Ghirlanda 2003

ANIMAL BEHAVIOUR, 2003, 66, 15–36
doi:10.1006/anbe.2003.2174
REVIEWS
A century of generalization
STEFANO GHIRLANDA & MAGNUS ENQUIST

Group for interdisciplinary cultural research, Stockholm University
(Received 21 September 2000; initial acceptance 15 November 2000;

final acceptance 20 November 2002; MS. number: RV-25)
We review data from both ethology and psychology about generalization, that is how animals respond to
sets of stimuli including familiar and novel stimuli. Our main conclusion is that patterns of general-
ization are largely independent of systematic group (evidence is available for insects, fish, amphibians,
reptiles, birds and mammals, including humans), behavioural context (feeding, drinking, courting, etc.),
sensory modality (light, sound, etc.) and of whether reaction to stimuli is learned or genetically inherited.
These universalities suggest that generalization originates from general properties of nervous systems, and
that evolutionary strategies to cope with novelty and variability in stimulation may be limited. Two
major shapes of the generalization gradient can be identified, corresponding to two types of stimulus
dimensions. When changes in stimulation involve a rearrangement of a constant amount of stimulation
on the sense organs, the generalization gradient peaks close to familiar stimuli, and peak responding is
not much higher than responding to familiar stimuli. Contrary to what is often claimed, such gradients
are better described by Gaussian curves than by exponentials. When the stimulus dimension involves a
variation in the intensity of stimulation, the gradient is often monotonic, and responding to some novel
stimuli is considerably stronger than responding to familiar stimuli. Lastly, when several or many familiar
stimuli are close to each other predictable biases in responding occur, along all studied dimensions. We
do not find differences between biases referred to as peak shift and biases referred to as supernormal
stimulation. We conclude by discussing theoretical issues.
 2003 The Association for the Study of Animal Behaviour
The study of how external stimuli affect behaviour has of generalization or theories that more or less directly
been referred to as the theory of stimulus selection in summarize observations (for example by incorporating
ethology and stimulus control in experimental psy- observed features of generalization into the theory, Hull
chology, and has played a key role in both disciplines 1943; Mackintosh 1974). In this review we aim to organ-
during the 20th century. A key finding of such research is ize existing data in a way useful to develop and test
generalization: if a behaviour has been established in theories of generalization. We also point out findings that
response to a stimulus, novel stimuli resembling the first conflict with existing theories, and we conclude with a
one will usually elicit the same response. Usually, modi- discussion of theoretical issues. Finally, we hope that this
fied stimuli are less effective than familiar ones, but review will be helpful as a guide of what to expect when
sometimes they are even more potent in evoking the reactions to novel stimuli are important, for example in
response. This finding has been referred to as ‘supernor- experimental design.
mal stimulation’ by ethologists, ‘peak shift’ by psycholo-
gists, and more recently ‘response bias’ (see, respectively, APPROACHES TO THE STUDY OF
Tinbergen 1951; Mackintosh 1974; Enquist & Arak 1998). GENERALIZATION
Interest in theories of generalization seems to have faded
in recent years, although our understanding is still unsat- Ethology and Experimental Psychology
isfactory (Mackintosh 1974; Ghirlanda & Enquist 1999).
Data about generalization come primarily from etho-
Behaviour is often ‘explained’ by merely empirical rules
logical and psychological studies of behaviour. Within
Correspondence: S. Ghirlanda, Kräftriket 7B, Stockholm University, experimental psychology, animals are typically trained to
S-106 91 Stockholm, Sweden (email Stefano.Ghirlanda@intercult. perform a response to one stimulus, called the positive
su.se). stimulus (S + ), and not to a second, negative stimulus
15
0003–3472/03/$30.00/0  2003 Published by Elsevier Ltd on behalf of The Association for the Study of Animal Behaviour.
16 ANIMAL BEHAVIOUR, 66, 1
(S ). The S may be absence of S + or a stimulus differing Table 1. Contribution of rearrangement and intensity of stimulation
from S + in characteristics such as visual size or sound along common stimulus dimensions
frequency. More complex arrangements, with several Contribution of
positive and negative stimuli, have also been used. After
training, the animals’ reactions to a set of test stimuli are Stimulus dimension Rearrangement Intensity
recorded. These are usually chosen from a ‘stimulus
dimension’ obtained by varying a physical variable such Intensity of sound L
as wavelength of light or intensity of sound. Data can Intensity of light L
thus be represented in the form of a response gradient Chemical concentration L
along the dimension; this is the chief analytical tool of (smell, taste)
the psychological tradition, from which key concepts Object size L L
Complex sound spectra L L
such as ‘peak shift’ and ‘stimulus control’ are defined (e.g.
Complex light spectra L L
Terrace 1966). The test stimuli typically include S + . If S Object shape L S
lies on the test dimension, too, it is customary to speak of Tone frequency L S
an ‘intradimensional’ discrimination, for example when Monochromatic light L S
light wavelength generalization is studied after a discrimi- Object orientation L
nation between two wavelengths. If S cannot be placed Object location L
on the test dimension, the term ‘interdimensional’ dis-
L: Large contribution; S: small contribution, or varying contribution
crimination is used, for instance if the test dimension is depending on the exact stimuli used; no symbol means negligible
sound frequency and S is silence or white noise. The contribution.
same discrimination can be both intra- and inter-
dimensional depending on the test stimuli chosen. We
thus prefer to speak of inter- and intradimensional tests, objective properties of stimuli. One drawback of this
rather than discriminations. approach is that it ignores how stimuli are received by
Whereas psychologists are almost exclusively based in the sense organs. For instance, focusing on wavelength
the laboratory, ethologists have mainly studied behaviour of light does not explain why ultraviolet light cannot
in nature. Although animals are seldom trained by the control behaviour in some animals. But if we consider
researcher, discriminations are common in the wild as photoreceptors, we discover that some animals have
well. For example, to incubate their eggs birds must none that react to ultraviolet light. Here we focus on two
discriminate between the egg and the nest background; to aspects of sense organ activation patterns. The first is the
pursue females male butterflies must discriminate their intensity of stimulation, by which term we mean the
presence from their absence, and so on. The main total activation of receptors. This is related to physical
research tool in ethology has been the use of dummies intensity but is not identical with it; for instance, a sound
resembling natural stimuli but with added, removed or of 100 kHz does not elicit any activation in human ears
modified features. For instance, the egg retrieval behav- regardless of its physical intensity. The second aspect is
iour of the herring gull, Larus argentatus, has been studied how a given amount of stimulation is distributed among
using dummy eggs of different sizes, colours and shapes receptors. We refer to stimulus dimensions along which
(Baerends 1982). intensity does not change as ‘rearrangement dimensions’,
We can thus summarize both the ethological and the since different stimuli along such dimensions correspond
psychological methods as the recording of animals’ reac- to a different arrangement of the same amount of
tions to stimulus sets including novel and familiar stimulation on the sense organs. Table 1 shows how
stimuli. In interpreting results from such experiments, it common stimulus dimensions can be classified according
is important to remember that behaviour is influenced to this scheme. The rationale for such a classification is
both by individual experiences and by the evolutionary that variation in intensity has significantly different
history of the species. Even if the stimuli used in the behavioural effects, compared with rearrangement of
laboratory have little significance for animals in the wild, stimulation (see below).
pre-existing responses and predispositions can influence
behaviour. For instance, novel stimuli can elicit fear. DATA SELECTION
Thus, laboratory experiments only approximate the ideal
situation of an empty memory modified by experience Although experimental paradigms in generalization
with only one or two stimuli, even when ‘naive’ individ- research can be summarized succinctly in their main
uals are used. The analysis of natural behaviour is even points, countless variations exist. We have tried to
more complex, both because the animals’ evolutionary include as many studies as possible in our analyses, but
history more directly influences behaviour and because we deemed some unsuitable. Our guidelines may be
experimental control over individual experiences is at summarized as follows. First, we included no study with
most partial. fewer than three subjects per treatment, except when we
considered the effect of group size. Second, responding
should be probed at a sufficient number of stimulus
Analysis of Stimulus Dimensions
locations to make inferences about gradient shape. When
In both ethology and comparative psychology, results fitting a curve to the data, we required that more than
of generalization tests are most often analysed in terms of 10 stimulus locations be probed. One exception is the
REVIEWS 17
analysis of response biases, where responding to a single Staddon & Reid 1990; Cheng et al. 1997). We analysed
test stimulus may serve to estimate a bias. Third, we have 223 rearrangement gradients and found that Gaussian
not included studies whose outcome was importantly curves account for about 3% more of the variance in
affected by unusual features of training or testing, apart observed data (estimated by r2; appendix 2 describes the
from when we discuss the specific effects of such features fitting procedure). This difference is small, but significant
(see especially intensity dimensions below). Fourth, we (Test 1, P<10 10). Both Gaussians and exponentials
could not include some studies in quantitative analyses account for more than 90% of the variance in most cases,
because they reveal generalization indirectly, for instance and about 25% of the gradients conform better to the
by studying how reproduction or position in a social exponential shape. Figure 2 illustrates these results.
hierarchy is affected by signals used in social interactions The experiments we surveyed were designed to give an
(e.g. Burley et al. 1982; Burley 1986; Johnson et al. 1993). overall picture of generalization gradients rather than to
Nevertheless, these are powerful examples of the biologi- decide between two specific hypotheses. The region
cal significance of generalization. Lastly, we do not around the peak, most important to discriminate expo-
include temporal generalization, owing to our lack of nentials from Gaussians, is often poorly sampled. Better
familiarity with the field and the additional space it sampling appears to favour Gaussian fits, as shown in
would require. Fig. 3 relative to light wavelength generalization in the
We turn now to reviewing the available data on gener- pigeon, Columba livia (see for instance the data from
alization, considering rearrangement dimensions first, Blough 1975 in Fig. 1). Note also that fitting attempts are
then intensity gradients, and lastly dimensions along evaluated only in the light of sampled values. Strictly, we
which both the amount and the arrangement of stimu- cannot say anything about other values. Yet we do not
lation on the sense organs vary. In order not to burden expect the actual gradient to depart systematically from a
the text with the description, results and data sources of good fit. Inspection of Gaussian and exponential fits (e.g.
statistical tests, we have collected this information in Fig. 1) shows that exponential fits often predict a con-
Appendices 1 and 2. In the following, we refer to tests by siderably taller gradient than actually observed (an aver-
their number in Appendix 1. A short summary concludes age of 20% higher, Fig. 4a). Predictions from Gaussian
each major section. fits, on the other hand, are distributed around the
observed values (Fig. 4b). This suggests that Gaussian fits
estimate gradient height more accurately. There is instead
REARRANGEMENT DIMENSIONS no difference between predictions of Gaussian and expo-
nential fits about the location of the peak (Test 3, NS).
Generalization gradients peaking at or near the positive
Finally, we note that gradients that are clearly neither
stimulus are considered the prototypical finding about
exponential nor Gaussian exist (e.g. the bottom gradients
generalization (Fig. 1). Such gradients have been found
in Fig. 1; Hoffman & Fleshler 1964; Blough 1972).
along diverse stimulus dimensions such as light wave-
length, tone frequency, object orientation and object
location (Table 2). Stimuli along these dimensions are Gradient Symmetry
best described as corresponding to a rearrangement of
stimulation with respect to the S + , without much change Nearly all theories of generalization assume or predict
in the total activation of sense organs. For instance, all that gradients obtained from interdimensional tests are
positions and orientations of lines or squares in the symmetrical around the S + (Spence 1937; Hull 1943;
(centre of the) visual field give rise to the same amount of Blough 1975; Shepard 1987). However, reproducible
stimulation in the eye. This is also true of tones of the asymmetries have been reported. When data from all
same physical intensity and not too different frequencies subjects taking part in an experiment are published, we
(e.g. human hearing, Coren et al. 1999). Variation in light can test whether they are consistently skewed towards
wavelength can be classified as a rearrangement dimen- one or the other side of the S + (see Appendix 2 for
sion as well, that is, total receptor activation is approxi- details). For instance, individual gradients in Hearst et al.
mately constant over considerable wavelength ranges in (1964, pigeon, line orientation, S + =vertical line, S =no
many species. line) are skewed on the side of clockwise rotations (Test 4,
P<0.05).
When individual data are not published, we can gather
Gradient Shape group averages from different studies conducted under
similar conditions, and look for a systematic across-study
Spence (1937), in his pioneering work on stimulus asymmetry. We can thus confirm that the the skewed
control, used parabolic functions to introduce the con- gradient reported by Hearst et al. (1964) has been consist-
cept of a generalization gradient. In later work, he also ently observed in other studies (Test 4, P<10 4);
assumed bell-shaped gradients (Spence 1942). He responses to anticlockwise tilted lines are on average 92%
acknowledged that these choices were purely illustrative, of responses to clockwise tilted ones (range 76–104%).
lacking at the time reliable data. Hull (1943), based on Similarly, an analysis of studies of light wavelength
Hovland’s (1937) data, incorporated an exponential generalization in pigeons reveals that generalization
function into his theory of behaviour. Nowadays around S + =550 nm is consistently skewed (Test 5,
researchers mostly hold gradients to be either expo- P<10 6). Wavelengths shorter than 550 nm total an
nential or Gaussian (e.g. Blough 1975; Shepard 1987; average of 67% of the responses to longer wavelengths
Blough (1975) Blough (1969)

0.15 S
+
= 597 S
+
= 582
– 0.3 –
S = no light S = no light
Gaussian fit
Gaussian fit
0.1 0.2 Exponential fit
Exponential fit
0.05 0.1
0 0
570 580 590 600 610 570 580 590 600 610
Light wavelength (nm) Light wavelength (nm)
Hanson (1959) Thomas & Bistey (1964)

+ +
S = 550 Gaussian fit S = 550 Gaussian fit
– 0.75 –
0.5 S = 560 Exponential fit S = no light Exponential fit
Relative responding
0.4 0.6
0.3 0.45
0.2 0.3
0.1 0.15
0 0
460 480 500 520 540 560 580 600 620 640 500 520 540 560 580 600
Light wavelength (nm) Light wavelength (nm)
Blough (1972) Nallan et al. (1979)

+ 0.4 +
S = 480 S = 45
– –
S = no light S = 15
0.3
0.05 0.2
0.1
0 0
450 460 470 480 490 500 510 0 20 40 60 80 100
Light wavelength (nm) Line orientation (degrees from vertical)
Figure 1. Examples of rearrangement generalization gradients, with Gaussian and exponential fits where appropriate. Sources are given above
each graph.
(range 48–94%). Observations of asymmetrical gradients versa. Considering the sense organs can help us under-
also exist for humans. For instance, Kalish (1958), stand why a gradient is symmetrical on a given scale. For
Thomas & Mitchell (1962) and Thomas & Bistey (1964), instance, in the case of sound, gradients appear most
reported marked asymmetries in every one of 12 groups symmetrical on a logarithmic frequency scale, in keeping
of 20 subjects each generalizing along the dimension of with physiological evidence that sound frequency must
light wavelength. Some gradients even show strong change exponentially to yield changes in activation pat-
response biases, despite resulting from interdimensional terns of constant magnitude in the ear (Kandel et al.
tests (see e.g. the data from Thomas & Bistey 1964 in 1991; Coren et al. 1999). For many dimensions de facto
Fig. 1). Furthermore, Thomas & Bistey (1964) reported standards about which scale to use have emerged
increasing asymmetry as more of the test dimension was that agree with these considerations based on sensory
sampled (symmetrically around the S + ), a challenging physiology. We return to this issue below.
result for theories of generalization (analysis of variance
in the original study: F4,95 =11.00, P<0.001).
Response Biases
Obviously, gradient symmetry depends on the scale
chosen along a dimension. Symmetry on a linear scale After training a discrimination between two stimuli
will be destroyed passing to a logarithmic one, and vice differing along the test dimension (intradimensional
REVIEWS 19
Table 2. Examples of generalization along rearrangement dimensions

Dimension Response bias Species Source
Light spectra
Monochromatic light No data Human Kalish 1958
Yes Pigeon, Columba livia Hanson 1959
Yes Goldfish, Carassius auratus Ames & Yarczower 1965
Nonmonochromatic light* Yes Goldfish Ohinata 1978
Colour of female dummy* Yes Glow-worm, Lampyris noctiluca Schaller & Schwalb 1961
Yes Glow-worm, Phausis splendidula Schaller & Schwalb 1961
No Butterfly, Argynnis paphia Magnus 1958
Colour of egg dummy* Yes Herring gull, Larus argentatus Baerends 1982
Orientation of
Line Yes Pigeon Bloomfield 1967
Head stripe Yes Fish, Haplochromis burtoni Heiligenberg et al. 1972
Rocket picture Yes Human children Nicholson & Gray 1971
Sound frequency
No data Goldfish Fay 1970
Yes Rat, Rattus r. norvegicus Brennan & Riccio 1972
Yes Human Baron 1973; Galizio 1985
Yes Pigeon Klein & Rilling 1974
Location in space Yes Pigeon Cheng et al. 1997
No data Honeybee, Apis mellifera Cheng 1999, 2000
For laboratory studies, the column ‘Response bias’ refers to intradimensional tests. See the main text for
information about biases in other conditions.
*Nonmonochromatic lights cannot be meaningfully aligned along a single dimension, and intensity effects may be
caused by different sensitivity of the receptors to different wavelengths of light.
Difference in % variance accounted for
100
Variance explained by fitted exponential curve (%)
10
95
r96 = –0.55
90 –9
–10 P < 10
0 5 10 15 20
85 Most common separation between test stimuli (nm)
Figure 3. Differences in percentage of variance accounted for
between Gaussian and exponential fits to light wavelength general-
ization data in pigeons, as a function of separation between test
stimuli (see Test 2, Appendix 1, for data sources and statistics).
80 Similar results are obtained when considering average separation
80 85 90 95 100 between test stimuli rather than the most common one. Note that
Variance explained by fitted Gaussian curve (%) most studies adopt a sampling step of about 10 nm, where a large
Figure 2. Comparison of Gaussian and exponential fits with empiri- variation is observed.
cal generalization gradients. Points above (below) the diagonal
represent empirical gradients better fitted by exponential (Gaussian)
curves. little attention. Sometimes a ‘negative’ bias is also
observed, that is, lower responding than to S to stimuli
that are further away from S + . This effect is apparent
tests), authors have often found response biases (Table 2). when S elicits a considerable number of responses even
That is, there exist stimuli that elicit stronger responding after discrimination training (Fig. 5b; Stevenson 1966;
than S + (Fig. 5a). These stimuli are, almost invariably, Wills & Mackintosh 1998). Response biases along rear-
located further away from S . Recall that response biases rangement dimensions also occur in nature, as reported
can also appear in interdimensional tests (e.g. Kalish in the ethological literature about ‘supernormal stimuli’
1958; Thomas & Bistey 1964), a finding that has received (Tinbergen 1951; Eibl-Eibesfeldt 1975; cf. Table 2 and
40 0.4
(a) (a)
Proportion of responses
Average = 0.81
30 0.3
20 0.2 +
S
10 0.1
–
Number of studies
S
0 0
0.5 0.75 1 1.25 500 520 540 560 580 600
Observed height/height of fitted exponential Wavelength (nm)
40
(b)
Responses in first minute of test

(b)
Average = 1.01 160
30
120
20
+
80 S
–
S
10
40
0
0.5 0.75 1 1.25 0
Observed height/height of fitted Gaussian 2 4 6 8 10 12 14 16 18 20
Stimulus number
Figure 4. Predictions about gradient height by (a) exponential and
(b) Gaussian fits to rearrangement gradients, compared with the Figure 5. Examples of response biases along rearrangement dimen-
observed height. Data sources in Test 1, Appendix 1. sions. (a) Hanson’s (1959) classical study. Pigeons were trained to
peck a key for food when it was lit with a 550-nm light, but not
when the wavelength was 570 nm. In a subsequent generalization
Generalization of inherited and learned behaviour, test the maximum number of responses occurred at the 540-nm
light. (b) A gradient showing response biases both left of S + and
below).
right of S − , from Guttman (1965). He first trained pigeons to peck at
In field studies it is often difficult to know what all the wavelengths to be tested, then introduced a discrimination
experience an animal has had. Under laboratory con- training by presenting only S + (still reinforced) and S − (unrein-
ditions, however, there are a few well-established facts forced). The following generalization tests reveal that some stimuli
about the effects of previous experience on biases. A to the right of S − are reacted to less than S − .
general finding is that both the strength of the bias
(maximum responding relative to responding to S + ) and
the distance of the most effective stimulus from S + Conclusions
increase when the S + and S come closer. In Fig. 6 we
show these effects in the pigeon, along the light wave- (1) Generalization gradients along rearrangement
length dimension. The pattern is the same in all studies dimensions are described better by Gaussian than by
where the separation between S + and S has been varied. exponential functions (Test 1, Fig. 4).
Examples are Hearst (1968, line-tilt, pigeon), Ohinata (2) When the rearrangement dimension includes one
(1978, wavelength, goldfish), Baron (1973, tone fre- S + but no S , the gradient typically peaks at S + (but not
quency, humans) and Cheng et al. (1997, spatial location, invariably, Fig. 1), and the gradient is typically symmetri-
pigeon). cal around S + ; however, reproducible asymmetries exist,
The variation in Fig. 6 reveals that similarity between and may be more common than usually assumed (Tests
test stimuli is not the sole determinant of response biases. 4, 5).
There are ample indications that training and testing (3) When the rearrangement dimension includes both
procedures, as well as the characteristics of stimuli used, one S + and one S , responding is biased: gradients
are important factors (Purtle 1973; Mackintosh 1974). For typically peak at a stimulus that is further away from S
instance, the so-called ‘errorless’ discrimination training, than S + (Fig. 5); and closer S + and S produce gradients
where the intensity of the S is increased gradually, does whose peak is both higher and further away from S +
not seem to produce response biases (Terrace 1964, 1966). (Fig. 6).
An example of how testing can affect generalization is the
finding that biases tend to recede when the test phase is INTENSITY DIMENSIONS
very long (Crawford et al. 1980; Cheng et al. 1997). This
may be the effect of the subjects learning that the test It has long been noted that, in contrast to rearrange-
stimuli are not reinforced (see also Blough 1975). ment dimensions, intensity dimensions yield strongly
REVIEWS 21
Maximum responding, relative to S +
Proportion of responses (S + = 1)
(a)
(a) 1.5
300% r35 = –0.47
P < 0.005
1
Whistles
200% Bells
Lights
0.5
100% –3 –2 –1 0 1 2 3
0 5 10 15 20 25 30 35 40 +
Distance from S (arbitrary units)
20
(b) (b)
40
Peak displacement (nm)
Percentage of responses
r35 = –0.48
15
P < 0.005
30
10
20
5
10
0 0
0 5 10 15 20 25 30 35 40 70 75 S1 80 85 S2 90 95
+ –
Distance between S and S (nm) Intensity of noise (dB)
Figure 6. Dependence of response biases on the difference between Figure 7. (a) Intensity generalization in dogs conditioned to salivate
S + and S − . Data come from studies in which pigeons were trained to to a light, whistle or bell (position 0 on the horizontal axis).
discriminate between two monochromatic lights (sources in Test 6, Responding is expressed in terms of proportion of responses relative
Appendix 1). (a) Ratio between peak responding and responding to to the S + . Data from Razran (1949), summarizing over 250 studies
the positive stimulus. (b) Displacement of gradient peak from the S + . from Pavlov’s laboratory. (b) Intensity generalization in rats follow-
Peak responding, peak position and responding to S + are estimated ing training with two noise intensities, S1 =78 dB and S2 =87 dB. The
from fitted Gaussians. gradient is reversed when the weaker stimulus is the positive one
(data from Huff et al. 1975).
asymmetrical gradients and strong response biases

(Mackintosh 1974; Fig. 7, Table 3). This was once attrib- Returning to gradient shape, if we consider ‘monotonic’ a
uted to an ‘energizing’ effect of intensity on behaviour gradient that does not drop past S + (see below), we find
(Hull 1949). However, it is not intensity per se that 16 out of 38 monotonic intensity gradients, but none in
controls responding, since gradients that are higher at 93 intradimensional rearrangement gradients (Test 7 (2),
lower intensities can follow from experiences with a faint P<10 6).
S + and an intense S (Fig. 7b; Pierrel & Sherman 1960; Theories of generalization must thus explain a number
Zielinski & Jakubowska 1977). of differences between intensity and rearrangement
The shape of intensity generalization gradients has generalization. The only models that so far have met with
been a debated subject. Claims that variations in intensity some success (reviewed by Ghirlanda 2002) predict
produce monotonic gradients (i.e. ever-increasing or ever- monotonic intensity gradients ‘under ideal conditions’
decreasing) have been contrasted with findings of non- (that is, in idealized generalization experiments probing a
monotonic gradients (Ernst et al. 1971; Thomas & Setzer memory modified by experiences with only two stimuli),
1972). It is clear that intensity gradients cannot grow or and may be able to explain nonmonotonic intensity
fall forever; for instance, intensities that are too high will gradients by appealing to determinants of gradient shape
harm the sense organs, and those that are too low cannot other than stimulus dimension (S. Ghirlanda, unpub-
be detected. However, important differences between lished data). In the following we discuss three such
intensity and nonintensity gradients exist independent of determinants.
statements about gradient shape. For instance, in 30 of 38 First, sampling errors lead to underestimates of the
intensity gradients surveyed, responding beyond S + does number of monotonic intensity gradients. An inaccurate
not drop below S + levels, even when the gradient is not estimate of responding at a single point can transform a
monotonic (intradimensional rearrangement gradients: 2 monotonic curve into a nonmonotonic one, but it is less
of 93, Test 7 (1), P<10 6). Furthermore, a quantitative likely that the reverse occurs. The above definition of
analysis shows that response biases along intensity ‘monotonicity’, considering only the S + and stimuli
dimensions are significantly stronger than along further away from S , reduces but does not eliminate the
rearrangement dimensions (Test 8, P<10 6; Fig. 8). effects of sampling. Indeed, monotonicity of intensity
Table 3. Examples of generalization along intensity dimensions

Intensity dimension Monotonicity Response bias Species Source
Sound
Tone Yes Yes Rat, R. norvegicus Pierrel & Sherman 1960
Yes/no* Yes Rat Thomas & Setzer 1972
Yes/no* Yes Guinea pig, Cavia porcellus Thomas & Setzer 1972
Yes Yes Rat Brennan & Riccio 1973
Yes Yes Rabbit, Oryctolagus cuniculus Scavio & Gormezano 1974
Bell Yes Yes Dog, Canis familiaris Razran 1949
Whistle Yes Yes Dog Razran 1949
White noise Yes Yes Rat Huff et al. 1975
Yes Yes Rat Zielinski & Jakubowska 1977
Light
White light Yes Yes Dog Razran 1949
Yes/no* Yes Pigeon, C. livia Ernst et al. 1971
No Yes Pigeon Lawrence 1973
Yes Yes Rat Brown 1942
No data Yes Earthworm, Lumbricus terrestris Gilpin et al. 1978
Brightness of egg dummy Yes Yes Herring gull, L. argentatus Baerends 1982
Brightness of female dummy Yes Yes Butterfly, Eumenis semele Tinbergen et al. 1942
Yes Yes Glow-worm, P. splendidula Schaller & Schwalb 1961
No No Glow-worm, L. noctiluca Schaller & Schwalb 1961
Chemical concentration
Odour No data Yes Bee, A. mellifera Bhagavan & Smith 1997
Taste No data Yes Rat Tapper & Halpern 1968
*Both monotonic and nonmonotonic gradients were found.
1 1500
Intensity (N = 39) Sessions 1 and 2
Sessions 3 and 4
Number of responses
Proportion of studies
0.8 Rearrangement (N = 93) Sessions 5 and 6

1000
0.6
0.4
500
0.2
0 0
100–200 200–300 300–400 400 or more 60 70– 80 90+ 100
Maximum responding, relative to S+ (%) Sound intensity (dB)
Figure 8. Distribution of response biases along intensity and Figure 9. Data on sound intensity generalization in rats (Pierrel &
rearrangement dimensions. The strength of bias is measured as the Sherman 1960), collected over 6 days of testing in extinction.
ratio of observed maximum responding to responding to S + . Note Gradients from test days 1 and 2 are both monotonic (their average
that strength of bias is underestimated along intensity dimensions, is shown in the figure); gradients from subsequent days are not.
since responding often does not start decreasing within the probed
stimulus range. Data sources in Test 8, Appendix 1.
Third, laboratory data are often analysed by taking
gradients increases with better sampling, that is, with into account only experimentally controlled stimuli.
larger experimental groups (Test 9, P<0.01). However, responding to both the low and high ends of
Second, a number of studies reporting nonmonotonic intensity continua is likely to be influenced by factors
intensity generalization were designed to explore the beyond experimental control. Stimuli of high intensity,
effects of very long test sessions (e.g. Newlin et al. 1979; for example very loud sounds or very bright lights, are
Thomas et al. 1991, 1992, excluded from the analyses often avoided by animals. Similarly, stimuli of very low
above; this research is summarized in Thomas 1993). intensity (silence, a dark response key) are usually not
Under these conditions, gradients can change from reacted to. In addition to such generic reactions, specific
monotonic to peaked during the test (Fig. 9). Such a responses may interfere. For instance, Baerends (1982)
shape change appear analogous to the disappearing or reported that lightly coloured egg dummies are preferen-
waning of response biases in the course of long test tially retrieved by herring gulls, Larus argentatus, but this
sessions, along rearrangement dimensions (e.g. Crawford preference does not extend to white dummies. Studies
et al. 1980, Cheng et al. 1997). with related species have shown that gulls usually remove
REVIEWS 23
Table 4. Examples of generalization along size dimensions

Response
Size dimension bias Species Source
Circle No data Rat, R. r. norvegicus Grice & Saltz 1950

No data Pigeon, C. livia Jenkins et al. 1958
No data Toad, Bufo bufo Ewert 1980
Yes Horse, Equus caballus Dougherty & Lewis 1991
Yes Butterfly, E. semele Tinbergen et al. 1942
Yes Human Berlyne 1950
Square No data Toad Ewert 1980
Rectangle No data Human Rosenbaum 1953
Egg Yes Herring gull, L. argentatus Baerends 1982
Yes Black-headed gull, L. ridibundus Baerends et al. 1975
Yes Oystercatcher, Haematopus ostralegus Tinbergen 1951
Yes Ringed plover, Charadrius hiaticula Koehler & Zagarus 1937
Tail Yes Widowbird, Euplectes progne Andersson 1982
Yes Swordtail fish, Xiphophorus helleri Basolo 1990; Basolo & Delaney
2001
Food item Yes Newborn chicks, Gallus g. domesticus Gamberale & Sillen-Tullberg 1998;
Gamberale-Stille 2000
Female dummy Yes Butterfly, Argynnis paphia Magnus 1958
Female lantern Yes Glow-worm, P. splendidula Schaller & Schwalb 1961
No Glow-worm, L. noctiluca Schaller & Schwalb 1961
white objects from the nest (a likely antipredatory bigger sizes. This is because the intensity and rearrange-
defence; Tinbergen et al. 1962; Baerends 1982). ment components have contrasting effects for bigger
stimuli (stimulating more receptors, but in a different
pattern), but work together in reducing responding to
Conclusions
smaller stimuli (stimulating fewer receptors and in a
(1) Gradients obtained along intensity dimensions different pattern). This prediction is confirmed by avail-
show larger response biases than rearrangement gradients able data (Test 10, P<0.05). The same data suggest that
(Fig. 8). size gradients are described better by Gaussian than by
(2) Many intensity gradients are monotonic (rather exponential curves (Test 11, P=0). Both conclusions
than peaked) over large ranges of intensity; more specifi- should be viewed as tentative in light of the small
cally: responding increases with intensity when S + is number of studies examined (N=7 and N=8, respect-
more intense than S (including when S is S + ‘turned ively). Another element in support of a rearrangement/
off’, e.g. a dark versus an illuminated key); and respond- intensity analysis of size dimensions is that size gradients
ing decreases with intensity when S + is less intense than show larger response biases than rearrangement gradients
S. (Test 8, P<0.02). In the small sample collected, biases
(3) Observed departures from monotonicity can, at towards bigger sizes appear comparable with biases along
least in some cases, be ascribed to: errors in sampling the intensity dimensions (Test 8, NS).
gradient (Test 9); long test sessions, leading to changes in Table 4 gives examples of response biases along size
gradient shape (Fig. 9); and pre-existing reactions (both dimensions. Similar regularities as those reported above
inherited and learned) to very intense stimuli (often for rearrangement dimensions seem to apply. For
avoided) or very weak ones (often ignored). instance, Weinberg (1973) found a stronger bias when S +
and S were closer in size.
VARIATION IN SIZE AND OTHER DIMENSIONS If a simple consideration of the arrangement and inten-
sity of stimulation is helpful in analysing size dimensions,
Along size dimensions, both the amount of stimulation it is not so for many other dimensions (Table 5). In some
and its arrangement on the sense organs vary. Consider a cases, we do not know enough about the sense organs
familiar stimulus of a given size. A bigger stimulus will act underlying perception along some dimensions, for
on more sensory cells than the familiar one, providing example floor tilt (Lyons et al. 1973) or arm movement
more stimulation. On the other hand, it will also provide (Hedges 1983; Dickinson & Hedges 1986). In other cases,
a different arrangement of stimulation. We can try to for example complex variations in shape (Ferraro &
understand size gradients as a trade-off between these two Grisham 1972; Wasserman et al. 1996), distinguishing
components. One immediate consequence is that the between intensity and nonintensity effects is simply not
gradient may be asymmetrical, higher on the side of sufficient. When we lack information about underlying
Table 5. Examples of generalization along dimensions that cannot be classified with the rearrangement/intensity
scheme
Response
Dimension bias Species Source
Visual shape
Female dummy Yes Butterfly, A. paphia Magnus 1958
Polygon Yes Pigeon, C. livia Ferraro & Grisham 1972
Egg No Herring gull, L. argentatus Baerends 1982
Visual contrast Yes Herring gull Baerends 1982
Yes Chicks, Gallus g. domesticus Osorio et al. 1999
Length of movement Yes Human Hedges 1983
No data Human Dickinson & Hedges 1986
Click rate Yes Pigeon Farthing & Hearst 1972
Yes Rat, R. r. norvegicus Weiss & Schindler 1981
Flicker rate Yes Pigeon Sloane 1964
Yes Butterfly, A. paphia Magnus 1958
Floor tilt Yes Pigeon Lyons et al. 1973
Yes Pigeon Riccio et al. 1966
Calls/songs Yes Monkey, Callimico goeldii Masataka 1983
Yes Blackbird, Turdus merula Wolffgramm & Todt 1982
Human faces Yes Human Rhodes 1996; Rhodes &
Zebrowitz 2002
Yes Chickens Ghirlanda et al. 2002
Checkerboard patterns Yes Human McLaren et al. 1995
Icon sets Yes Human Wills & Mackintosh 1998
Drawings of rotated objects No data Pigeon Wasserman et al. 1996
‘Aggressiveness’ of verbal stimuli No data Human Buss 1961; 1962
‘Fearfulness’ of snake pictures No data Human Buss et al. 1968
sensory processes, we can try to infer some characteristics (2) Response biases have been found along all
of such processes by analysing the experimental data in dimensions investigated so far.
the light of what we know from other, better studied (3) Generalization along size dimensions is influenced
dimensions. For example, Dickinson & Hedges (1986) let by both intensity and rearrangement effects: size gener-
blindfolded humans move a sliding handle for a given alization gradients are typically peaked; they are better
distance, and tested them with movements of different approximated by Gaussian than by exponential functions
lengths by asking if they matched the training one. Their (Test 11); they exhibit larger biases than gradients along
results seem to suggest that length of movement is per- rearrangement dimensions, and comparable to intensity
ceived as a rearrangement dimension. Monotonic gradi- gradients (Test 8); and when S is the absence of S + ,
ents have been reported along dimensions that cannot be responding is biased towards bigger sizes (Test 10).
readily identified as intensity ones (Ghirlanda & Enquist,
1999); for instance rate of stimulus presentation (Magnus
1958; Weiss & Schindler 1981) or femininity/masculinity FACTORS AFFECTING THE AMOUNT OF
of human faces (Enquist et al. 2002b). The case of changes GENERALIZATION
in shape is particularly interesting. For instance, Magnus A fundamental question is what regulates the amount of
(1958), studying Argynnis paphia butterflies, found that generalization along a dimension. The width of a peaked
certain shapes attract males more than female-shaped gradient provides a measure of the amount of general-
dummies, but also that males respond very little to other ization, and one factor related to gradient width is
shapes (Fig. 10). Data of this kind can be used to explore discriminability. Guttman & Kalish (1956) noted that
how similarity is perceived across species. generalization is measured by the change in behaviour
arising from a change in stimulation, whereas discrimi-
Conclusions nability is defined as the change in stimulation necessary
to yield a behavioural change (cf. Lashley & Wade 1946).
(1) Both peaked and monotonic gradients have been They thus suggested that generalization and discrimi-
found along dimensions where both the intensity and the nability should be inversely related, but because they
arrangement of stimulation vary. lacked reliable data on discriminability they failed to
REVIEWS 25
1969) to about 27 nm (Thomas & King 1959), indicating

that choice of stimuli and details of experience have a
large impact on amount of generalization. Adding nega-
tive stimuli along the test dimensions tends to make
gradients narrower. For instance, pigeons generalize less
after discrimination between two wavelengths than after
discrimination between one wavelength and a dark
stimulus (Test 12, P<0.001). This effect has also been
observed in goldfish, along the wavelength dimension
(Ames & Yarczower 1965; Ohinata 1978), and in humans,
along the dimension of sound frequency (Baron 1973).
So far we have mainly discussed generalization after
experiences with only one S + and one S . In reality, a
particular external condition, such as presence of food,
produces many different stimuli on the sense organs
(many positive stimuli for feeding behaviour). For
example, food items differ in appearance and a single
food item gives rise to different receptor activations
because of variations in distance, light conditions, orien-
tation and so on. In such situations there is variation
within as well as between the sets of positive and negative
stimuli. Despite the obvious biological importance of
the subject, our knowledge of generalization following
experiences with many stimuli is unsystematic. Relevant
data come from a variety of sources such as studies of
0 0.1 0.2 0.3 0.4 summation, generalization, memory and perceptual con-
Proportion of responses stancy (Walsh & Kulikowski 1998). In Fig. 11 we show
some data from studies of generalization. When two or
Figure 10. Reactions of male Argynnis paphia butterflies to female
more positive stimuli are close enough we may observe
models of different shapes (data from Magnus 1958). The effective-
ness of each model is expressed as the proportion of flights towards
stronger reactions to intermediate stimuli (Fig. 11a), a
it on a total of 135 flights, in an open area where all models were result found also in studies of summation and stimulus
visible simultaneously. All models had an area of approximately compounds (Wagner 1971; Weiss 1972; Kehoe 1986) and
22 cm2 and were of the same colour (orange-red). in evaluations of the prototype model of memory (Rosch
et al. 1976; Homa et al. 1981; Shanks 1995). When the
positive stimuli are more distant the intermediate ones
observe this relation studying light wavelength general- may result in less responding (Fig. 11a). This may be the
ization in the pigeon. Later work has confirmed that same as the failure of revealing prototype effects in some
pigeons generalize more in regions of the light spectrum memory studies (Lea & Harrison 1978; Pearce 1987;
where discriminability is poorer (reviewed in Honig & Watanabe 1988). A fundamental question, for both
Urcuioli 1981). The same conclusion was reached by behaviour and evolution, is whether response biases are
Kalish (1958) in a study of wavelength generalization in found after the same amount of experience with many
humans. Similarly, Fay (1970) showed that width of stimuli (Enquist & Johnstone 1997; Enquist & Arak 1998).
sound frequency generalization gradients in goldfish Flat gradients (Fig. 11c) and small (Fig. 11b) and large
increases approximately linearly with the frequency of biases (Fig. 11d) have all been reported. The strongest
the S + , in line with the decreasing discriminability along biases seem to appear along intensity dimensions (Fig.
this dimension (Hawkins 1981). Furthermore, discrimi- 11d, from Scavio & Gormezano 1974; see also Kessen
nability can change differently on the two sides of the S + , 1953; Bass 1958; Murray & Kohfeld 1965; Birkimer &
thus affecting gradient symmetry (Blough 1972; Honig & James 1967; Blue et al. 1971).
Urcuioli 1981). We stress that discriminability and gener-
alization are both behavioural measures, and thus one
Conclusions
cannot explain the other. Rather, we should understand
how both emerge from characteristics of nervous systems (1) Gradient width and symmetry can be related to
such as the response properties of receptors and how discriminability along the test dimension: finer dis-
signals from receptors are processed. criminability is correlated with narrower gradients; and
A second factor influencing the amount of general- when discriminability is different on the two sides of S +
ization is the experiences an animal had along the test asymmetrical generalization gradients are observed.
dimension. Gradients obtained from interdimensional (2) Along rearrangement dimensions, discriminations
tests along the same dimension can vary considerably in between stimuli leads to decreased generalization
width. An analysis of light wavelength generalization in (narrower gradients).
the pigeon, for instance, shows that the standard devi- (3) Along rearrangement dimensions, training on sev-
ation of Gaussian fits varies from about 2.5 nm (Blough eral or many equally rewarded S + increases generalization
0.2 0.2
(a) (b)
0.15 0.15
0.1 0.1
0.05 0.05
Proportion of responses
0 0
–20 –10 0 10 20 500 520 540 560 580
Deviation from middle wavelength (nm) Wavelength (nm)
0.1 0.5
(c) (d)
0.4
0.3
0.05
0.2
0.1
0 0
4 8 12 16 65 72 79 86
Stimulus number Tone intensity (dB)
Figure 11. Four examples of generalization after experience with many stimuli. x: Training stimuli; C: test stimuli. Subjects had the same
experience with all training stimuli. (a) Two-stimulus training can yield gradients with both one or two peaks depending on the difference
between stimuli (data from Blough 1969, pigeon, wavelength). (b) Three-stimulus training can produce less responding to the intermediate
stimulus (data from Kalish & Guttman 1959, pigeon, wavelength). (c) In this case no response biases are found after extensive experience with
a set of contiguous stimuli (data from Guttman 1965, pigeon, wavelength). (d) Responding can be stronger to more intense stimuli, even
when all stimuli were followed by the same consequences during training (data from Scavio & Gormezano 1974, rabbit, tone intensity).
(broader gradients): if the S + ’s are close to each other, tions of generalization typically make direct reference to
a flat or almost flat gradient develops over the range individual learning (Kalish 1969). These attitudes prob-
covered by the training stimuli; departing from the S + on ably stem from early ideas within the two disciplines:
both sides, the gradient falls down (Fig. 11a, b, c); the classical psychologists often ignored innate determinants
gradient shows multiple peaks if the distance between of behaviour (Watson 1924), and early ethologists
training stimuli is sufficiently increased (Fig. 11a); and claimed that inherited and learned behaviour are
responding to all training stimuli is the same or similar governed by different mechanisms (Von Uexkull 1928;
(small biases, Fig. 11b, c). Lorenz 1937).
(4) Along intensity dimensions, substantial biases in Although such rigid ideas have been abandoned
responding persist even when two or more stimuli are (Hogan & Bolhuis 1994; Bolhuis & Hogan 1999), the idea
equally rewarded. The more intense S + ’s elicit stronger that inherited and learned behaviour generalize differ-
reactions (Fig. 11d; Kessen 1953; Bass 1958; Murray & ently seems to have survived longer (Baerends & Krujit
Kohfeld 1965; Birkimer & James 1967; Blue et al. 1971). 1973; Dawkins & Guilford 1995). In particular, it has
been claimed that innate behaviour results in ‘open-
GENERALIZATION OF INHERITED AND LEARNED ended’ generalization (i.e. monotonic gradients), while
BEHAVIOUR individual learning does not (Baerends & Krujit 1973;
Hogan et al. 1975; Lorenz 1981). The data do not support
Both genetically inherited and individually learned this statement. Note first that ethological studies of super-
responses generalize to novel stimuli (Tables 2–5). Within normality are not always about behaviour that is inde-
ethology, however, there has been a tendency to separate pendent of individual learning (e.g. egg retrieval in gulls,
the study of innate and learned generalization. For Baerends 1982). Furthermore, some dimensions cannot
instance, ethologists have often claimed that super- by their nature result in open-ended generalization,
normality and peak shift are distinct phenomena, at the regardless of ontogeny of behaviour. For instance, the
same time that similarities have been acknowledged response of male Haplochromis burtoni cichlids to the
(Baerends & Krujit 1973; Hogan et al. 1975; Staddon orientation of the head stripe (Heiligenberg et al. 1972)
1975; Dawkins & Guilford 1995). Within psychology, the can only come back to its original value after the stripe
question is seldom addressed explicitly; however, defini- has turned a full circle.
REVIEWS 27
Pure rearrangement dimensions are rare in ethological rearrangement of stimulation appears valid for both
studies, as test stimuli most often vary in complex ways. inherited and learned behaviour; and the claim that
One exception is the just cited study by Heiligenberg biases in inherited behaviour are open ended (mono-
et al. (1972), where the supernormal effect of a rotated tonic) whereas biases in learned behaviour are limited is
head stripe decreases after only a 90 rotation. In other unsupported.
cases a rearrangement dimension can be defined by vary-
ing only one characteristic of a complex stimulus. For DISCUSSION
instance, many of the stimuli in Fig. 10 are rectangles of
different length and constant area. It is clear that the Empirical data gathered in about 100 years of research
preferences of A. paphia males along this rearrangement establish generalization as a fundamental behavioural
dimension are not open ended. The case of colour is more phenomenon, whose basic characteristics appear univer-
complex. First, different light spectra of the same physical sal. Birds and mammals are most studied, but fish, insects,
intensity can elicit different amounts of activation in amphibians and reptiles generalize in the same ways. It
receptors (and differences between species exist). Second, seems to matter little, for generalization, whether a
physical intensity is seldom controlled for in ethological behaviour has been acquired phylogenetically or through
studies. Anyway, ethologists report that dummies of individual learning, and the nature of sensory con-
unnatural colour can be both more and less effective in tinua is an important determinant of gradient shape.
eliciting an innate behaviour (Magnus 1958; Schaller & Furthermore, generalization seems little dependent on
Schwalb 1961). In some cases the most effective colour is the context in which a given behaviour is used. That is, if
clearly not at the extremes of the spectrum (L. noctiluca a discrimination between green and red is established,
glow-worms: Schaller & Schwalb 1961; herring gulls: generalization to other colours will follow independent of
Baerends 1982) or closely matches the natural colour whether the discrimination is about food items or poten-
(A. paphia butterflies: Magnus 1958). tial partners (e.g. Ghirlanda et al. 2002), or on whether
Open-ended generalization of innate behaviour has behaviour is performed by children to obtain ‘points’ or
been reported almost exclusively along intensity dimen- by pigeons pecking for food. The generality of the
sions, for instance by Tinbergen et al. (1942, brightness of findings reviewed suggests that generalization arises
female butterfly dummy), and Schaller & Schwalb (1961, from basic and universal characteristics of behaviour
brightness of glow-worm female dummy), with a few mechanisms (cf. Hogan 1994).
exceptions along size dimensions (where strong biases are This review has focused on empirical findings, but we
expected, see above). For example, Baerends (1982) end with some theoretical considerations. A number of
showed that oversized eggs are preferred by incubating attempts have been made to understand the causes of
gulls up to giant sizes. Magnus (1958) provided mixed generalization (reviewed in Kalish 1969; Mackintosh
evidence, showing that male A. paphia butterflies prefer 1974). Theorizing about mechanisms has considered
four-fold enlarged female dummies to dummies twice the properties of stimuli, sense organs and neural processing,
normal size when the dummies are stationary, but not and how these factors interact. Physical similarity
when they imitate flight. Similarly, Schaller & Schwalb between stimuli is one cause of generalization. Stimuli
(1961) reported that in glow-worms, P. splendidula males may be similar because they share common components,
prefer dummy female lanterns four times bigger than and generalization may follow because novel stimuli
normal, but L. noctiluca males do not. There are further include components also present in familiar stimuli (e.g.
reports that the supernormal effect of bigger stimuli often Thorndike 1911; Guthrie 1930, 1935; Blough 1975;
quickly ceases. One instance is the response of male Rescorla 1976). However, not all stimuli are made up of
P. splendidula and L. noctiluca to circular lights, which ‘components’ in this sense (e.g. light and sound spectra).
declines for circles several times bigger than the female’s In general, what is similar and different to an organism
lantern (Schaller & Schwalb 1961). The same authors depends also on properties of receptors and the organiz-
showed that L. noctiluca males prefer a dummy lantern ation of sense organs (including early processing of neural
containing three horizontal segments to one with six signals within sense organs). These factors determine how
(two is normal). Koehler & Zagarus (1937) found that physical similarity translates into similarity of nervous
ringed plovers, Charadrius hiaticula, retrieve eggs weigh- signals to the brain, and will thus contribute to general-
ing 17 g, but not those above 35 g (normal eggs weigh ization. Receptors and sense organs have often been
about 11.5 g). In this last case individual learning may ignored, especially in contemporary psychological
play a role (cf. Baerends 1982). Finally, Ewert (1980) has models (but not always in early ones, Hull 1943;
shown that both naïve and experienced toads, Bufo bufo, Schlosberg & Solomon 1943; Hebb 1949). By considering
clearly prefer catching objects within a restricted size them it may be possible to account for both rearrange-
range. ment generalization and intensity generalization within
the same model, by recognizing that similarity depends
upon which receptors are stimulated and to what degree
Conclusions
(Ghirlanda 2002).
We could find no difference between generalization of Generalization is also modulated centrally in the
genetically inherited and individually learned behaviour, nervous system. Suggestion about how this occurs vary in
with respect to either gradient shape or response biases: detail, but the core idea is that processing of stimuli that
the distinction between the effects of intensity and are distinctly different can rely, at least to some extent, on
28
ANIMAL BEHAVIOUR, 66, 1
Table 6. Basic predictions by models of generalization

Rearrangement generalization Intensity generalization
Biases along
dimensions Monotonic Monotonic
Bell including increase decrease Stronger
Theory shape S + and S − (S + >S − ) (S + <S − ) biases Source
Early ideas suggesting overlap or interactions Consequences unclear or not studied Pavlov 1927; Hebb 1949; Horne 1965;
among nerve cells* Thompson 1965; Lorenz 1981; Baerends 1982
Gradient interaction, original formulation Assumed Yes† No No No Spence 1937; Hull 1943
Gradient interaction, later developments Assumed Yes† Assumed No Assumed Hull 1949; Perkins 1953; Logan 1954
Exponential generalization in ‘psychological No No No No No Shepard 1987; Cheng et al. 1997
space’§
Gaussian generalization in ‘psychological spaces’§ Assumed Yes No No May be assumed Shepard 1987
Configural theory Assumed Yes No No May be assumed Pearce 1987
Blough’s model, original formulation Assumed Yes No No May be assumed Blough 1975
Blough’s model, reinterpretation* Yes Yes Yes Yes Yes Ghirlanda & Enquist 1999; Ghirlanda 2002
Feed forward networks, with idealized inputs‡ No/assumed No/yes No No No Gluck 1991; Pearce 1994; S. Ghirlanda,
unpublished data
Feed forward networks, with realistic inputs* Yes Yes Yes Yes Yes Ghirlanda & Enquist 1998
Overlap theory, based on receptor activations* Yes Yes Yes Yes Yes Ghirlanda & Enquist 1999
Derivations for some of these predictions not in the original sources can be found in Ghirlanda & Enquist 1999; Ghirlanda 2002.
*Empirical knowledge about the nervous system is used.
†Responding underestimated.
‡See cited papers for details of assumed stimulus representations.
§For possible improvements, see Ennis (1988), Shepard (1988), Staddon & Reid (1990), Shepard (1990).
REVIEWS 29
the same nerve cells and connections (Pavlov 1927; Hebb these models provide a powerful explanation for how
1949; Horne 1965; Thompson 1965; Baerends & Krujit generalization is generated, including the consequences
1973; Blough 1975; Lorenz 1981). Distinct stimuli may of learning (Blough 1975; Ghirlanda & Enquist 1998;
thus come to elicit similar responses. Generalization of Ghirlanda 2002).
this kind is strongly dependent on experience (includ-
ing the species’ experience, coded in the genes). Often
generalization is substantial along dimensions with Acknowledgments
which the organism has little experience (Peterson 1962; We thank Susanne Stenius for valuable help in collecting
Rubel & Rosenthal 1975; Kerr et al. 1979). Along references. Ken Cheng, Björn Forkman, Birgitta Sillén-
familiar dimensions organisms generalize less: latent Tullberg and the anonymous referees provided helpful
learning, perceptual learning and discriminations comments on the manuscript. Laura, Giovanni and
between similar stimuli all decrease generalization Daniele Ghirlanda made the acquisition of many data
(Mackintosh et al. 1991; Bennett et al. 1994; see also sets possible by kindly allowing use of computer equip-
above). Discrimination learning, in particular, can sub- ment. S.G. thanks the Institute for the Science and
stantially lower generalization, presumably up to sensory Technologies of Cognition, National Research Council,
limits. Rome, for kind hospitality during some of the time spent
There is of course also a functional side to general- working on the manuscript. This work has been partly
ization. Evolution has favoured those behaviour mech- funded by Marianne och Marcus Wallenberg Stiftelse and
anisms that are ‘intelligent’ towards the real world. For by the Jubileumsfund of the Bank of Sweden.
instance, stimuli that are similar to one another often
share some causal relation with events in the outside
world. Animals detect and use such regularities, general- References
izing knowledge about familiar situations to novel ones.
In addition, animals use general methods to cope with Akins, F. R., Drew Gouvier, W. & Lyons, J. E. 1981. The effects of S +
and S − alternation on stimulus control and the peak shift. Journal
novelty, including exploratory and avoidance behaviour.
of General Psychology, 105, 101–114.
Observed similarities in generalization across taxa may
Ames, L. L. & Yarczower, M. 1965. Some effects of wave-
indicate that evolutionary strategies to deal with novelty length discrimination on stimulus generalization in the goldfish.
are limited. Note that, without generalization, learning Psychonomic Science, 3, 311–312.
itself would be seriously limited: by trying out different Andersson, M. 1982. Female choice selects for extreme tail length
responses to novel stimuli animals can adapt by learning, in a widowbird. Nature, 299, 818–820.
but without generalization possibly productive responses Arbib, M. A. (Ed.) 2003. The Handbook of Brain Theory and Neural
would never be tested. Whether everything about Networks. 2nd edn. Cambridge, Massachusetts: MIT Press.
generalization is functional is more uncertain (Enquist & Baerends, G. 1982. The hering gull and its egg. Part II. The
Arak 1998; Enquist et al. 2002a). Characteristics that seem responsiveness to egg-features. Behaviour, 82, 358–363.
difficult to explain as adaptive include, for instance, the Baerends, G. P. & Krujit, J. P. 1973. Stimulus selection. In:
two- or three-fold increase in responding along intensity Constraints on Learning (Ed. by R. A. Hinde & J. Stevenson-Hinde).
dimensions, and biases within sets of stimuli with the New York: Academic Press.
same consequences. Baerends, G., Beer, C. & Manning, A. (Eds) 1975. Function and
A number of models provide predictions about Evolution of Behaviour. Oxford: Clarendon Press.
generalization (Table 6). Gradient interaction models Baron, A. 1973. Postdiscrimination gradients of human subjects
predict responding after training on several stimuli by on a tone continuum. Journal of Experimental Psychology, 101,
combining gradients relative to each training stimulus 337–342.
(Spence 1937; Hull 1943; Kalish & Guttman 1957, 1959). Baron, M. R. & Harris, C. A. 1968. Generalization tests along a
The latter, however, are not predicted by the model but light intensity dimension following three S-R-reinforcement
assumed or obtained from experiments. Similar assump- contingencies. Psychological Record, 18, 261–266.
tions about generalization are also present in other Bartoshuk, A. K. 1964. Human neonatal cardiac responses to sound:
models (Pearce 1987; Shepard 1987). This incompleteness a power function. Psychonomic Science, 1, 151–152.
is avoided in some recent ‘artificial neural network’ Basolo, A. 1990. Female preference predates the evolution of the
sword in swordtail fish. Science, 250, 808–810.
models. Here gradient shapes emerge from the inter-
action of input patterns with model mechanisms that Basolo, A. L. & Delaney, K. 2001. Male biases for male character-
istics in females in Priapella olmecae and Xiphophorus helleri (Family
attempt to capture properties of real nervous systems.
Poeciliidae). Ethology, 107, 431–438.
Some earlier models, originally based on assumptions
Bass, B. 1958. Gradients in response percentage as indices of
about generalization, can also be reinterpreted as network
nonspatial generalization. Journal of Experimental Psychology, 56,
models taking into account actual perceptual processes 278–281.
(for instance, the model in Blough 1975; see Ghirlanda & Bennett, C. H., Wills, S. J., Wells, J. O. & Mackintosh, N. J. 1994.
Enquist 1999, Ghirlanda 2002). Artificial neural networks Reduced generalization following preexposure: latent inhibition
are promising to students of behaviour (Haykin 1999; of common elements or a difference in familiarity? Journal of
Arbib 2003). In combination with models of sense organs, Experimental Psychology: Animal Behavior Processes, 20, 232–239.
neural networks can potentially integrate all the factors Berlyne, D. E. 1950. Stimulus intensity and attention in relation to
discussed above: properties of stimuli, reception, neural learning theory. Quarterly Journal of Experimental Psychology, 2,
processing and learning. Preliminary results indicate that 71–75.
Bhagavan, S. & Smith, B. H. 1997. Olfactory conditioning in the Crawford, L. L., Steele, K. M. & Malone, J. C. 1980. Gradient form
honey bee, Apis mellifera: effects of odor intensity. Psychological and sequential effects during generalization testing in extinction.
Bullettin, 61, 107–117. Animal Learning and Behavior, 8, 245–252.
Birkimer, J. C. & James, J. P. 1967. Stimulus intensity dynamism Dawkins, M. & Guilford, T. 1995. An exaggerated preference for
and generalization of inhibition in a discriminated lever press simple neural network models of signal evolution? Proceedings of
avoidance situation. Psychonomic Science, 8, 377–378. the Royal Society of London, Series B, 261, 357–360.
Bloomfield, T. M. 1967. A peak shift on a line-tilt continuum. Dickinson, J. & Hedges, D. G. 1986. Adaptation level as an
Journal of the Experimental Analysis of Behavior, 10, 361–366. explanation of the peak shift in generalization with movement
Blough, D. 1969. Generalisation gradient shape and summation in stimuli. Journal of Mathematical Biology, 18, 101–110.
steady-state tests. Journal of the Experimental Analysis of Behavior, Dougherty, D. M. & Lewis, P. 1991. Stimulus generalization,
12, 91–104. discrimination learning, and peak shift in horses. Journal of the
Blough, D. S. 1975. Steady state data and a quantitative model of Experimental Analysis of Behavior, 56, 97–104.
operant generalization and discrimination. Journal of Experimental Dukhayyil, A. & Lyons, J. E. 1973. The effect of overtraining on
Psychology: Animal Behavior Processes, 104, 3–21. behavioral contrast and the peak-shift. Journal of the Experimental
Blough, P. M. 1972. Wavelength generalization and discrimination Analysis of Behavior, 20, 253–263.
in the pigeon. Perception and Psychophysics, 12, 342–348. Dysart, J., Marx, M., McLean, J. & Nelson, J. A. 1974. Peak shift as
Blue, S., Sherman, G. J. & Pierrel, R. 1971. Differential responding a function of multiple schedules of reinforcement. Journal of the
as a function of auditory stimulus intensity without differential Experimental Analysis of Behavior, 22, 463–470.
reinforcement. Journal of the Experimental Analysis of Behavior, 15, Eibl-Eibesfeldt, I. 1975. Ethology. The Biology of Behavior. New York:
371–377. Holt, Rinehart & Winston.
Bolhuis, J. J. & Hogan, J. A. (Eds) 1999. The Development of Animal Ennis, D. M. 1988. Confusable and discriminable stimuli: comment
Behavior: a Reader. Malden, Massachusetts: Blackwell. on Nosofksy (1986) and Shepard (1986). Journal of Experimental
Brennan, J. F. & Riccio, D. C. 1972. Stimulus control of shuttle Psychology: General, 117, 408–411.
avoidance in young and adult rats. Canadian Journal of Psychology, Enquist, M. & Arak, A. 1998. Neural representation and the
26, 361–373. evolution of signal form. In: Cognitive Ethology (Ed. by R. Dukas),
Brennan, J. F. & Riccio, D. C. 1973. Stimulus control of avoidance pp. 1–420. Chicago: Chicago University Press.
behavior in rats following differential or nondifferential pavlovian Enquist, M. & Johnstone, R. 1997. Generalization and the evolution
training along dimensions of the conditioned stimulus. Journal of of symmetry preferences. Proceedings of the Royal Society of
Comparative and Physiological Psychology, 85, 313–323. London, Series B, 264, 1345–1348.
Brown, J. S. 1942. The generalization of approach responses as a Enquist, M., Arak, A., Ghirlanda, S. & Wachtmeister, C.-A. 2002a.
function of stimulus intensity and strength of motivation. Journal Spectacular phenomena and limits to rationality in genetic and
of Comparative Psychology, 33, 209–226. cultural evolution. Transactions of the Royal Society of London.
Brown, J. S., Bilodeau, E. A. & Baron, M. R. 1951. Bidirectional Enquist, M., Ghirlanda, S., Lundqvist, D. & Wachtmeister,
gradients in the strength of a generalized voluntary response C.-A. 2002b. An ethological theory of attractiveness. In: Facial
to stimuli on a visual-spatial dimension. Journal of Experimental Attractiveness. Evolutionary, Cognitive and Social Perspectives (Ed.
Psychology, 41, 52–61. by G. Rhodes & L. A. Zebrowitz), pp. 127–157. Westport,
Brown, J. S., Clarke, F. R. & Stein, L. 1958. A new technique for Connecticut: Ablex.
studying spatial generalization with voluntary response. Journal of Ernst, A. J., Engberg, L. & Thomas, D. R. 1971. On the form of
Experimental Psychology, 55, 359–362. stimulus generalization curves for visual intensity. Journal of the
Brush, F. R., Jenkins, W. O., John, W. F. & Whiting, J. W. M. 1952. Experimental Analysis of Behavior, 16, 177–180.
Stimulus generalization after extinction and punishment: an Ewert, J.-P. 1980. Neuroethology. Berlin: Springer-Verlag.
experimental study of displacement. Journal of Abnormal and Farthing, G. W. & Hearst, E. 1972. Stimulus generalization and
Social Psychology, 47, 633–640. discrimination along the click-frequency (flutter) continuum in
Burley, N. 1986. Sexual selection for aesthetic traits in species with pigeons. Perception and Psychophysics, 12, 176–182.
biparental care. American Naturalist, 127, 415–445. Fay, R. 1970. Auditory frequency generalization in the goldfish
Burley, N., Krantzberg, G. & Radman, P. 1982. Influence of (Carassius auratus). Journal of the Experimental Analysis of Behavior,
colour-banding on the conspecific preferences of zebra finches. 14, 353–360.
Animal Behaviour, 30, 444–455. Ferraro, D. P. & Grisham, M. G. 1972. Discrimination and
Buss, A. H. 1961. Stimulus generalization and aggressive verbal generalization of complex visual shape variations in pigeons.
stimuli. Experimental Psychology, 61, 469–473. Perceptual and Motor Skills, 35, 915–927.
Buss, A. H. 1962. Stimulus generalization with aggressive verbal Friedman, H. 1963. Wave-length generalization as a function of
stimuli: a new paradigm. Journal of Psychology, 53, 417–424. spacing of test stimuli. Journal of Experimental Psychology: General,
Buss, A. H. & Daniell, E. F. 1967. Stimulus generalization and 65, 334–338.
schizophrenia. Journal of Abnormal Psychology, 72, 50–53. Frieman, J. & Thomas, D. R. 1970. Peak shift resulting from a
Buss, A. H., Murray, E. N. & Buss, E. 1968. Stimulus generalization chained reinforcement schedule. Psychonomic Science, 18, 315–
and fear of snakes. Journal of Personality and Social Psychology, 10, 316.
134–141. Galizio, M. 1980. Conditional gradient displacements: the effects of
Cheng, K. 2000. Shepard’s universal law supported by honeybees in conditional discrimination training on human auditory frequency
spatial generalization. Psychological Science, 11, 403–408. generalization. Journal of Experimental Psychology: Human Learning
Cheng, K. 2002. Generalisation: mechanistic and functional and Memory, 6, 766–777.
explanations. Animal Cognition, 5, 33–40. Galizio, M. 1985. Human peak shift: analysis of the effects of
Cheng, K., Spetch, M. L. & Johnson, M. 1997. Spatial peak shift three-stimulus discrimination training. Learning and Motivation,
and generalization in pigeons. Journal of Experimental Psychology: 16, 478–494.
Animal Behavior Processes, 23, 469–481. Gamberale-Stille, G. 2000. Decision time and prey gregariousness
Coren, S., Ward, L. & Enns, J. 1999. Sensation and Perception. 3rd influence attack probability in naive and experienced predators.
edn. Fort Worth, Philadelphia: Harcourt Brace. Animal Behaviour, 60, 95–99.
REVIEWS 31
Gamberale, G. & Sillen-Tullberg, B. 1998. Aposematism and Hearst, E. & Poppen, R. 1965. Steeepened generalization gradients
gregariousness: the combined effect of group size and coloration after massed extinction to the CS. Psychonomic Science, 2, 83–84.
on signal repellence. Proceedings of the Royal Society of London, Hearst, E., Koresko, M. B. & Poppen, R. 1964. Stimulus general-
Series B, 265, 889–894. ization and the response-reinforcement contingency. Journal of the
Gewirtz, J. L., Lones, L. V. & Waerneryd, K.-E. 1956. Stimulus Experimental Analysis of Behavior, 7, 369–380.
units and range of experienced stimuli as determinants of Hebb, D. O. 1949. The Organization of Behaviour. London: J. Wiley.
generalization-discrimination gradients. Journal of Experimental Hedges, D. G. 1983. Stimulus generalization and the peak shift with
Psychology, 52, 51–57. movement stimuli. Journal of Mathematical Biology, 15, 280–296.
Ghirlanda, S. 2002. Intensity generalization: physiology and Heiligenberg, W., Kramer, U. & Schulz, V. 1972. The
modelling of a neglected topic. Journal of Theoretical Biology, 214, angular orientation of the black eye-bar in Haplochromis burtoni
389–404. (Cichlidae: Pisces) and its relevance to aggressivity. Zeitschrift fur
Ghirlanda, S. & Enquist, M. 1998. Artificial neural networks as Vergleichende Physiologie, 76, 168–176.
models of stimulus control. Animal Behaviour, 56, 1383–1389. Hoffman, H. S. & Fleshler, M. 1964. Stimulus aspects of aversive
Ghirlanda, S. & Enquist, M. 1999. The geometry of stimulus controls: stimulus generalization of conditioned suppression fol-
control. Animal Behaviour, 58, 695–706. lowing discrimination training. Journal of the Experimental Analysis
Ghirlanda, S., Jansson, L. & Enquist, M. 2002. Chickens prefer of Behavior, 7, 233–239.
beautiful humans. Human Nature, 13, 383–389. Hogan, J. A. 1994. Structure and development of behavior systems.
Gilpin, A. R., Ratner, S. C. & Glanville, B. B. 1978. Stimulus Psychonomic Bulletin and Review, 1, 439–450.
generalization of contraction response to light in earthworm. Hogan, J. & Bolhuis, J. (Eds) 1994. Causal Mechanisms of
Perceptual and Motor Skills, 47, 230. Behavioural Development. Cambridge: Cambridge University Press.
Gluck, M. 1991. Stimulus generalisation and representation in Hogan, J., Kruijt, J. & Frijlink, J. 1975. ‘Supernormality’ in a learning
adaptive network models of category learning. Psychological situation. Zeitschrift für Tierpsychologie, 38, 212–218.
Science, 2, 50–55. Homa, D., Sterling, S. & Trepel, L. 1981. Limitations of examplar-
Grant, D. A. & Schiller, J. J. 1953. Generalization of the conditioned based generalisation and the abstraction of categorical infor-
galvanic skin response to visual stimuli. Journal of Experimental mation. Journal of Experimental Psychology: Human Learning and
Psychology, 46, 309–313. Memory, 7, 418–439.
Grice, G. R. & Saltz, E. 1950. The generalization of an instrumental Honig, W. K. & Urcuioli, P. J. 1981. The legacy of Guttman and
response to stimuli varying in the size dimension. Journal of Kalish (1956): 25 years of research on stimulus genralization.
Experimental Psychology, 40, 702–708. Journal of the Experimental Analysis of Behavior, 36, 404–445.
Grusec, T. 1968. The peak shift in stimulus generalization: equiva- Honig, W. K., Thomas, D. R. & Guttman, N. 1959. Differential
lent effects of errors and noncontingent shock. Journal of the effects of continuous extinction and discrimination training on the
Experimental Analysis of Behavior, 11, 239–249. generalization gradient. Journal of Experimental Psychology, 58,
Guthrie, E. R. 1930. Conditioning as a principle of learning. 145–152.
Psychological Review, 37, 412–428. Honig, W. K., Boneau, C. A., Burstein, K. R. & Pennypacker, H. S.
Guthrie, E. R. 1935. The Psychology of Learning. New York: Harper. 1963. Positive and negative generalization gradients obtained
Guttman, N. 1959. Generalization gradients around stimuli after equivalent training conditions. Journal of Comparative and
associated with different reinforcement schedules. Journal of Physiological Psychology, 56, 111–116.
Experimental Psychology, 58, 335–340. Horne, G. 1965. Physiological and psychological aspects of selective
Guttman, N. 1965. Effects of discrimination formation on general- perception. In: Advances in the Study of Behavior. Vol. 1 (Ed. by
ization from the positive-rate baseline. In: Stimulus Generalization D. S. Lehrman, R. A. Hinde & E. Shaw), pp. 155–215. New York:
(Ed. by D. Mostofsky), pp. 210–217. Stanford, California: Stanford Academic Press.
University Press. Hovland, C. I. 1937. The generalization of conditioned responses: II.
Guttman, N. & Kalish, H. 1956. Discriminability and stimulus The sensory generalization of conditioned responses with varying
generalization. Journal of Experimental Psychology, 51, 79–88. intensities of tone. Journal of Genetic Psychology, 51, 279–291.
Hall, G. & Honig, W. K. 1974. Stimulus control after extradimen- Huff, R. C., Sherman, J. E. & Cohn, M. 1975. Some effects of
sional training in pigeons: a comparison of response contingent response-independent reinforcement in auditory generalization
and noncontingent training procedures. Journal of Comparative gradients. Journal of the Experimental Analysis of Behavior, 23,
and Physiological Pscyhology, 87, 945–952. 81–86.
Hall, J. F. & Prokasy, W. F. 1961. Stimulus generalization to Hull, C. L. 1943. Principles of Behavior. New York: Appleton-Century-
absolutely discriminable tones. Perceptual and Motor Skills, 12, Crofts.
175–178. Hull, C. L. 1949. Stimulus intensity dynamisn (V) and stimulus
Hanson, H. 1959. Effects of discrimination training on stimulus generalization. Psychological Review, 56, 67–76.
generalization. Journal of Experimental Psychology, 58, 321–333. James, J. P. & Hughes, G. R. 1969. Generalization of habituation of
Hanson, H. M. 1961. Stimulus generalization following three- the GSR to white noise of varying intensities. Psychonomic Science,
stimulus discrimination training. Journal of Comparative and 14, 163–164.
Physiological Psychology, 54, 181–185. Jenkins, H. M. & Harrison, R. H. 1960. Effect of discrimination
Hawkins, A. D. 1981. The hearing abilities of fish. In: Hearing and training on auditory generalization. Journal of Experimental
Sound Communication in Fishes (Ed. by W. N. Tavolga, A. N. Psychology, 59, 246–253.
Popper & R. R. Fay), pp. 109–137. New York: Springer-Verlag. Jenkins, W. O., Pascal, G. R. & Walker, R. W. J. 1958. Deprivation
Haykin, S. 1999. Neural Networks: A Comprehensive Foundation. and generalization. Journal of Experimental Psychology, 56, 274–
2nd edn. New York: Macmillan. 277.
Hearst, E. 1968. Discrimination training as the summation of Johnson, K., Dalton, R. & Burley, N. 1993. Preferences of female
excitation and inhibition. Science, 162, 1303–1306. American goldfinches (Carduelis tristis) for natural and artificial
Hearst, E. 1969. Excitation, inhibition, and discrimination learn- male traits. Behavioral Ecology, 4, 138–143.
ing. In: Fundamental Issues in Associative Learning (Ed. by N. J. Kalish, H. I. 1958. The relationship between discriminability and
Mackintosh & W. K. Honig), pp. 1–41. Halifax, Nova Scotia: generalization: a re-evaluation. Journal of Experimental Psychology,
Dalhousie University Press. 55, 637–644.
Kalish, H. 1969. Chapters 9–12. In: Learning. Vol. 1 (Ed. by Marx, M. & McLean, J. 1971. Postdiscrimination gradients as a
M. Marx), pp. 259–275. London: MacMillan. function of variation in reward schedules. Psychonomic Science, 23,
Kalish, H. & Guttman, N. 1957. Stimulus generalization after equal 55–56.
training on two stimuli. Journal of Experimental Psychology, 53, Masataka, N. 1983. Categorical responses to natural and synthe-
139–144. sized alarm calls in Goeldi’s monkeys (Callimico goeldii). Primates,
Kalish, H. & Guttman, N. 1959. Stimulus generalisation after equal 24, 40–51.
training on three stimuli: a test of the summation hypothesis. Mednick, S. A. & Lehtinen, L. E. 1957. Stimulus generalization as a
Journal of Experimental Psychology, 57, 268–272. function of age in children. Journal of Experimental Psychology, 53,
Kalish, H. I. & Haber, A. 1963. Generalization: I. 180–183.
Generalization gradients from single and multiple stimulus points. Moye, T. B. & Thomas, D. R. 1982. Effects of memory reactivation
II. Generalization of inhibition. Journal of Experimental Psychology, treatment on postdiscrimination generalization performance in
65, 176–181. pigeons. Animal Learning and Behavior, 10, 159–166.
Kandel, E., Schwartz, J. & Jessell, T. 1991. Principles of Neural Murray, H. G. & Kohfeld, D. L. 1965. Role of adaption level stimulus
Science. 3rd edn. London: Prentice Hall. intensity dynamism. Psychonomic Science, 3, 439–440.
Kehoe, E. J. 1986. Summation and configuration in conditioning of Nallan, G. B., McCoy, D. F., Pace, G. M. & Welch, R. 1979.
the rabbit’s nictitating membrane response to compound stimuli. Generalization gradients following differential intradimensional
Journal of Experimental Psychology: Animal Behavior Processes, 12, autoshaping. Perceptual and Motor Skills, 48, 671–677.
186–195. Newlin, R. J., Rodgers, J. P. & Thomas, D. R. 1979. Two determi-
Kerr, L. M., Ostapoff, E. M. & Rubel, E. W. 1979. Influence of nants of the peak shift in human voluntary stimulus generalization.
acoustic experience on the ontogeny of frequency generalization Perception and Psychophysics, 25, 478–486.
gradients in the chicken. Journal of Experimental Psychology: Animal Nicholson, J. N. & Gray, J. A. 1971. Behavioural contrast and peak
Behavior Processes, 5, 97–117. shift in children. British Journal of Psychology, 62, 367–373.
Kessen, W. 1953. Response strength and conditioned stimulus Ohinata, S. 1978. Postdiscrimination shift of the goldfish (Carassius
intensity. Journal of Experimental Psychology, 45, 82–86. auratus) on a visual wavelength continuum. Annual of Animal
Klein, M. & Rilling, M. 1974. Generalization of free-operant avoid- Psychology, 28, 113–122.
ance behavior in pigeons. Journal of the Experimental Analysis of Olson, G. & King, R. A. 1962. Supplementary report: stimulus
Behavior, 21, 75–78. generalization gradients along a luminosity continuum. Journal of
Klipec, W. D., Akins, F. R. & Koerner, A. 1979. The effects Experimental Psychology, 63, 414–415.
of D-9-THC on wavelength generalization in the pigeon. Osorio, D., Vorobyev, M. & Jones, C. 1999. Colour vision of
Physiological Psychology, 7, 153–155. domestic chicks. Journal of Experimental Biology, 202, 2951–2959.
Koehler, O. & Zagarus, A. 1937. Beiträge zum Brutverhalten Pavlov, I. P. 1927. Conditioned Reflexes. Oxford: Oxford University
des Halsbandregepfeifers (Charadrius h. hiaticula L.). Beiträge zur Press.
Fortpflanzung der Vögel, 13, 1–9. Pearce, J. M. 1987. A model for stimulus generalization in Pavlovian
Lashley, K. S. & Wade, M. 1946. The Pavlovian theory of conditioning. Psychological Review, 94, 61–73.
generalization. Psychological Review, 53, 72–87. Pearce, J. M. 1994. Similarity and discrimination: a selective review
Lawrence, C. 1973. Generalization along the dimension of sound and a connectionist model. Psychological Review, 101, 587–607.
intensity in pigeons. Animal Learning and Behavior, 1, 60–64. Penkower Rosen, A. & Terrace, H. S. 1975. On the minimal
Lea, S. E. G. & Harrison, S. N. 1978. Discrimination of conditions for the development of a peak shift and inhibitory
polymorphous stimulus sets in pigeons. Quarterly Journal of stimulus control. Journal of the Experimental Analysis of Behavior,
Experimental Psychology, 30, 521–537. 23, 385–414.
Logan, F. A. 1954. A note on stimulus intensity dynamism (V). Perkins, C. C. J. 1953. The relation between conditioned stimulus
Psychological Review, 61, 77–80. intensity and response strength. Journal of Experimental
Lorenz, K. 1937. Ueber die Bildung des linstinktsbegriff. Psychology, 46, 225–231.
Naturwissenschaft, 25, 289–300, 307–318, 324–331. Peterson, N. 1962. Effect of monochromatic rearing on the control
Lorenz, K. 1981. The Foundations of Ethology. New York: Springer- of responding by wavelength. Science, 136, 774–775.
Verlag. Pierrel, R. & Sherman, J. G. 1960. Generalization of auditory
Lyons, J., Klipec, W. D. & Steinsultz, G. 1973. The effect of intensity following discrimination training. Journal of the
chloropromazine on discrimination performance and the peak Experimental Analysis of Behavior, 3, 313–322.
shift. Physiological Psychology, 1, 121–124. Pryke, S. R. & Andersson, S. 2002. A generalized female bias for
Mackintosh, N. 1974. The Psychology of Animal Learning. London: long tails in a short-tailed widowbird. Proceedings of the Royal
Academic Press. Society of London, Series B, 269, 2141–2146.
Mackintosh, N. J., Kaye, H. & Bennett, C. H. 1991. Perceptual Purtle, R. B. 1973. Peak shift: a review. Psychological Bulletin.
learning in flavour aversion conditioning. Quarterly Journal of 408–421.
Experimental Psychology, B43, 297–322. Razran, G. 1949. Stimulus generalisation of conditioned responses.
McLaren, I.-P., Bennett, C., Guttman-Nahir, T. & Kim, K. 1995. Psychological Bulletin, 46, 337–365.
Prototype effects and peak shift in categorization. Journal of Rescorla, R. A. 1976. Stimulus generalization: some predictions
Experimental Psychology: Learning, Memory, and Cognition, 21, from a model of pavlovian conditioning. Journal of Experimental
662–673. Psychology, 2, 88–96.
Magnus, D. 1958. Experimentelle Untersuchung zur Bionomie and Rhodes, G. 1996. Superportraits: Caricatures and Recognition. Hove:
Ethologie des Kaisermantels Argynnis paphia L. (Lep. Nymph.) I. The Psychology Press.
Uber optische Auslöser von Anfliegereaktion und ihre Bedeutung Rhodes, G. & Zebrowitz, L. A. (Eds) 2002. Facial Attractiveness.
für das Sichfinden der Geschlechter. Zeitschrift fur Tierpsychologie, Evolutionary, Cognitive and Social Perspectives. Westport,
15, 397–426. Connecticut: Ablex.
Marsh, G. 1972. Prediction of the peak shift in pigeons from Riccio, D. C., Urda, M. & Thomas, D. R. 1966. Stimulus control in
gradients of excitation and inhibition. Journal of Comparative and pigeons based on proprioceptive stimuli from floor inclination.
Physiological Psychology, 81, 262–266. Science, 153, 434–436.
REVIEWS 33
Rohrbaugh, M., Brennan, J. F. & Riccio, D. C. 1971. Control of Thomas, D. R. & Bistey, G. 1964. Stimulus generalization as a
two-way shuttle avoidance in rats by auditory frequency and function of the number and range of generalization test stimuli.
intensity. Journal of Comparative and Physiological Psychology, 75, Journal of Experimental Psychology, 68, 599–602.
324–330. Thomas, D. R. & King, R. A. 1959. Stimulus generalization as a
Rosch, E., Simpson, C. & Miller, R. S. 1976. Structural bases function of level of motivation. Journal of Experimental Psychology,
of typicality effects. Journal of Experimental Psychology: Human 57, 323–328.
Perception and Performance, 2, 491–502. Thomas, D. R. & Mitchell, K. 1962. Instructions and stimulus
Rosenbaum, G. 1953. Stimulus generalization as a function of categorizing in a measure of stimulus generalization. Journal of the
level of experimentally induced anxiety. Journal of Experimental Experimental Analysis of Behavior, 5, 375–381.
Psychology, 45, 35–43. Thomas, D. R. & Setzer, J. 1972. Stimulus generalization gradients
Rubel, E. W. & Rosenthal, M. H. 1975. The ontogeny of auditory for auditory intensity in rats and guinea pigs. Psychonomic Science,
frequency generalization in the chicken. Journal of Experimental 28, 22–24.
Psychology: Animal Behavior Processes, 1, 287–297. Thomas, D. R. & Switalski, R. W. 1966. Comparison of stimulus
Scavio, M. J. J. & Gormezano, I. 1974. CS intensity effects on rabbit generalization following variable-ratio and variable-interval
nictitating membrane, conditioning, extinction and general- training. Journal of Experimental Psychology: General, 71, 236–240.
ization. Pavlovian Journal of Biological Science, 9, 25–34. Thomas, D. R., Ost, J. & Thomas, D. 1960. Stimulus generalization
Schaller, F. & Schwalb, H. 1961. Attrappenversuche mit Larven and as a function of the time between training and testing procedures.
Imagines einheimischer Leuchtkäfer (Lampirinae). Zeitschrift für Journal of the Experimental Analysis of Behavior, 3, 9–14.
Tierpsychologie, 24, 154–166. Thomas, D. R., Mood, K., Morrison, S. & Wiertelak, E. 1991. Peak
Schlosberg, H. & Solomon, R. L. 1943. Latency of response in a shift revisited: a test of alternative interpretations. Journal of
choice discrimination. Journal of Experimental Psychology, 33, Experimental Psychology: Animal Behavior Processes, 17, 130–140.
22–39. Thomas, D. R., Lusky, M. & Morrison, S. 1992. A comparison of
generalization functions and frame of reference effects in different
Shanks, D. S. 1995. The Psychology of Associative Learning.
training paradigms. Perception and Psychophysics, 51, 529–540.
Cambridge: Cambridge University Press.
Thompson, R. F. 1965. The neural basis of stimulus generalization.
Shepard, R. N. 1987. Toward a universal law of generalization for
In: Stimulus Generalization (Ed. by D. I. Mostofsky), pp. 154–178.
psychological science. Science, 237, 1317–1323.
Stanford, California: Stanford University Press.
Shepard, R. N. 1988. Time and distance in generalization and Thorndike, E. L. 1911. Animal Intelligence: Experimental Studies.
discrimination: reply to Ennis (1988). Journal of Experimental New York: MacMillan.
Psychology: General, 117, 415–416. Tinbergen, N. 1951. The Study of Instinct. New York: Oxford
Shepard, R. N. 1990. Neural nets for generalization and classifi- University Press.
cation: comment on Staddon & Reid (1990). Psychological Review, Tinbergen, N., Meeuse, B. J. D., Boerema, L. K. & Varossieau,
97, 579–580. W. W. 1942. Die Balz des Samtfalters, Eumenis (=Satyrus) semele
Sloane, H. N. J. 1964. Stimulus generalization along a light flicker (L.). Zeitschrift fur Tierpsychologie, 5, 182–226.
rate continuum after discrimination training with several S-s. Tinbergen, N., Kruuk, H. & Paillette, M. 1962. Egg shell removal
Journal of the Experimental Analysis of Behavior, 7, 217–221. by the black-headed gull (Larus r. ridibundus L.) II. The effects of
Spence, K. 1937. The differential response in animals to stimuli experience on the response to colour. Bird Study, 9, 123–131.
varying within a single dimension. Psychological Review, 44, 430– Tosti, D. T. & Ellis, H. C. 1964. Stimulus generalization in
444. the absence of discrimination factors. Journal of Experimental
Spence, K. W. 1942. The basis of solution by chimpanzees of the Psychology, 68, 595–598.
intermediate size problem. Journal of Experimental Psychology, 31, Von Uexküll, J. J. 1928. Teoretische Biologie. Berlin: Springer-Verlag.
257–271. Wagner, A. R. 1971. Elementary associations. In: Essays in
Staddon, J. 1975. A note on the evolutionary significance of Neobehaviorism: a Memorial Volume to Kennet W. Spence (Ed. by
supernormal stimuli. American Naturalist, 109, 541–545. H. H. Kendler & J. T. Spence), pp. 187–213. East Norwalk,
Staddon, J. E. & Reid, A. K. 1990. On the dynamics of Connecticut: Appleton-Century-Crofts.
generalization. Psychological Review, 97, 576–578. Walsh, V. & Kulikowski, J. 1998. Perceptual Constancy. Cambridge:
Steinshneider, A., Lipton, E. L. & Richmond, J. B. 1966. Auditory Cambridge University Press.
sensitivity in the infant: effect of intensity on cardiac and motor Wasserman, E. A., Gagliardi, J. L., Cook, B. R., Kirkpatrick-Steger,
responsivity. Child Development, 37, 233–252. K. & Astley, S. L. 1996. The pigeon’s recognition of drawings of
depth-rotated stimuli. Journal of Experimental Psychology: Animal
Stevenson, J. G. 1966. Stimulus generalization: the ordering and
Behavior Processes, 22, 205–221.
spacing of test stimuli. Journal of the Experimental Analysis of
Watanabe, S. 1988. Failure of visual prototype learning in the
Behavior, 9, 457–468.
pigeon. Animal Learning and Behavior, 16, 147–152.
Tapper, D. N. & Halpern, B. P. 1968. Taste stimuli: a behavioral
Watson, J. B. 1924. Behaviorism. New York: Peoples Institute.
categorization. Science, 161, 708–709.
Weinberg, S. G. 1973. Some determinants of the peak shift in
Tempone, V. J. 1965. Stimulus generalization as a function of stimulus generalization. Psychological Reports, 32, 47–58.
mental age. Child Development, 36, 229–235. Weiss, S. J. 1972. Stimulus compounding in free-operant and
Terrace, H. S. 1964. Wavelength generalization after discrimination classical conditioning. Psychological Bulletin, 78, 189–208.
training with and without errors. Science, 144, 78–80. Weiss, S. J. & Schindler, C. W. 1981. Generalization and peak shift
Terrace, H. 1966. Stimulus control. In: Operant Behavior. Vol. 1 in rats under conditions of positive reinforcement and avoidance.
(Ed. by W. K. Honig), pp. 271–344. New York: Appleton-Century- Journal of the Experimental Analysis of Behavior, 35, 175–185.
Crofts. Wheatley, K. L. & Thomas, D. R. 1974. Relative and absolute
Thomas, D. R. 1962. The effects of drive and discrimination training density of reinforcement as factors influencing the peak shift.
on stimulus generalization. Journal of Experimental Psychology: Journal of the Experimental Analysis of Behavior, 22, 409–418.
General, 64, 24–28. Wilkie, D. M. 1972. The peak shift and behavioral contrast: effects of
Thomas, D. R. 1993. A model for adaptation-level effects on discrimination training with delayed reinforcement. Psychonomic
stimulus generalization. Psychological Review, 100, 658–673. Science, 26, 257–258.
Wills, S. & Mackintosh, N. J. 1998. Peak shift on an artificial (1975, 3), Dukhayyil & Lyons (1973, 1), Dysart et al.
dimension. Quarterly Journal of Experimental Psychology, 51B, 1–31. (1974, 5), Friedman (1963, 3), Frieman & Thomas
Wolffgramm, J. & Todt, D. 1982. Pattern and time specificity in (1970, 2), Grusec (1968, 4), Guttman & Kalish (1956, 4),
vocal responses of blackbirds Turdus merula L. Behaviour, 81, Guttman (1959, 2), Hanson (1959, 5), Hanson (1961, 2),
264–286. Honig et al. (1959, 12), Kalish & Haber (1963, 7), Klipec
Zielinski, K. & Jakubowska, E. 1977. Auditory intensity general- et al. (1979, 6), Marsh (1972, 2), Marx & McLean (1971,
ization after CER differentiation training. Acta Neurobiologiae 1), Moye & Thomas (1982, 2), Penkower Rosen & Terrace
Experimentalis, 37, 191–205.
(1975, 1), Terrace (1964, 3), Thomas & King (1959, 5),
Thomas et al. (1960, 3), Thomas (1962, 3), Thomas &
Appendix 1: Statistical Tests Switalski (1966, 2), Wheatley & Thomas (1974, 4). Total:
98 gradients.
Data from the sources listed below were acquired from
published tables, by millimetre paper readouts or by Test 3
computer scanning and readout with the g3data software Gaussian and exponential fits do not make differ-
(freely available from http://beam.helsinki.fi/frantz/ ent predictions about peak location in rearrangement
software/g3data.php). The number following the year in gradients. Wilcoxon signed-ranks test: Z= 0.87, N=223,
each citation indicates the number of data sets in P=0.38. Sources: as Test 1.
the cited study relevant to the test. Statistical tests
were computed using the statistical package for the Test 4
GNU/Octave software (freely available from http://www. Line-tilt generalization around a vertical line in pigeons
octave.org). All tests are two tailed. is skewed towards clockwise rotations. One-sample t test:
t9 = 2.37, N=10, P<0.05. Sources: individual gradients
Test 1 from the 0 group in Hearst et al. (1964). One sample t
Gaussian fits to rearrangement gradients account for test, t19 = 5.21, N=20, P<10 4. Sources: average group
about 3% more of the variance than exponential fits gradients from Bloomfield (1967, 1), Hall & Honig (1974,
(Appendix 2). Sign test: Z=163, N=223, P<10 10. 4), Hearst et al. (1964, 3), Hearst & Poppen (1965, 3),
Sources: Akins et al. (1981, 2), Ames & Yarczower (1965, Hearst (1968, 4), Hearst (1969, 3), Honig et al. (1963, 2).
2), Baron (1973, 6), Bloomfield (1967, 2), Blough (1969, Total: 20 gradients.
3), Blough (1972, 11), Blough (1975, 3), Brown et al.
(1951, 2), Brown et al. (1958, 2), Cheng et al. (1997, 15), Test 5
Crawford et al. (1980, 4), Dukhayyil & Lyons (1973, 1), Light wavelength generalization around 550 nm in
Dysart et al. (1974, 5), Fay (1970, 6), Friedman (1963, 3), pigeon is skewed towards longer wavelengths. One-
Frieman & Thomas (1970, 2), Galizio (1980, 2), Galizio sample t test: t17 = 8.40, N=18, P<10 6. Sources: aver-
(1985, 24), Gewirtz et al. (1956, 7), Ghirlanda et al. age group gradients from Friedman (1963, 3), Guttman &
(2002, 2), Grusec (1968, 4), Guttman & Kalish (1956, 4), Kalish (1956, 1), Hanson (1959, 1), Hanson (1961, 1),
Guttman (1959, 2), Hall & Honig (1974, 4), Hanson Honig et al. (1959, 4), Marsh (1972, 1), Thomas & King
(1959, 5), Hanson (1961, 2), Hearst et al. (1964, 4), Hearst (1959, 5), Thomas & Switalski (1966, 2). Total: 18
& Poppen (1965, 3), Hearst (1968, 4), Hearst (1969, 6), gradients.
Honig et al. (1959, 12), Honig et al. (1963, 2), Jenkins &
Harrison (1960, 1), Kalish (1958, 4), Kalish & Haber Test 6
(1963, 7), Klipec et al. (1979, 6), Lyons et al. (1973, 8), (1) Strength of response bias (measured as the ratio of
Marsh (1972, 2), Marx & McLean (1971, 1), Mednick & maximum responding to S + responding) in light wave-
Lehtinen (1957, 2), Moye & Thomas (1982, 2), Nallan length intradimensional tests with pigeons increases as
et al. (1979, 1), Ohinata (1978, 5), Penkower Rosen & separation between S + and S decreases. r35 = 0.47,
Terrace (1975, 1), Tempone (1965, 3), Terrace (1964, 3), P<0.005.
Thomas & King (1959, 5), Thomas et al. (1960, 3), (2) Distance of gradient peak from S + in light wave-
Thomas (1962, 3), Thomas & Bistey (1964, 3), Thomas & length intradimensional tests with pigeons increases as
Switalski (1966, 2), Wheatley & Thomas (1974, 4), Wilkie separation between S + and S decreases. r35 = 0.48,
(1972, 1). Total: 223 gradients. These studies comprise P<0.005.
both inter- and intradimensional tests, as well as tests Sources: Akins et al. (1981, 2), Dukhayyil & Lyons
where two positive stimuli were subject to different (1973, 1), Frieman & Thomas (1970, 2), Grusec (1968, 4),
reinforcement shcedules (e.g. Dysart et al. 1974; Guttman (1959, 1), Hanson (1959, 4), Honig et al. (1959,
Wheatley & Thomas 1974), producing a gradient with a 4), Klipec et al. (1979, 6), Marsh (1972, 1), Marx &
single peak. McLean (1971, 1), Moye & Thomas (1982, 2), Terrace
(1964, 3), Thomas et al. (1960, 3), Thomas (1962, 3).
Test 2 Total: 37 gradients.
Advantage of Gaussian fits to light wavelength
generalization gradients in pigeons increases with finer Test 7
sampling. r96 = 0.55, P<10 9. Sources: Akins et al. (1) Responding along intensity dimensions keeps above
(1981, 2), Blough (1969, 3), Blough (1972, 11), Blough S + levels, for stimuli further away from S , more often
REVIEWS 35
than along rearrangement dimensions. Fisher’s exact (1966, 2), Thomas & Setzer (1972, 4), Wills & Mackintosh
probability test: P<10 6. Along intensity dimensions, (1998, 5), Zielinski & Jakubowska (1977, 3). Total: 39
responding to the most extreme intensity sampled gradients. ‘Degree of monotonicity’ is defined as the
was above S + level in 30 of 38 gradients, and along absolute Spearman rank correlation between responding
rearrangement dimensions in two of 93. observed to a stimulus and its position along the stimulus
(2) More monotonic intensity gradients are found dimension (absolute values are used to cover both
along intensity dimensions than along rearrangement increasing and decreasing gradients). The correlation
ones. Fisher’s exact probability test P<10 6. Along inten- reported above is the Spearman rank correlation between
sity dimensions, 16 of 38 gradients were monotonic, and group size and average degrees of monotonicity. The
along rearrangement dimensions none of 93. A gradient value N=11 reported above refers to the number of group
is considered monotonic if it never drops beyond S + sizes found in the surveyed studies (in the range 2–45).
(see text).
Sources of intensity data: Baron & Harris (1968, 1), Test 10
Bartoshuk (1964, 1), Brennan & Riccio (1973, 4), Ernst
et al. (1971, 4), Huff et al. (1975, 2), Lawrence (1973, 4), Size generalization gradients obtained in inter-
Olson & King (1962, 4), Pierrel & Sherman (1960, 2), dimensional tests appear to peak at sizes larger than S + .
Razran (1949, 3), Rohrbaugh et al. (1971, 1), Steinshneider Wilcoxon signed-ranks test: T=2.20, N=7, P<0.05.
et al. (1966, 2), Thomas & Setzer (1972, 4), Wills & Sources: Brush et al. (1952, 1), Buss & Daniell (1967, 3),
Mackintosh (1998, 3), Zielinski & Jakubowska (1977, 3). Dougherty & Lewis (1991, 1), Grant & Schiller (1953, 1),
Total: 39 gradients. Sources of rearrangement data: Akins Tosti & Ellis (1964, 1). Total: 7 gradients. The result holds
et al. (1981, 2), Ames & Yarczower (1965, 1), Baron (1973, when peak position is estimated by either Gaussian or
4), Bloomfield (1967, 1), Cheng et al. (1997, 8), Crawford exponential curves. The linear dimensions of the stimuli
et al. (1980, 4), Dukhayyil & Lyons (1973, 1), Frieman & were used as the independent variable (e.g. circle
Thomas (1970, 2), Galizio (1985, 14), Ghirlanda et al. diameter or square side). Using stimulus area, which is
(2002, 2), Grusec (1968, 4), Guttman (1959, 1), Hanson another natural choice for size dimensions, would have
(1959, 3), Hearst (1969, 2), Honig et al. (1959, 8), Kalish & yielded even bigger asymmetries.
Haber (1963, 1), Klipec et al. (1979, 6), Lyons et al. (1973,
8), Marsh (1972, 1), Marx & McLean (1971, 1), Moye & Test 11
Thomas (1982, 2), Nallan et al. (1979, 1), Ohinata (1978, Gaussian fits to size gradients account for about 3%
4), Terrace (1964, 3), Thomas et al. (1960, 3), Thomas more of the variance than exponential fits (Appendix 2).
(1962, 2), Wilkie (1972, 1), Wills & Mackintosh (1998, 3). Sign test: Z=8, N=8, P=0. Sources: Brush et al. (1952, 1),
Total: 93 gradients. Buss & Daniell (1967, 3), Dougherty & Lewis (1991,
Sources of size data: Andersson (1982, 1), Baerends 2), Grant & Schiller (1953, 1), Tosti & Ellis (1964, 1).
(1982, 1), Basolo & Delaney (2001, 1), Dougherty & Lewis Total: 8 gradients. These studies comprise both inter- and
(1991, 1), Magnus (1958, 3), Pryke & Andersson (2002, 1), intradimensional tests.
Schaller & Schwalb (1961, 2). Total: 10 gradients.
Test 12
Test 8: Pigeons generalize less, along the light wavelength
(1) Intensity generalization produces stronger response dimension, after a discrimination between two wave-
biases than rearrangement dimensions. Mann–Whitney U lengths than between one wavelength and a dark stimu-
test: Z=4.89, N1 =38, N2 =93, P<10 6. lus. Kolmogorov–Smirnov two-sample test: Z=1.67,
(2) Size dimensions produce stronger biases than N1 =48, N2 =40, P<0.001. Sources: Akins et al. (1981, 2),
rearrangement dimensions. Mann–Whitney U test: Blough (1969, 3), Blough (1972, 11), Blough (1975, 3),
U=4.21, N1 =10, N2 =93, P<10 4. Dukhayyil & Lyons (1973, 1), Dysart et al. (1974, 1),
(3) Size and intensity dimensions appear to produce Friedman (1963, 3), Frieman & Thomas (1970, 2), Grusec
biases of comparable size. Mann–Whitney U test: (1968, 4), Guttman & Kalish (1956, 4), Guttman (1959,
U= 0.61, N1 =10, N2 =38, P=0.54. 1), Hanson (1959, 5), Hanson (1961, 1), Honig et al.
Sources: see Test 7. Strength of bias is measured as the (1959, 12), Kalish & Haber (1963, 7), Klipec et al. (1979,
ratio of maximum observed responding to S + responding. 6), Marsh (1972, 2), Marx & McLean (1971, 1), Moye &
Thomas (1982, 2), Penkower Rosen & Terrace (1975, 1),
Terrace (1964, 3), Thomas & King (1959, 5), Thomas et al.
Test 9 (1960, 3), Thomas (1962, 3), Thomas & Switalski (1966,
Better sampling of the gradient (more experimental 2). Total: 88 gradients.
subjects) leads to higher degree of monotonicity in inten-
sity gradients. rs =0.58, N=11, P<0.01. Sources: Baron &
Harris (1968, 1), Bartoshuk (1964, 1), Brennan & Riccio Appendix 2: Data Analysis
(1973, 4), Ernst et al. (1971, 4), Hall & Prokasy (1961, 1),
Huff et al. (1975, 2), James & Hughes (1969, 1), Lawrence Fitting procedure
(1973, 4), Olson & King (1962, 4), Pierrel & Sherman We describe briefly the fitting procedure leading to the
(1960, 2), Rohrbaugh et al. (1971, 1), Steinshneider et al. results on gradient shape reported in the main text.
By ‘exponential’ and ‘Gaussian’ shapes we mean the Gaussian fits to each gradient. These can be used in
following functions, respectively: statistical tests as detailed in Appendix 1.
fE(x)=a exp(/xb//c)+d (1)

Testing for gradient symmetry
2 2
fG(x)=a exp((xb) /2c )+d (2) To test for symmetry of gradients obtained from inter-
dimensional tests, we divided the gradient into a ‘left’
where, in both functions, the maximum value of the and a ‘right’ side, using S + as the reference point. We then
function is a+d, the peak position is b, c determines how summed all responses to stimuli on the same side, obtain-
quickly the gradient drops from the peak, and d is the ing a left sum L and a right sum R (responses to S + are
minimum function value (approximately reached very far excluded from both L and R). R and L were combined as
from the peak). The parameter d has been included follows to yield the test statistics:
because not all empirical gradients fall to zero within the
probed stimulus range. However, d was not fitted; rather,
it was set equal to the minimum gradient value, consist-
ent with its meaning in equations (1) and (2). Using a, b
and c as fitted parameters, we fitted both equations (1) which has the meaning of proportion of responses to the
and (2) to the studies listed as sources for Test 1. For left of S + . By construction, its value is independent of
fitting we used the Marquardt–Levenberg algorithm pro- whether published data are reported as absolute numbers
vided with the Gnuplot software, version 3.7.1. (available of responses or as proportions of responses to each stimu-
via anonymous ftp from ftp://ftp.ucc.ie/pub/gnuplot). lus. Under the null hypothesis that a set of gradients
For each data set and each fit, we then computed the shows no systematic asymmetry, the expected value of l is
Pearson correlation r, and we used r2 as an estimate of 0.5. In testing this hypothesis by one-sample t tests (see
how much of the variance in empirical data is accounted Appendix 1), we considered the same number of stimuli
for by the fitted curve (a measure of goodness of fit). We on each side of S + (a few times discarding a data point),
thus had two samples of r2’s, from the exponential and and covering the same range.

Ghirlanda 2003

Uploaded by

Copyright:

Available Formats

Ghirlanda 2003

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Ghirlanda 2003

Uploaded by

Copyright:

Available Formats

ANIMAL BEHAVIOUR, 2003, 66, 15–36

STEFANO GHIRLANDA & MAGNUS ENQUIST

(Received 21 September 2000; initial acceptance 15 November 2000;

Blough (1975) Blough (1969)

Hanson (1959) Thomas & Bistey (1964)

Blough (1972) Nallan et al. (1979)

Table 2. Examples of generalization along rearrangement dimensions

Responses in first minute of test

asymmetrical gradients and strong response biases

Table 3. Examples of generalization along intensity dimensions

*Both monotonic and nonmonotonic gradients were found.

0.8 Rearrangement (N = 93) Sessions 5 and 6

Table 4. Examples of generalization along size dimensions

Circle No data Rat, R. r. norvegicus Grice & Saltz 1950

1969) to about 27 nm (Thomas & King 1959), indicating

Table 6. Basic predictions by models of generalization

fE(x)=a exp(/xb//c)+d (1)

You might also like