ( t 2 )after the signal goes off. Thus, the mean time during
which pulses are gated into the accumulator is DT = T - To,where Tois the expected

CLOCK
PROCESS

FIGURE 2. Information-processing model.

The top row describes a clock process with a
MEMORY pacemaker generating pulses that are
PR0 CESS
switched into an accumulator for working
memory (middlerow). After reinforcement,
working memory contents are stored in a
reference memory for later comparison on
subsequent trials with the current working
DECISION memory value.
PROCESS

difference between the latencies to close and open the switch. In principle, Tomay be
negative as well as positive. The graph above the switch shows this linear relation
between DT and T. The intercept is the minimum signal duration below which counts
do not register in the accumulator. For signal durations less than this minimum, the
switch is opened before it has closed so no pulses are switched into the accumulator.
Conversely, if the latency to reopen the switch exceeds the initial latency to close it
(r2 > tl), even a very short signal duration suffices to allow counts to register during the
f2 latency. If t 1and t2 were precisely the same, the switch duration would mimic exactly
the signal duration. In general, this situation is unlikely, however, and these two
latencies constitute one of the sources of variance that we will consider later.
The accumulator in this scheme simply records the number of impulses gated to it.
The mean accumulated number, MT,therefore, is just the impulse rate times the

‘We generally refer to random variables with lowercase letters and their expectations with the
-
corresponding uppercase letter (for example, DT E(dT)).An exception is A, which can play a
dual role in the Poisson pacemaker.
GIBBON et al.: SCALAR TIMING IN MEMORY 55

A =ECh)

tl-fl-tz
I
'EW I< Dl
r -jt,i
I: 1

jtrj
,

- d T -L - Dr --+-

1 PACEMAKER kl SWITCH PI ACCUMULATOR1

FIGURE 3. Clock process. The pacemaker generates impulses at intervals of T and these
impulses are gated into an accumulator by the switch. The switch has a latency to close at the
beginning of the stimulus and to open at the end of the stimulus. The accumulator records the
number of impulses gated to it during the interval that the switch is closed. (After Gibbon and
Church.')

duration that the gate is closed, as shown in the lower graph directly above the
accumulator. The upper graph shows the accumulation value as a function of signal
duration. The switch introduces the non-zero intercept.
The memory system that we propose here is much simpler than is probably realistic
for memory models (for example, that of Heinemann4-'). However, it is quite
complicated enough to analyze, even in this oversimplified form. We propose that the
working memory directly reflects the accumulated count as shown in the proportional
plot above working memory in FIGURE 4 (and in FIGURE 3). When, for a given

- DT

WORKING
-
MENORY: ________ _______
-M i -

REFERENCE
MEMORY,

Ti7
~

FIGURE 4. Memory process. The accumulator in working memory directly reflects counts gated
to it from the pacemaker. On reinforced trials the working memory record for that trial is stored in
a more permanent reference memory. (After Gibbon and Church.')
56 ANNALS NEW YORK ACADEMY OF SCIENCES

condition (series of trials), S + is fixed at some value, and when on a particular trial
S + is presented ( T = S+)and a response is made and reinforced, the value stored in
reference memory, m$+,is a proportional representation of the value on that trial
recorded in working memory (ms+). The comparison, then, on subsequent trials, is
between the current value in working memory, m , and a stored value in reference
memory, m3+.The judgment whether or not to respond is based on some kind of
comparison between these two values.
Several major features of this account are present in an early proposal of
Treisman: in which a pacemaker, counter, store (memory), and comparator occur. To
our knowledge, that was the first model to use a form of scalar timing (equation 11)
explicitly in a timing system. Creelman’ some years ago also proposed a timing model
in which a pacemaker and a counter were involved, but with a Poisson rather than
scalar form of variance. These distinctions will come in for more discussion later.
In our earlier analysis’ we discuss briefly the need for separate lines of evidence to
establish the existence or utility of each of the processes in this account.* In the present
article we will concentrate on features of the account that may be applied in a similar
manner to similar and dissimilar timing tasks.
First, however, a brief summary of our earlier analysis is necessary. We first
examined a case in which there was no variance in the system whatever, so that the
values in the accumulator in working memory and in reference memory were all
one-to-one with the appropriate signal durations.
Two response rules were examined, one of which compared the current time value
in working memory with the remembered time value in reference memory, and
dictated a response when these two values were “close enough,” that is, when the
absolute discrepancy between them lay below a certain threshold. The absolute rule for
the no-variance case is shown in FIGURE 5 in the top row. The tics on the abscissa in the
upper left panel indicate respectively, To,S, and 2S, corresponding to the minimum
signal duration, a given S + duration, and a second S + duration set at double the first.
These might be appropriate, say, to the 2-sec and 4-sec S + conditions in FIGURE1. The
absolute difference between M; and MT is shown decreasing to 0 at S and increasing
beyond it as the solid line function. The positive diagonal hatching shows the
acceptance region within which responses are required when the absolute discrepancy
falls below the threshold value indicated by the horizontal line, B. When the S + value
is doubled (at 2S), the dashed line shows the discrepancy function for this case, and the
negative diagonal hatching indicates the acceptance region.
In the middle panel of the top row the response consequences are plotted in real
time. In the first case for S + = S, responding occurs whenever T is within an absolute
window of S, and the same rule for S + = 2 s produces the same spread around the
reinforced S + value. In the upper right panel these step functions for response
probability are replotted in relative time, as a function of T / S + . The efficiency of the
absolute rule is seen to increase considerably in relative time as S + is increased. The
spread in relative time around S + = 2 s is much smaller than that around S + = S in
this plot. This plot is comparable to that in the bottom panel of FIGURE 1. It shows that
an absolute comparison rule is untenable under these assumptions.
In the bottom row of FIGURE 5, we show an alternative response rule, which
compares the absolute discrepancy between the current time and the remembered
reinforced time, to the remembered reinforced time. This relative rule requires a
response whenever
GIBBON et al.: SCALAR TIMING IN MEMORY 51

The subjective discrepancy is taken as a proportion of the reference memory value for
reinforcement. In the panel in the lower left, the solid line function for S+ = S is shown
decreasing from 1.0 down to 0 and back up for T > S as in the panel above. Now,
however, the threshold, B, is a proportion. When the discrepancy is less than this
proportion, responding is dictated. Again the acceptance region is indicated by positive
diagonal hatching.
The dashed function shows the relative subjective discrepancy for S+ = 2s. When
S+ is doubled, the window size nearly doubles also, as the negative diagonal hatching
shows. In the middle panel the response consequences plotted in absolute time show the
increasing spread around S+ = 2S, and in the right panel the two acceptance regions

n
w I .
> w
- >
+ F
0 ;B
Y E
a
3
cn
Ta S
T-
2s To
-T -
S

TIME
2s
- T/S+
I .0
-
FIGURE 5. Two comparison rules. The left column shows subjective discrepancies between
working memory and reference memory as a function of signal duration with two different
placements of S + . The right two columns shows response probability. The top row describes an
absolute discrepancy rule, in which the absolute dimerence between working and reference
memory values forms the basis of the response decision. The botrom row describes a relative
discrepancy rule in which the absolute discrepancy is taken as a proportion of the representation
ofS+. (After Gibbon and Church.')

are nearly equivalent when plotted in relative time. They are not precisely equivalent
because the To intercept in the accumulation of subjective time counts more heavily
when S+ is smaller than when it is larger. However, even for To relatively large, the
difference is not great. I n this example Tohas been chosen equal to S/4.
Thus, even when there is no variance in the timekeeping and mnemonic system, a
relative response rule is needed to accommodate the superposition of generalization
gradients in relative time.
Of course, the real response gradients are not square waves, and so a no-variance
account will not suffice as a realistic description. In our earlier analysis we studied
several sources of variance which would generate smooth shoulders on the gradients,
58 ANNALS NEW YORK ACADEMY OF SCIENCES

but we found that in no case was the absolute discrepancy rule tenable. Even when
variation increased substantially with S+,the absolute discrepancy rule is too efficient
to accommodate a realistic description of the data as S+ is increased.
With the relative discrepancy rule, however, a number of different sources of
variance were compatible with both the superposition of the data in relative time and
with smooth shoulders around S + (FIG. 1). The analysis showed that at least four
distinct sources of variation in perceiving, remembering, and discriminating time
intervals were possible. Each of these sources was scalar, that is, multiplicative with the
memory for time. Each was associated with different components of the information-
processing scheme. Two additional sources of variation in the system are also potential
contributors to variability, but these sources were shown to be not feasible as a sole
source of variation, were there but one.
One of our current goals is to deepen this analysis by allowing these latter two
sources of variation to be imbedded in variation from multiple points in the system. The
earlier analysis is summarized with respect to the relative discrepancy rule only. The
way in which the two nonscalar sources of variation violate the superposition
requirement will be described briefly first.

Switch: Constant

The simplest of these two sources of variation is the variability introduced by

varying latency to open and close the switch gating pulses into short-term memory. It is
likely that even in highly practiced subjects there is some variation in the perceptual

C0NSTANT:SWITCH

> POISS0N:PACEMAKER
I-
n 1 1
-I
.8
4
m
0
CY
a .4

-1 TIME -T/S+ v
FIGURE 6. Response probability gradients resulting from variation in the switch only (top) or
from Poisson variation in the pacemaker only (bottom). At the /eji gradients are plotted in real
time and at the right in relative time.
GIBBON et al.: SCALAR TIMING IN MEMORY 59

FIGURE 7. Working memory ac-

cumulator. Distributions of counts
associated with two different S +
durations are shown when the only
source of variance is the scalar fluc-
tuation in pacemaker rate. (After
Gibbon and Church.’)

response to external stimulation. A gating mechanism introducing variance of this sort

is similar to an early proposal of Kristofferson and his colleagues which argued that in
some ranges the timing system might reflect only constant, low v a r i a n ~ e . ~A~ ~ ~ ~ ”
realistic form for this latency difference distribution might be most likely back-to-back
exponentials (Laplace distribution). A similar alternative is the triangular distribution
analyzed by Kristofferson as the result of a difference between two rectangular
variates. If the switch is complex, however, and requires the execution of several
components in order to open and close, it might approach a normal form. The niceties
of distinctions between forms will not concern us in the present account, since we will
rely on rather grosser distinctions which are relatively insensitive to distribution form.
We assume a normal form in what follows.
The top row of FIGURE6 shows the effect on response probability of introducing
variance in switch latency only. In the upper left panel response gradients are plotted
against absolute time for S+ = S a n d S+ = 2s. Increasing S+ results both in greater
accuracy at S + ,and in a broader spread around S+. In the panel in the upper right,
these effects counterbalance each other somewhat in relative time. There the largest
discrepancy is a t S+,while the increase in the spread with increasing S+ is seen to be
not sufficient to produce superposition in the wings. Constant variance, which does not
change with increasing size of the interval being judged, results in too efficient a
system at large values of S+. Accuracy is increased at the target duration, and the
spread around it does not increase sufficiently to accommodate superposition.

Pacemaker: Poisson

A second source of variation in this system is one that has received classical
attention in modeling perceptual systems, namely Poisson variation in the discharge of
the pacemaker. From a variety of considerations, variance in the interpulse interval in
a neural pacemaker ought to follow the Poisson I ~ W . ~ . ’ * ’ ~ - ’Interpulse
~ intervals are
exponentially distributed in a steadily varying stream with intensity X = A. The
accumulator in memory is then a Poisson counter. with a variance that increases
directly with the mean.
60 ANNALS NEW YORK ACADEMY OF SCIENCES

In the lower row of FIGURE 6 response gradients for the Poisson pacemaker source
are shown. The gradients in absolute time in the lower left show an increase in
accuracy at S + when S + is doubled. However, this increase is not as large as that for
the constant variance case, and the increase in spread at S + = 2S is somewhat
broader. Nevertheless, when plotted in relative time on the right, the gradients still
deviate substantially from superposition. Thus, this system, which does increase
variance with increasing S+,does not do so fast enough to accommodate superposi-
tion, particularly near S+.
These two kinds of variability introduced by different stages of information
processing are to be contrasted with two other kinds of variability, both of which are
scalar and induce the approximate superposition seen in the data. These are discussed
next.

Pacemaker: Scalar

An alternative source of pacemaker variance is a drifting rate. For simplicity we

imagine that the time between pulses, T , is fixed on any trial, but that from trial to trial
the pulse rate, X = 1 / ~ ,varies normally around a mean, 8.'
In FIGURE 7 the result for the accumulated count in memory is shown for S + = S
and S + = 2s. The rising solid line reflects the mean value of the rate, A, comparable to
the rising line in the upper right panel in FIGURE 3. The dashed rays emanating from To
indicate plus and minus 1 standard deviation of A. When S + is doubled, the
distribution of accumulated counts is nearly a scale transform of the distribution
associated with S , hence the name scalar timing. The solid function distribution for 2S
is generated by our information-processing system with a fixed To. The dashed
function which is nearly identical to it is the simple scale transform of the S
distribution that would result from strict proportionality, an axis multiplier of 2. The
small discrepancy reflects the role of To. Again, To was chosen rather large, equal to
S/4. Even with an intercept of this size, the multiplicative property dominates the
result.
This kind of multiplicative variance in combination with the relative discrepancy
rule results in the response gradients shown in the top row of FIGURE 8. In absolute
time (on the left), it is clear that accuracy when S + = S a n d when it equals 2S remains
the same, but variance increases considerably. In the proportional plot in the upper
right panel, the functions now show near superposition, like the data. Thus, if the only
source of variance were a drifting rate in the pacemaker, the smooth shoulders,
constant discriminability at the positive value, and rough superposition are accommo-
dated by this scalar source.

Memory: Scalar

In our earlier analysis we also studied the possibility of additional noise introduced
when the accumulated counts from the pacemaker are stored in working memory, and
again when these records are transferred to relatively permanent storage in reference
memory. If both of these storage mechanisms involve a proportional transform, then
they too may introduce variability which has the scalar form shown for pacemaker rate

bA more realistic version might allow both kinds of variation simultaneously. Our later analysis
(APPENDIX) allows Poisson variance within trials at a given intensity, A, which in turn is a random
variable that drifts over trials.
GIBBON et al.: SCALAR TIMING IN MEMORY 61

variance in the top row of FIGURE8. The first of these potential sources, that involved
in translation to a short-term storage in working memory, will be ignored i n what
follows here, since with the ratio discrimination rules that we employ, the mean value
of the translation process is cancelled, and its variance may be absorbed in the
pacemaker variance term. Storage in reference memory, however, plays an important
role, and may contribute variance as well as distortion between remembered and
current working memory values. Qualitatively, however, these two sources of variabil-
ity show similar results for the response gradients as long as the reference memory
translation is unbiased. Hence we do not treat them separately here.

SCALAR:PACEMAKER OR MEMORY

W
v,
z
.4
v,
W
0:

0
SCALAR :THRESHOLD
t

TIME T/S+ _3

FIGURE 8. Response probability gradients for two scalar sources. The rop row shows the results
of variation in either the pacemaker or the memory constant. The bottom row shows the results of
variation in the threshold.

Threshold: Scalar

Still another source of variance is realistic for our processing system. This is
variance in the relative proximity of T to S+ that subjects deem “close enough.” One
can readily imagine momentary fluctuations in this level. Multiplying through by the
norming value in Equation 1, variation in b induces the scalar property on remembered
time. The manner in which this property is expressed, however, differs from that for
clock or memory variance.
In the bottom row of FIGURE 8 response gradients are plotted assuming a normally
distributed threshold about some positive value b. The gradients in the lower left panel
are strictly symmetric, have identical accuracy for S+ = S and S+ = 2S, and show a
prominent discontinuity at the positive value. In the lower right panel the relative time
gradients approximately superpose. Thus, this source of variance, like the scalar source
62 ANNALS NEW YORK ACADEMY OF SCIENCES

for clock or memory in the upper row, induces superposition in relative time, as the
data require. The manner in which it is accomplished, however, differs in two ways:
First the upper gradients have a slight, positive skew. The lower gradients in contrast
are strictly symmetric. Second, the upper gradients are smooth, roughly bell-shaped
-
around S+,while the lower gradients are discontinuous at T S+.The discontinuity
results from a feature of our assumptions that is open to question here, namely, a
normal form for the threshold distribution that includes negative values. In practical
terms this assumes that on some trials the threshold is so conservative that not even a
comparison of identical values (no subjective discrepancy) would warrant a response.
The defect may be remedied by truncating threshold distributions to be always
positive, but the result is equally difficult for theory if this is the only source of
variation. This modification induces perfect accuracyat T = S + for all S + values, and
the data show clearly otherwise.
These features of the threshold variance account might prove troublesome for a
description involving only this source of variance. However, in combination with the
clock and memory variability we will see that the symmetry in threshold variability is
an important feature of our account.

Simultaneous Sources of Variation: General Case

From the analysis thus far, we were able to conclude that were one forced to pick a
single source of variation, it would have to be a scalar source, and probably would have
to be located in pacemaker or memory. However, while the smooth character of these
gradients looks like that of our temporal generalization gradients, and while our
temporal generalization gradients have, as do these, some slight skew with a higher
right than left wing, the skew in theory is generally larger than the skew in reality. The
data fall somewhere between strict symmetry around S+, and the scalar forms for
clock or memory variance alone.
This discrepancy leads us to analyze the mixture generated by allowing all of these
potential sources of variability to operate a t once. In the APPENDIX we pursue this
analysis, obtaining an approximation for the underlying random variables in working
and reference memory, and their discrimination via a variable threshold in a
comparator. The result may be summarized as follows: Allowing random variation in
constant, Poisson, and scalar sources results, not surprisingly, in constant, Poisson, and
scalar components of variance in the composite variables underlying the spread of the
response probability gradients. For reasonable choices of parameter values for the
constant and Poisson variance sources, however, the scalar sources dominate variance
in the seconds to minutes ranges studied here. This may be seen in FIGURE 9. In the
APPENDIX the argument is developed showing that under our assumptions the response
probability gradients reflect two random variables, one corresponding to each edge of
the acceptance region for our comparison statistics. The standard deviations of the
window edge variables grow (nearly) linearly with T and S+. In FIGURE 9 the
standard deviation of the upper window edge is shown as a function of increasing T =
S + signal durations for a variety of choices of pacemaker rate, A = 5 , 10, 15, and 20
-
per sec. The figure assumes a switch with a mean and variance of To uo = 0.5 sec,
variance in the threshold of ub = 0.1, and a scalar pacemaker or memory coefficient of
variation of y = 0.2. The point of the figure is that the scalar sources of variance
rapidly come to dominate the standard deviation of the window edge. These functions
are approximately linear in S + for S+ values above about 5 sec (dashed line). Since
the ranges we study here are usually well above this level, we cannot say whether
significant contributions from a Poisson source and a constant source might not be
GIBBON et al.: SCALAR TIMING IN MEMORY 63

present, but masked. The domination by the scalar sources in these mixtures may be
used in the future to specify more precisely the range of parameter values for
pacemaker rate and switch variance that are possible and yet still accommodate our
data. It is clear from the work presented in this volume that the temporal ranges that
interface between milliseconds and hours will be important in the future.

PEAK

A procedure that is very similar to temporal generalization in conception and results,

but quite different in the manner of assessing response strength, is the “peak”
procedure developed by Seth Roberts” and studied by him and Church and Meck.16 It
is a modified discrete-trial free-operant fixed-interval schedule. At the onset of a
signal, usually a light or a noise, rats may respond on a lever, which occasionally will
“pay off” for the next response after S+ sec have elapsed from the beginning of the

n
t
v,
” 4 . 9

Z
0
H
I-
4.
H
>

TIME, S+ CSEC)
FIGURE 9. Standard deviation of a window-edge decision statistic as a function of the size of
S+.The parameter is pacemaker rate.

signal. On some trials, however, responding is not reinforced and the signal remains on
for a long time. Subjects come to anticipate completion of the S+ interval. On trials in
which reinforcement is not programmed, they show maximum responding close to that
time. Typical results are shown in FIGURE 10. The top row shows responding of two
pigeons studied under the peak procedure using a color change on a response key as the
S+ duration signal. On half of the trials the first response after 15 sec was reinforced.
Response rate increases up to a maximum near 15 sec and decreases thereafter.
In the second row, data from a rat studied by Roberts is shown for two different
conditions. In the first condition on 50% of the trials the rat was reinforced for
responding after 20 sec had elapsed since the onset of an auditory signal. The leftmost
function shows the performance. Response rate peaks a little beyond 20 sec and
declines at longer times. The broader function on the right was obtained from the same
rat under a later condition with S+ = 40 sec. These data are quite similar to the
64 ANNALS NEW YORK ACADEMY OF SCIENCES

TIME

FIGURE 10. Response rate gradients from two pigeons (top row) and one rat (bottom row)
studied under the peak procedure.

temporal generalization data in two respects: Accuracy at the peak time is about the
same in both functions, and the spread around the larger S + is approximately
proportionally increased. Indeed, the peak procedure might be thought of as a
generalization procedure which delivers all possible durations on every unreinforced
trial.

Memory Distortion

All of the above cases show some slight displacement of the peak time from the
nominal S+ value. This is not unusual with the peak procedure. The peak for bird No.
3 106 is about 13 or 14 sec and the peak for rat No. 2 studied under S + = 20 sec occurs
near 24 sec.
In FIGURE11 in the left hand column the effect of deviation from S + is shown for
the data of rat 2. The top functions show the peak rate as a function of absolute time as
in FIGURE10. In the middle panel, these functions are shown plotted in relative time,
T / S + , as in FIGURE1 for temporal generalization. Now superposition is not achieved.
The S + = 20 sec curve shows a broader right wing in relative time.
In the model, the information-processing account allows for distortion of the mean
in the translation constant between working and reference memory.” Working
memory is assumed to reflect pacemaker accumulation directly, while the storage into
reference memory may introduce both additional variance of its own, and systematic
distortion in the mean (upwards or downwards). The translation constant, K*, required
for these data is somewhat above I .O, hence the failure of superposition.
In the lower left panel, a fit to the peak data has been accomplished and the data
are now plotted relative to K * S + , that is, relative to the mean of the memory
distribution for the reinforced time. When adjusted in this manner, superposition
succeeds.
A more extreme example of memory distortion which nevertheless preserves
GIBBON et al.: SCALAR TIMING IN MEMORY 65

superposition in the appropriate metric is obtained from an experiment studying two

peak procedures in rats simultaneously. A peak procedure value of 10 sec in the
presence of a light and 30 sec in the presence of a noise were studied on randomly
intermixed trials. Median performance of a group of ten rats is shown in the right hand
column of FIGURE11. In the upper right panel the two peak functions plotted in real
time are seen to be centered close to 10 sec and 30 sec, as would be expected from the
preceding data. The functions differ somewhat from those of the individual subjects in
FIGURE10. Both functions are fairly flat near the peak, leaving some ambiguity as to
just which time value corresponds to the “true maximum.” However, even with this
flatness it is clear that the 10-sec data lie mostly above 10 sec, while the 30-sec data lie

75

75

25
25

20 48 68 20 40 60
T 1

1
W
75 if%.
v,
z
0
a
v,
w
&
25

1.0 2.0 3.0

25

/ , t L,, ,
2 0 4.0 6 0
T/S+ T/S-

1.59 3.17 4.76 I 46 2 92 4 38

T/K”S+
TIME -r/<*s+

FIGURE 11. Response rate gradients for rat 2 in the left column, and for a group of rats studied
under two different S + values on the right. Gradients in the top row are plotted in real time, in the
middle row in time relative to S + time, and in the bottom row in time relative to the empirically
determined memory representation of S + .
66 ANNALS NEW YORK ACADEMY OF SCIENCES

mostly below 30 sec. It is as though the subjective representations of the 10-sec and
30-sec S+ experiences have become mixed or perhaps mutually attracted in memory.
In the middle panel of FIGURE 11 these data are plotted in relative time, and the
systematic distortion is again evident. The 10-sec curve appears broader than the
30-sec curve in relative time, because its “true” S+ time is larger than 10 sec and the
“true” 30-sec S+ time is less than 30 sec. This is similar to the deviation on the left, but
more extreme. In the bottom panel, the results of fitting these data, allowing for
memory distortion, show clear superposition when plotted in time relative to the
memory time ( T / K * S + ) .
Thus, the peak procedure provides temporal gradients with a maximum near the
time of reinforcement, but deviations appear both idiosyncratically and because of
potential interactions in reference memory when more than one time value must be
retained. The results of this analysis imply that this sort of distortion is a multiplicative
one. The scalar property applies to the memory for time, not necessarily to real time.

TIME LEFT

Our final application of these ideas is to a procedure that differs considerably from the
temporal generalization or peak procedures. In a previous report, Gibbon and Churchla
studied choice procedures with rats and pigeons; these procedures were designed to
reveal preference for the subjectively shorter of two delays to food, when one of these
delays was elapsing. The aim of that work was to examine these choices parametrically
a t a variety of delays, since parametric data should reveal curvature or linearity in
subjective time. The linearity result will not be reviewed here for that implication, but
rather some new data from the same procedure under several conditions will be
examined for the superposition property.
The procedure studied with pigeons is shown in FIGURE12. At the beginning of the
trial (initial link), two keys are lit with different colors, say white and red, and birds
distribute responding across the two keys as time elapses during the trial. At some
point (T), the next response produces mutually exclusive (terminal link) consequences
on either key. If the next response is the white key, the red key is extinguished, and
responding on the white key may continue for the remainder of the white-key interval,

FIGURE If. “Time left” procedure. Two keys

are lit with different colors (white or red) at
the beginning of the trial. At some randomly
chosen point, T, the next response results in
one or the other of two mutually exclusive
delays to reinforcement (bull’s-eyes).

0 T T+30 68
GIBBON et al.: SCALAR TIMING IN MEMORY 67

fY
0
L
W
0
z
W
fY
W
G
W
fY
a

15 T i n 45

T I M E CT)
FIGURE 13. Psychometric preference function. Preference, the proportion of “C” choices at
successive points during the choice period, increases from favoring “S” at the start of the trial to
favoring “C” as the trial elapses. The point of indifference is indicated by the dashed line over
TI,,.

in this example 60 sec, when reinforcement is made available. Thus, on the white key,
reinforcement is available after a total of 60 sec from the beginning of the trial.
If, at the choice point, the next response is to the red key, the white key is
extinguished, the red key changes color to green, and responding will be reinforced for
pecking the green key after 30 sec have elapsed since the color change. Thus, this delay,
called the standard ( S ) ,is fixed a t 30 sec. The entry times ( T )vary from trial to trial.
Early in the trial, it behooves subjects to respond to the red key since the delay to
food there is shorter than on the comparison 60-sec interval. However, as the choice
period elapses, the time left on the comparison (C = 60 sec) side becomes shorter than
the 30-sec standard, and now it behooves subjects to respond on the white key. The
typical result is shown in FIGURE13, in which choice responding begins heavily
favoring the standard, and at some point in the trial switches over to favor the
comparison time-left side as that delay becomes more favorable. The point at which the
switch is made, T,,,, is a datum of primary interest. If subjective time is linear and
subjects are unbiassed in their appreciation of this time, then they should switch over to
preferring the time-left side of the choice at precisely C - S sec into a C-sec
comparison interval-at 30 sec into the 60 sec interval in the above example. In fact, it
is common for subjects to switch somewhat before the midpoint value for C = 2S, and
the bias in favor of the elapsing interval may be due to a preference for a key color
paired with primary reinforcement. In the analysis that follows, the crossover point,
T,,,, plays a fundamental role.
The information-processing account for this procedure has three distinct condi-
tions: the choice period, and each of the mutually exclusive terminal link consequences.
The scheme we use to model choice responding is shown in FIGURE14. When the choice
period begins, the switch gates pacemaker pulses into a working memory accumulator
for trial time, T. The difference between the memory for current time and the memory
68 ANNALS NEW YORK ACADEMY OF SCIENCES

for the comparison ( C ) interval represents the subjective time left on the comparison
side. This remaining time is compared in the comparator with memory for the time
‘‘left’’ on the standard side-the standard delay to food were the terminal link to come
on immediately.
As time accumulates during the trial in working memory, the remaining time on
the comparison side is continually updated, and compared with memory for the
standard delay. At the outset of the trial, the standard delay appears shorter than the
remaining time on the comparison delay and hence preference for the comparison
interval is low. At some point subjects cross over and prefer the remaining time on the
comparison side, when their subjective assessment of this delay makes it appear more
favorable, by a potentially biased threshold. Note that this threshold, b’, is distinct
from the proportional threshold defining a response window in the temporal general-

CLOCK

ZKfl!
I
N
K
R
0
N
R
E
MEMORY

DECISION

J.
“ s “

1 I
T
TIME
FIGURE 14. Information-processing model for choice in the “time left” task. During the choice
period the switch gates pacemaker pulses to working memory for elapsed trial time, T. A
comparison is made between the discrepancy between C and the current time-time left to food
on the comparison side-and the standard delay to food, S. Preference is for the shorter of these
two delays.

ization case, Here b‘ indicates a bias in favor of the time-left or standard side, with an
unbiased mean value represented by B = 1.0. We expect, however, that variance
associated with the decision rule for both cases may be quite comparable.
The two terminal link conditions establish the memory values for the comparison
and standard delays. In FIGURE 15, processing of the time-left terminal link is shown.
The switch continues to gate pulses into the working memory for T, until reinforcement
occurs when T = C a t the end of the trial. During this terminal link subjects are in a
go/no-go situation precisely comparable to the left wing of a peak procedure (a fixed
interval schedule). Hence their decision to respond or not to the white key is assumed to
be based upon the time left to reinforcement normalized by the overall time to
reinforcement, as shown in the decision rule in the lower right. After reinforcement,
the reinforced value from working memory (mc) is stored in reference memory (mr)
for the comparison interval.
GIBBON et al.: SCALAR TIMING IN MEMORY 69

-R

0 T C
I I I

TIME

FIGURE 15. Information-processing model for the time-left terminal link. The switch continucs
to gate pulses to the trial-time accumulator, and the comparison is a go/no-go decision, as in the
peak procedure.

When on other trials the choice at the entry point is in favor of the standard, the
standard terminal link is entered. FIGURE 16 shows the switch gating pulses into a new
working memory accumulator (m,) as subjects begin timing the standard interval.
Go/no-go responding is again assumed to be comparable to the left wing of a peak
procedure, and the decision rule reflects the relative proximity to food. After
reinforcement, the value in the working memory store (ms) is transferred to a reference
memory for the standard interval (m:).In this way the memories for C and S are built

DECISICN

1 I I I I
0 T Tt1 TtS C
TIME

FIGURE 16. Information-processing model for the standard terminal link delay. During this
delay the switch gates pulses into an accumulator for the standard delay in working memory, and
the comparison is again a go/no-go decision, as in a peak procedure.
70 ANNALS NEW YORK ACADEMY OF SCIENCES

up over training and then become available for sampling during the choice period of
later trials. In principle it should be possible to predict on individual trials the details of
responding in the terminal links from features of responding in the choice link. We will
pursue this line in later work. However, in what follows, we concentrate on analysis of
preference, pooled over the choice period.
The time, TI/*, at which subjects cross over during a sufficiently lengthy choice
period from preferring S to preferring time-left is revealing. The decision rule implies
that at this point the delays are subjectively equal. On average,

Mr - M,,,,
= BM;. (2)
Translating these mean values according to Equation 1 gives us

TI12 =
:(- -
9
B K*S + To[1 + K*(B' - l)].
Indifference points should be linear in S, for constant ratios of C to S, as described
(3)

previou~ly.'~ Note that for To = 0 and no bias (B'= 1) or memory distortion (K* = l ) ,
TI,, = C - S , when the actual time left, C - T,,2requals S . Deviations from physical
equivalence may be introduced by bias and memory distortion. TI,, may be thought of
as a set-point at which these factors have been subjectively equalized.
The performance of three subjects studied under three combinations of S = C/2
(7.5, 15; 15, 30; 30, 60) are shown in the three panels in FIGURE 17. The S value is

b
L
W
-I
1 .a
E
H
I- FIGURE 17. Psychometric preference
E .8
functions for time left for three subjects
0 at three different pairs of S = C/2
L
values. The functions are plotted
W against real time in the trial. The S
u
z .2 parameter values are indicated in the
r
Y
W top panel.
W
L 1.0
W
E
a
.6

.2

15 30 45 80

T I M E CT)
GIBBON et al.: SCALAR TIMING IN MEMORY 71

t-

H
I-
CY
FIGURE 18. Preference functions for these three 0
LI
subjects replotted from FIGURE 17 as a function of
time relative to the indifference point, T,,,. The W
0
symbol code is the same as in FIGURE 17. Z
W
CY
w
L L I
W
CY
a

1 .8 2.8

T I M E CT/T 1/2 )

indicated above each curve in the top panel. These preference functions, like those
published earlier,” show a shallower rise as S = C / 2 is increased, and T I , ,values are
approximately linear in S , as expected (Equation 3).
Our interest centers on a superposition property of the account, which is developed
in the APPENDIX. There it is shown that if these curves are normalized by their T,,*
values, the effects of changing bias (B’ # I ) and memory encoding (K* # 1) are
cancelled in much the same way as the effects of changing K* may be cancelled in the
peak procedure through normalizing by K * S + . Here normalization by T I , , entails
superposition as long as the scalar variance sources dominate. This will be true if the
time ranges are large relative to To and the Poisson rate parameter is greater than
about 10 per sec. Plotting the preference functions at T / T , , , results in the superposi-
tion of these functions shown in FIGURE18. Thus, again a Weber’s law property is
revealed, but with the important proviso that it operate on subjective, not objective
time. Distortions in memory and innate preferences operate to rescale objective time.

ACKNOWLEDGMENTS

We gratefully acknowledge the assistance of Stephen Fairhurst and Amy Waring

in all of the data and theory analysis presented here.
12 ANNALS NEW YORK ACADEMY OF SCIENCES

REFERENCES

I. GIBBON, J. & R. M. CHURCH. Sources of variance in an information processing theory of

timing. Presented at the Harry Frank Guggenheim Conference on Animal Cognition,
Columbia University, June 2-4, 1982. In press.
2. GIBBON, J. 1977. Scalar expectancy theory and Weber’s law in animal timing. Psychol. Rev.
8 4 279-325.
3. CHURCH, R. M. & J. GIBBON.1982. Temporal generalization. J. Exp. Psychol. Animal
Behav. Processes. 8: 165-1 86.
4. HEINEMANN, E. G. 1982. A memory model for decision processes in pigeons. In Quantita-
tive Analyses of Behavior. Vol. 3: Acquisition. M. Commons, R. J. Herrnstein, and A. R.
Wagner, Eds. Ballinger. Cambridge, MA.
5. HEINEMANN, E. G. 1982. The presolution period and the detection of statistical associa-
tions. In Quantitative Analyses of Behavior. Vol. 4: Discrimination Processes. M.
Commons, R. J. Hermstein, and A. R. Wagner, Eds. Ballinger. Cambridge, MA.
6. TREISMAN, M. 1963. Temporal discrimination and the indifference interval: Implications
for a model of the “internal clock.” Psychol. Monogr. 77: 13.
7. CREELMAN, C. D. 1962. Human discrimination of auditory duration. J. Acoust. SOC. Am.
34 582-593.
8. CHURCH, R. M. 1984. Properties of the internal clock. This volume.
9. KRISTOFFERSON, A. B. 1967. Attention and psychophysical time. Acta Psychol. 27: 93-
100.
10. ALLAN.L. G., A. B. KRISTOFFERSON& E. W. WIENS.1971. Duration discrimination of
brief light flashes. Percept. Psychophys. 9 3 2 4 3 2 7 .
II. WING,A. M. & A. B. KRISTOFFERSON. 1973. Response delays and the timing of discrete
motor responses. Percept. Psychophys. 1 4 5-1 2.
12. GETTY,D. J. 1976. Counting processes in human timing. Percept. Psychophys. 2 0 191-
197.
13. MCGILL,W. J. 1967. Neural counting mechanisms and energy detection in audition. J.
Math. Psychol. 4 351-376.
14. GREEN,D. M. & R. D. LUCE. 1974. Counting and timing mechanisms in auditory
discrimination and reaction time. In Contemporary Developments in Mathematical
Psychology. Vol. 2: Measurement, Psychophysics, and Neural Information Processing.
D. H. Krantz, R. C. Atkinson, R. D. Luce & P. Suppes, Eds. W. H. Freeman. San
Francisco, CA.
15. ROBERTS, S. 1981. Isolation of an internal clock. J. Exp. Psychol. Animal Behav. Processes
7:242-268.
16. MECK,W. H. & R. M. CHURCH. 1981. Simultaneous temporal processing. J. Exp. Psychol.
Animal Behav. Processes 7:242-268.
17. MECK,W. H. 1981. Selective adjustment of the speed of internal clock and memory
processes. J. Exp. Psychol. Animal Behav. Processes 7: 242-268.
18. GIBBON, J. & R. M. CHURCH. 1981. Time left: Linear versus logarithmic subjective time. J.
Exp. Psychol. Animal Behav. Processes 7: 87-108.

APPENDIX
Temporal Generalization and Peak

W e require t h e total probability of t h e event in Equation 1, t h e decision rule, when

all stages in t h e information-processing s c h e m e may contribute variability. T h a t is, w e
require t h e total probability of
GIBBON et al.: SCALAR TIMING IN MEMORY 73

where m:, mT, and 6 are each independent random variables sampled from reference
memory, working memory, and from a threshold distribution on each trial. We develop
this calculation by obtaining the mean and variance of the memory random variables,
when these receive variance contributions from each of the postulated sources of
variance in FIGURES2 through 4. Then the analysis follows that of Church and
Gibbon.’ At several points in the development normal approximations for distribution
forms that are demonstrably slightly skewed will be used. We have in each case run
computer simulations to satisfy ourselves that the approximations are reasonable. Our
strategy is to add successive sources of variance moving through the information-
processing chain from the pacemaker to the comparator.
The pacemaker generates pulses with Poisson variability a t an intensity, A. Imagine
first that the effective switch closure time, dn is fixed at the mean value, Dpc
The number of counts accumulated in working memory associated with this switch
closure time is then a Poisson-distributed variate with a mean and variance
E(m) = Var(rn) = AD.
We now allow the intensity parameter to vary. The moment generating function for the
Poisson variate with fixed intensity is
Mm,(B)= exp[XD(ee - I)],
where we have replaced the subscript T by X to indicate dependence on X at any fixed T .
If we now let X vary around a mean A with standard deviation yA, the moment
generating function of the mixture then becomes

J-,

Assuming normality off, we may complete the square under the integral and factor so
that X is integrated out of (A3). With some rearrangement the moment generating
function may be shown to be
M,(O) = exp [AD(eo - 1 ) + 1 / 2 (yLiD)*(ee- (‘44)
Taking first and second derivatives and setting 0 = 0 gives the first and second
moments as
p; = E(m) = AD,
pi = AD + (AD)’(I + y’), (A54
so that the variance of counts i n working memory under both Poisson and scalar
sources of variation is given by
Var(m) = pi - p\2 = AD + (TAD)*. (A5b)
The distribution form corresponding to the moment-generating function for the
mixture (A4) is not normal, but approaches normaIity rapidly as Tgrows. In the ranges
we will discuss, the forms are reasonably well approximated by normal distributions
with mean and variance given by Equations A5a and 5b.

‘Subscripting will be dropped henceforth where the parameter dependence is obvious, and
picked up again where needed. In the present case we assume the stimulus is on for an arbitrary
but fixed period of time, Tsec, and thus dependence on Twill be suppressed.
14 ANNALS NEW YORK ACADEMY OF SCIENCES

We now wish to allow variability in the latency to open and close the switch. That
is, we wish to obtain the mean and variance of the mixture that results from allowing d
to vary normally about a mean, D = T - Towith standard deviation a,. The conditional
rules for mean and variance of x conditioned on y" may be written

For the present case, allowing switch variance results in a mean and variance for
working memory given by

E(m) = M = AD, ('47a)

Var(m) = uf, = a(Au0)*+ AD + (a - l ) ( ~ i D ) ~ , (A7b)

where

Working memory may be construed as involving a proportional translation for

short-term storage of the count in the accumulator. Such a translation might add
another source of variance. This source was considered in our earlier piece' and is not
discriminable in form from variance induced by pacemaker rate variability, and hence
we do not analyze it separately here.
Translation to reference memory occurs only on a subset of the trials, but after long
training we assume the reference memory has built up a distribution of remembered,
reinforced working memory values via the proportional translation indicated in FIGURE
4. The conditional rule (A6a) gives, for m,*. = k*ms+,

And a lengthy but straightfoward application of (A6b) gives, for uk. = y * K * ,

Var(m*) = E[Var(m$)] + Var[E(m,*.)],

or
uf = + a*ADs+
K*2[aa*(Auo)2 + (aa* - l)(ADs+)2], (AW
where

We are now in a position to analyze the decision rule, assuming independent,

normally distributed memory variates.'
If we assume that the three variables in (Al) are independent and (essentially)
dFeller, W. 1966. An Introduction to Probability Theory and Its Applications. Vol. I: 164.
Wiley. New York. NY.
eThe assumption of normality in the (product) memory variates violates somewhat the true
character of these distributions, which have some positive skew. Again, however, in the range of
seconds to minutes, this skew is not substantial and normal approximations are reasonable.
GIBBON et al.: SCALAR TIMING IN MEMORY 75

positive, the decision rule may be written

m
I-b<-<l+b. ('49)
m*
The correlated upper and lower limits (window edges) do not pose special difficulties
here and may be treated as independent variables (compare the Appendix in Church
and Gibbon').
Defining w, = 1 + ( - I ) % with mean and variance,
E(w,)= w, = 1 + (-I)'&
Var(w,) = ui, i = I , 2,
Equation A9 becomes
m*w, < m < m*w2. ('410)
Letting x, = rn - m*w,,

E(x,)= M - M,*+
W,. (Alla)
Repeated application of Equation A6b shows
Var(x,) = ui+ uf.(ui + wf) + M,*: u:, (Allb)
where and m i . are given by Equations A7b and A8b. Assuming the x, are normal, we
have

WI T ) = W Z 2 ) - NZ,), ('412)
where

and Q, is the unit normal distribution function. It is convenient to divide numerator and
denominator of the 2,by M,*+, giving

Some algebraic rearrangement in Equations A7 and A8 gives

and

(z) =
1
[ U*ad + a* DS+
A
-
1
+ (a*a - I)D;+ ,
76 ANNALS NEW YORK ACADEMY OF SCIENCES

where a* and a are as defined in Equations A7 and A8. The denominator of Equation
-
A1 3 with i 2 was used for the window-edge standard deviation in FIGURE 9. The form
(A12), with definitions (A13), (A14), and (A15), was used to plot the generalization
gradients in the text. The peak procedure gradients require the additional scale factor
-
R,,,, translating response probability into response rate,
-
RT = RmaxP(R I 0. (A14

Time-Lef i Procedure

The time-left procedure is analyzed in a similar fashion, but with two memory
variates, m: and m$, corresponding to reinforced exposure to the S and C delays. The
memories are built up in the same way as in temporal generalization and (A8) and (A15)
hold for S + = S, C.
The decision rule for choice for time-left may be written:
m$ - mT < b'm;. ('417)
Defining

where

(A19a)

and
E(x) = M$ - M? - B M & (A19b)
Var(x) = uk + u i r + u i ; (uf + B2)+ M,*2ui,. (A19c)
Again, it is convenient to divide numerator and denominator of (A19a) by M:,
giving

(zy(&),[
where (umS/M$)'is defined by (A15) and

= n*au; + a* Dc
- + (a*a - 1)D: .
A I
This form (A18, A20, and A21) was used in fitting the psychometric functions in
figures in the text.
To see why near superposition is achieved when P("C") is normalized by T,,,, we
GIBBON er al.: SCALAR TIMING IN MEMORY 77

consider only large S , C, with SIC constant, so that Tomay be neglected. From (A17)
T I 1is2 given by

T I , ,= ;( - B’) K*S.

The numerator of (A20) becomes:

which is constant at TIT1,*.

The denominator terms may be analyzed separately. As S , C becomes large, the
constant and Poisson components of variance become negligible and, from (A21)
(+ (a*a - 1) );( 2

From (A15)

(ZY- (a*a - l ) ,

and from (A14),

Thus, the variance term approaches constancy for constant T / T I i 2also, resulting in
superposition in (A18).