Journal of Experimental Psychology

Vol. 5!, No. 1, 1956

DISCRIMINABILITY AND STIMULUS GENERALIZATION 1

NORMAN GUTTMAN AND HARRY I. KALISH 2
Duke University

Three different forms for the stimu- example,

lus generalization gradient have been
proposed: concavity (Hull, 6), con- AS
DL = = .75] - [_p(R) = .50]
vexity (Spence, 12) and linearity
(Schlosberg and Solomon, 10). In AS
the formulation of each of these pro-
posals, however, there has been a
tacit or explicit assumption of some On the other hand, the slope of the
relationship between discriminability, generalization gradient, i.e., the gen-
in the psychophysical sense, and the eralization decrement, is given by
generalization decrement. This as-
sumption is manifested in Hull's This implies that the gen-
AS
postulate that generalization gradi- eralization decrement should be pro-
ents are decreasing exponential func- portional to the reciprocal of the DL.
tions on a j.n.d. scale, and also in It is a way of saying that the organism
Schlosberg and Solomon's proposal generalizes to the extent that it can-
that generalization gradients are not discriminate or that generalization
straight lines on an equal-appearing is the inverse of discrimination. In
interval scale. Spence's use of a essence, this proposal is a reformula-
logarithmic stimulus plot appears to tion of Schlosberg and Solomon's
stem from similar considerations. hypothesis (10).
This assumption does not only ap- The foregoing analysis may be sub-
pear to be intuitively reasonable, but jected to empirical verification by in-
can also be shown to have a rational vestigating a continuum for which the
discriminability function is known.
basis by comparing the dimensions of For the present experiment, wave
the difference threshold and the gen- length of light was chosen as a dimen-
eralization decrement (from which the sion along which to test generalization
generalization gradient can be derived because the DL for wave length
by integration). The difference limen (AX) is not constant over the spec-
(DL) is ordinarily defined by the slope trum. By an appropriate selection of
of the psychophysical function. For CS values, therefore, it should be
1
This study was supported by the Duke possible to produce a set of gradients
University Research Council and by grants whose slopes reflect the characteristics
M-631C and M-629C from the National Insti- of the AX function. Where AX is
tute of Mental Health, United States Public small, the gradient should be sharp;
Health Service. The research program of
which this is a part is currently supported by where AX is large, the gradient should
grant MH-1002 from the National Institute of be relatively flat; and if the CS is
Mental Health, United States Public Health fixed at a value where AX is either in-
Service. The authors are indebted to Harley creasing or decreasing, the generaliza-
M. Hanson, Werner K. Honig, and Mrs. tion gradient should show a corre-
Deborah H. Nickerson for their assistance in
this research. sponding asymmetry. This contin-
2
Now at University of Missouri. uum is also interesting because of the
79
80 NORMAN GUTTMAN AND HARRY I. KALISH

qualitative changes which occur The magazine was situated directly below the
within it and which would lead to the key and consisted of a metal door hinged at the
top. This door was actuated by a cam on a con-
expectation, on the basis of subjective stant-speed motor which completed a cycle in 5
experience, that generalization would sec. When a reinforcement was presented, the
occur extensively within a hue but cam allowed the door to open abruptly, exposing
would decrease abruptly as the transi- the food and switching on a 7.5-w. lamp behind
tions between spectral hues are ap- the food tray. The door remained open for 3.5
sec. and then closed gradually, turning off the
proached. light as it shut.
The technique of the present study, The external side of the key was illuminated
which utilizes the aperiodically rein- by a 6-v., 18-amp. ribbon filament lamp whose
forced key-pecking response, permits beam was directed through a tunable wave-
length filter and whose filament image was cast
an examination of the generalization slightly out of focus on the key. The filter was a
gradients for individual Ss. More- Cambridge Thermionic Corp. Monochromator
over, this technique makes it possible (Model B). This instrument utilizes the differ-
to investigate the generalization gradi- ential rotation of quartz crystals for various
ents at different levels of response wave lengths to produce dispersion of the spec-
trum. Its transmission spectrum at a given
strength during the course of extinc- setting consists of a centroid of high relative
tion, as well as variations in the gen- intensity flanked symmetrically by a series of
eralization gradient attributable to periodically spaced bands of lower intensity.
individual differences in response Although the instrument may not be considered
a true monochromator, it produces a stimulus
strength. The obtaining of general- patch of high apparent purity for the human ob-
ization gradients for individual Ss in server. The band width of the centroid increases
this experiment of an outcome of linearly from 14 to 24 MH over the range of set-
the fact that aperiodic reinforcement tings from 450 to 640 M/K. The percentage
greatly increases resistance to extinc- transmission of incident light is approximately
flat over this range, varying between 2 and 4%.
tion, such that the introduction of a No attempt was made to produce an equal energy
test stimulus for a brief interval during spectrum or alternatively to equate the spec-
experimental extinction reduces the trum lights in terms of the visibility function.
response strength by a small fraction Procedure.—The procedure used in the pres-
ent study derives from the work of Skinner (11)
of its total extent, e.g., in a 30-sec. and Ferster (3), and is similar to that used by
test, 50 responses may be subtracted Brush, Bush, Jenkins, Johns, and Whiting (2).
from a "reserve" of several thousand. When the Ss were reduced to 80% of their ad
lib. weight, they were trained to eat from the
food magazine and were conditioned by the
METHOD method of successive approximation to peck at
Subjects.—The Ss were 24 experimentally the key. The 24 Ss were divided into groups of
naive pigeons maintained by restricted feeding 6, and each group was trained on a given CS.
to 80% of their body weight under ad libitum The CS wave lengths were 530, 550, 580, and
feeding. 600Mp. Fifty consecutive reinforcements were
Apparatus.—A modified version of the given on each of two days. Following this, ad-
Skinner automatic key-pecking apparatus was ditional sessions of aperiodic reinforcement were
used. The S's compartment consisted of a ply- administered, using a mean inter-reinforcement
wood box 14 in. high, 12 in. wide, and 12 in. long interval of approximately 1 min. During APR,
and was situated in an air-conditioned, darkened, 60-sec. stimulus-on intervals were alternated with
soundproof chamber. The roof of S's box was 10-sec. stimulus-off intervals, and 30 stimulus-on
clear Lucite and the floor was a perforated metal intervals constituted each daily training session.
grating covered with a sheet of translucent During the stimulus-on interval a shutter in the
plastic. The upper half of one of the walls was light path of the monochromator was opened
made of aluminum with a f-in. round aperture at and the key was illuminated with colored light.
the center 6.5 in. above the floor of the box. At the same time, the interior of S's box was
The S's key, a rectangle of translucent plastic illuminated by a 25-w. lamp reflected through
(Insurok) lightly sprung against the wall of the the translucent plastic floor, resulting in an illu-
box, was exposed through this opening and con- minance of <1.00 ft.-candle on the vertical
trolled counting and reinforcing circuits. walls. In the stimulus-off condition, similar to
 DISCRIMINABILITY AND STIMULUS GENERALIZATION 81

TABLE 1
GENERALIZATION TEST STIMULI IN M/i

Group -70 -60 -50 -40 -30 -20 -10 cs +10 +20 +30 +40 +50 +60

530 470 490 500 510 520 530 540 550 560 570 590
550 490 510 520 530 540 550 560 570 580 590 610
580 520 540 550 560 570 580 590 600 610 620 640
600 530 550 560 570 580 590 600 610 620 630 640

the "blackout" situation used by Ferster (4), various CS's appear to be principally
both the key and box illumination were termin- in terms of vertical displacement, i.e.,
ated. The blackout condition was introduced
to facilitate the changing of the key color during total level of responding. The slopes
the stimulus generalization tests. of the generalization gradients, on the
Generalization testing was carried out under other hand, are very similar over the
extinction and was preceded by six 30-sec. inter- major portions of the curves and do
vals of responding to the CS during which three
reinforcements were administered. The wave
not appear to conform to the expecta-
lengths used in the generalization tests are given tion of marked changes corresponding
in Table 1. The 11 different stimuli were ran- to the transitions between spectral
domized within a series and 12 different random hues. The most conspicuous illustra-
series were presented to each S resulting in a tion of this point is the left half of the
schedule of 132 stimulus presentations. Twelve
different schedules were constructed for Ss given curve for 600 M/n which passes from
the 550 and 580 MM CS's and the same 12 orange through yellow to green. It
schedules were subsequently used for Ss given
the 530 and 600 M/J, CS's. Each stimulus pres- 450 |CS'6
entation was 30 sec. and was followed by a 10- l\
sec. stimulus-off interval. 400-
Following generalization testing, three addi-
tional APR sessions were given using the original 380 •
CS. After this a second generalization test GS;530 CS.580J
identical to the first was administered. During 300

both testings, the number of key pecks for each \CS.550 // ^

l\
290
30-sec. test stimulus presentation was recorded.
200
RESULTS
150
Characteristics of the generalization
gradient.—The mean generalization 100

gradients obtained in the first test are

shown in the upper portion of Fig. 1.
For each CS group (530, SSO, 580, and
600 MJW), the gradient was obtained
by plotting the mean total number of
responses for each test stimulus against
wave length in arithmetic units.
It will be seen that the various gen-
eralization gradients have highly
comparable forms. On either side
470 490 510 S30 550 570 590 610 630
of the CS, the rate of responding de- WAVELENGTH(Mu>
clines first in a nearly linear fashion,
FIG. 1. Upper: Mean generalization gradi-
and as the rate approaches zero, the ents, first test. Lower: Hue discrimination as a
curves become negatively accelerated. function of wave length for pigeons (adapted
The differences among the curves for from [5]), and humans (adapted from £1]).
82 NORMAN GUTTMAN AND HARRY I. KALISH

exhibits approximately the same for the two species appear to be

changes in slope as the other curves highly similar with minima and max-
which do not pass through as many ima at approximately the same wave
hues. lengths. Three characteristics of the
The similarity among the slopes of AX function might be reflected in the
the curves is further substantiated by generalization gradients. The first
tests of trend (Type I analysis of vari- is the absolute value of AX which
ance [7]) for the right and left halves determines the jnd interval between
of the gradients separately. For the any two values on the continuum,
left half of all gradients, over the and hence might be related to the
stimulus range from the CS to —40 slope of the gradient. The other two
M/t, the interaction between wave- characteristics are the first and second
length of test stimuli and CS groups derivatives of AX with respect to X,
was not significant (F = 1.02; 12, 80 which might be related to rate of
df), indicating that the hypothesis of change of curvature. For the pur-
parallel slope cannot be rejected over pose of the present analysis, we shall
this range. For the right side of these be concerned mainly with the absolute
gradients, over the stimulus range value of AX.
from the CS to +40 M/t, the inter- In terms of the discriminability
action was significant beyond the .1% function for the pigeon, the general-
level (F = 3.84; 12, 80 df). A fur- ization gradients for CS values of 530
ther analysis of the differences in slope and 580 M/t should be relatively sym-
between pairs of gradients on the metrical. For values near the CS,
right side revealed that the interaction however, the gradient for 580 MX
previously obtained resulted from a should be steeper than for any other
slope1 difference between the 600 and CS. The gradients centered at 550
580 M/t CS groups (F =5.10; 4, 40 and 600 M/t should be asymmetrical,
df; P < .005). The other slope differ- with 550 M/t being steeper in the
ences on the right side were not statis- direction of increasing wave length,
tically significant. and vice versa in the case of 600 M/t.
Evidence for differences in vertical In general, these expectations are
displacement of the gradients was also not borne out. Figure 1 and the pre-
furnished by the Type I analysis of ceding statistical analyses of slope
variance used to test slope. The F differences suggest that for the most
obtained for the mean differences in part the gradients are of uniform slope.
rate of responding for the left sides of The major exceptions to uniformity
the gradient is 5.55 (3, 20 df) which is of slope are the relative steepness
significant beyond the 1% level. The of the 600 M/t gradient and the flat-
corresponding F for the right sides is ness of the 580 M/t gradient, both
2.80 (3, 20 df) which is significant of which features are in the direction
between the 5 and 10% levels. The contrary to the hypothesis based on
mean differences on the right side are the AX function.
presumably reduced by the interaction The bidirectional symmetry of the
involving the 600 and 580 M/t CS gradients was examined by means of
groups. separate analyses of variance (A X
The lower portion of Fig. 1 presents B X S [7]) for each CS group. In
the spectral difference threshold (AX effect, one side of the gradient was
=
/M) for both pigeons (5) and superimposed on the other and a
humans (1) adjusted to the same ap- test of trend difference was obtained
proximate ordinates. The functions from the A X B (Right-Left Sides
 DISCRIMINABILITY AND STIMULUS GENERALIZATION 83

400

300

200
100

400

300

200-

100
FIG. 2. Mean generalization gradients,
second test.

X Test Stimuli) interaction. The

values entering into this analysis were 200-
obtained by subtracting response rate 100
for each test stimulus from the re-
sponse rate at the CS. The hypothe- 400
sis of symmetry could not be rejected 300
for the 530 and 580 gradients, since
200
the interaction F in each case was less
than 1.00. This appears to conform 100
to the AX hypothesis. On the other
-40 0 +40 -40 0 + 40 -40 0 +40
hand, however, significant asym- WAVELENGTH
metries contrary to the direction indi-
cated by the AX function are revealed FIG. 4. Individual generalization gradients
for 12 Ss representing various CS groups and
in the analyses of variance for the 550 levels of response strength. Solid lines desig-
and 600 Mju groups. The interaction nate the first test, broken lines the second test.
F for the 550 MH gradient is 8.39,
which is significant beyond the .1% level for 4, 20 df; the F for the 600
M/i gradient is 5.85, which is signifi-
cant beyond the 1% level for 3, 15
df*
The results of the second generaliza-
tion test, shown in Fig. 2, are highly
similar to those obtained in the first
test. In general, all gradients are
,2nd GENERALIZATION
reduced in height and the differences
TEST
in vertical displacement and slope are
attenuated. Neither the F for verti-
cal displacement nor the F for slope
difference is significant (P 5 .20).
The hypothesis of symmetry can be
rejected only for the 600 M/j. gradient,
since F = 3.44 (3, 15 df; .025 < P
.70 -60 -50 -40 .30 -20 -10 CS *IO »20 *30 +40 *60 *60
< .05). The relationships among the
WAVELENGTH [ M*) 3
The number of df is reduced for the 600 MM
FIG. 3. Mean generalization gradients, first and group because only four points to the right of the
second test, for all CS groups combined. CS were available for analysis.
-84 NORMAN GUTTMAN AND HARRY I. KALISH

FIG. 5. Mean generalization gradients for groups of eight Ss differing in

total number of responses, first test.

gradients for the second generalization tional gradient may be obtained from
test are such that the curves can be any one series of 11 stimulus presenta-
nearly superimposed by translating tions. A second interesting feature
them along the wave-length axis to a of these curves is the extent to which
common CS point. each S produces a gradient in the
The relationships between the mean second test which is almost a replica
generalization gradient for all groups of the first.
in the first test and the mean for all An examination of the individual
groups in the second test are shown in gradients suggests that the averaged
Fig. 3. gradients in Fig. 3 are not entirely
Individual generalization gradients. representative of the generalization
—Twelve pairs of individual general- phenomenon for the single S. For
ization gradients for the first and sec- some Ss the curves are bilaterally
ond tests are shown in Fig. 4. These convex, for some, concave, and for
gradients are arranged in order of others, concave on one side and con-
ascending response rate, and for each vex on the other. Certain Ss, such as
CS group the Ss presented are the Animal E (Fig. 4, second test), ex-
lowest, the highest, and the S nearest hibit linearity over a major portion of
the group median. Perhaps the most the gradient. The linearity observed
salient aspect of these curves is the over the central range of the averaged
orderliness and reproducibility of the curve (Fig. 3) may well be the result
generalization process within the in- of a random distribution of concavi-
dividual S. Although these gradients ties and convexities.
are summed over the individual's It may also be noted in Fig. 4 that
schedule of 132 stimulus presentations, certain Ss show maximum responding
it is interesting that a similar bidirec- at a point either ± 10 M/u from the CS.
 DISCRIMINABILITY AND STIMULUS GENERALIZATION 85

In some instances this contributes to Additional information concerning

the departure from symmetry of the changes in the generalization gradient
averaged group curves (Fig. 1). associated with changes in level of
The generalization gradient as a responding may be obtained from an
function of differences in response analysis of the extinction process oc-
strength.—The averaged gradients in curring during the generalization test.
Fig. S were obtained by dividing the The generalization gradients-in Fig. 6
24 Ss (without respect to CS group) are based upon the total responses of
into three relatively homogeneous sub- all Ss and represent the average per-
groups in terms of total responses for formance for successive blocks of 33
the first generalization test. As the stimulus presentations. The gradient
total response level is reduced, the for the first quarter of the extinction
generalization gradient becomes uni- series is relatively steep, and the gradi-
formly flatter. Tests of trend differ- ents for the succeeding quarters be-
ences over the right and left halves of come progressively flatter. These
the gradients separately permit re- changes bear a strong resemblance to
jection of the hypothesis that the those shown by the response-strength
curves are parallel. On the right subgroups in Fig. 5. Furthermore,
side, the interaction F — 22.4 (8, 84 the gradients for the successive stages
df; P < .001) while F for vertical dis- of extinction (Fig. 6) display a de-
placement is 19.8 (2, 21 df; P < .001). parture from parallelism on both sides
The interaction F for the left halves is of the CS. The interaction F's for
5.15 (8, 84 df; P < .001), while the the right and left halves of the gradi-
vertical displacement F — 19.5 (2, ents are 8.22 and 6.40, both significant
21 df; P < .001). In a later section beyond the .1% level for 12, 276 df.
we shall analyze in greater detail the The similarities between the changes
nature of these interactions. in the form of the generalization
I20i

FIG. 6. Mean generalization gradients for successive fourths of first test.

86 NORMAN GUTTMAN AND HARRY I. KALISH

40 80 120 200 240

TOTAL RESPONSE RATE
RESPONSES PER WIN.
FIG. 7. Rate of responding for various test stimuli as a function of total response strength.

gradient during extinction (Fig. 6) spending (in responses per minutej

and the changes associated with indi- was determined for each gradient.
vidual differences in response strength In Fig. 7 the values from each
(Fig. 5) strongly suggest that both unidirectional gradient are plotted
sets of changes may be represented by on ordinates erected at points cor-
a common function. In order to ob- responding to the total response
tain a graphical comparison of these strength for that gradient. The lines
changes, each bidirectional gradient in this figure connect the same test
for the first generalization test in Fig. stimulus values for the various gradi-
5 and 6 was first converted to a uni- ents. Thus, the uppermost pair of
directional gradient by averaging the lines connect the CS values for the
response rates for ±10 M/u, ±20 M^u, gradients in Fig. 5 and 6, the next
etc., and converting to a common pair of lines connect the values for
scale of responses per minute. Next, ±10 Mj«, etc. These lines may be
an index4 of the total rate of re- visualized as the contours of the sur-
4
The index used was
faces in Fig. 5 and 6. The high de-
where R = rate of responding in responses per
minute. This index provides that the values
for the CS will lie along a line with a 0,0 inter-
cept, and permits a rescaling of the total ex-
tinction curve on a response strength axis.
 DISCRIMINABILITY AND STIMULUS GENERALIZATION 87

gree of correspondence between these judgment. In more formal terms, the

lines indicates that changes in general- analogy we have suggested between the
ization as a function of response generalization decrement and the differ-
strength are the same whether the ence limen may be misleading, because
response strength varies within the p(R) in one situation refers to a simple
conditioned response, while in the psy-
individual, as in the course of extinc- chophysical experiment, p(R) is associ-
tion, or whether response strength is ated with a complex of discriminatory
considered as a parameter character- responses.
izing the differences between indi- The foregoing considerations, however,
viduals. fall far short of excluding the possibility
The transformation of the data to that a relationship does exist and that
the form shown in Fig. 7 suggests the the present experiment has failed to dis-
mathematical operation necessary to close it. Our results are compatible
predict generalization at one level of with the notion that the wave-length dis-
response strength given the general- criminability function for the pigeon is
much flatter than has been reported (5),
ization function at a different level. flatter in fact than that for humans. It
The operation proposed by Hull (6) is may also be that the discriminability
multiplication by a constant. Miller differences in our situation are attenu-
(9) has suggested, in the context of a ated because the pigeon works in a dim
discussion of the effect of motivational light. As a check on the latter possibility
changes on the gradient, that the con- a crude psychophysical experiment was
version can be accomplished by add- performed on four human tfs using the
ing a constant at every point on the method of minimal changes. The AX
gradient. If the multiplicative rela- function obtained was not appreciably
tion holds, then all the lines in Fig. 7 different from that generally reported; it
showed minima in the blue and yellow
should pass through the point 0, 0, but regions, a maximum in the green, and
if the additive relation obtains, then differences on the order of 5:1. The
all the lines in this figure should be same results were obtained in the dark
parallel to each other. The present and in the presence of dim light.
results indicate that neither hypothe- Another factor in our experiment
sis alone provides a sufficient de- which might act to distort the expected
relationship is brightness variation, since
scription of the situation. The addi- no attempt was made to equate the test
tive relation appears to hold for values stimuli for brightness or incident energy.
near the CS, but the multiplicative This might have the consequence of in-
relation must be invoked for values creasing the slope of gradients passing
further removed from the CS. into regions of lower visibility, and
might account for the asymmetry of the
gradient for 600 M/J. If, however,
DISCUSSION brightness were an important factor over
The results of the present experiment all, it would be difficult to account for the
do not appear to provide evidence for a high degree of similarity among the
correspondence between the discrimin- gradients plotted on a wave-length con-
ability function for wave length of light tinuum. The extent to which brightness
(AX vs. X) and the generalization process. can be invoked as a variable is inversely
Since some relationship is reasonable on proportional to the degree of relationship
logical and intuitive grounds, our results between changes in response rate and
must be interpreted in the light of a wave length.
number of possibilities. In the first Finally, serious consideration must be
place, the relationship may not, in fact, given to the possibility that the relation-
exist, because during generalization test- ship between generalization and dis-
ing the S is not making a comparative crimination is much more complex than
 NORMAN GUTTMAN AND HARRY I. KALISH

we have supposed. Before any general parable forms for the various spectral regions
statement can be made concerning the tested. The results open the possibility of an
independence of discriminability and independence between the generalization decre-
stimulus generalization, the problems ment and the discriminability of stimuli.
Changes in the generalization gradient accom-
raised by the present experiment must panying changes in response strength were in-
be explored with other continua, species, vestigated. A close correspondence was found
and techniques. between changes in the form of the gradient
With respect to the issue of changes in during extinction and changes associated with
the form of the generalization gradient individual differences in response strength. In
with changes in response strength, our both cases the changes appear to be describable
findings may be compared most directly by means of a combined additive and multiplica-
with those of Margolius (8), who investi- tive relationship.
gated changes in generalization as a func-
tion of number of conditioning trials. REFERENCES
In terms of the rate of responding meas- 1. BORING, E. G., LANGFELD, H. S., & WELD,
ure, Margolius found increases in the H. P. Foundations of psychology. New
absolute and relative amounts of gen- York: Wiley, 1948.
eralization as the number of training 2. BRUSH, F. R., BUSH, R. R., JENKINS, W. O.,
trials increased. Margolius' interpreta- JOHN, W. F., & WHITING, J. W. M.
Stimulus generalization after extinction
tion of his findings is equivalent to what and punishment: an experimental study
we have designated as the additive rela- of displacement. /. abnorm. soc. Psy-
tion. It will be recalled that for the chol, 1952, 47, 633-640.
present results, the additive hypothesis 3. FERSTER, C. B. The use of the free operant
was tenable for the major area of the in the analysis of behavior. Psycho!.
curve including values near the CS, and Bull., 1953, SO, 263-274.
high response strengths, but for remote 4. FERSTER, C. B. Sustained behavior under
values of the stimuli and for low response delayed reinforcement. /. exp. Psychol.,
strengths the additive .relation fails. 1953, 45, 218-224.
This suggests the possibility that the 5. HAMILTON, W. F., & COLEMAN, T. B. Tri-
chromatic vision in the pigeon as illus-
empirical gradient is comprised of two trated by the spectral hue discrimination
components, one which is linear on a curve. /. comp. Psychol., 1933, IS, 183-
wave-length scale and changes addi lively, 191.
and another which is curvilinear and 6. HULL, C. L., Principles of behavior. New
changes multiplicatively. It may be York: D. Appleton-Century, 1943.
that the second component is contributed 7. LINDQUIST, E. F. Design and analysis of
by random responding and responding to experiments in psychology and education.
New York: Houghton Mifflin, 1953.
extraneous stimuli accompanying the 8. MARGOLIUS, G. Stimulus generalization of
beginning of a trial. This separation of an instrumental response as a function of
components implies that if the contribu- the number of reinforced trials. /. exp.
tion of the second component was re- Psychol., 1955,49,105-111.
duced, either experimentally or analyti- 9. MILLER, N. E. Experimental studies of
cally, the additive relation would become conflict. In J. McV. Hunt (Ed.),
more conspicuous. Personality and the behavior disorders.
New York: Ronald, 1944.
10. SCHLOSBERG, H., & SOLOMON, R. L. Lat-
SUMMARY ency of response in a choice discrimina-
tion. /. exp. Psychol, 1943,33, 22-39.
An experiment was performed to test the 11. SKINNER, B.W. Are theories of learning
hypothesis of a relationship between the dis- necessary?' Psychol. Rev., 1950, 57, 193-
criminability of spectrum colors and stimulus 216.
generalization along the wave-length continuum. 12. SPENCE, K. W. The differential response in
Four groups of six pigeons were trained to peck animals to stimuli varying within a single
at a disc illuminated by monochromatic lights dimension. Psychol. Rev., 1937, 44, 430-
corresponding to 530, 550, 580, and 600 Mp. 444.
Bidirectional generalization gradients were ob-
tained from measures of response rate during ex- (Received for early publication
tinction. These gradients were of highly com- August 8, 1955)