PSYCHOLOGICAL SCIENCE

Research Article
REPRESENTATION OF COLORS IN THE BLIND, COLOR-BLIND, AND
NORMALLY SIGHTED
Roger N. Shepard 1 and Lynn A. Cooper2
IStanford University and 2Columbia University

Abstract-Human adults with normal vision, with three types of significant ways have emerged. The first is a follow-up study by
color-blindness, or with complete absence of vision since birth Marmor (1978), which presented our nine color words to a
rank-ordered the similarities of all pairs of colors correspond- larger sample of blind individuals, including cases oflate as well
ing to nine hue names. When presented with the names only, as early blindness. The second is an independent Soviet study
subjects with any color vision produced rankings for which mul- reported in the accompanying article by Izmailov and Sokolov,
tidimensional scaling yielded Newton's color circle. When sub- in which normally sighted subjects first learned arbitrary asso-
jects were presented with the colors themselves, the recovered ciations between artificial words and hues around the color cir-
color circle remained the same for the normally sighted but cle and then judged the similarities among the colors when only
collapsed along the red-green dimension for the color-blind. their newly associated names were presented. This publication
Based on their rankings by color names, the totally blind sub- of Izmailov and Sokolov's results seemed to us a fitting occa-
jects all fell outside the range of the color-normal subjects but sion for a fuller presentation of our own earlier findings, to-
partly overlapPl!d the color-deficient subjects; a rare rod mono- gether with some comparisons between our findings, theirs, and
chromat roughly approximated the color-normal subjects. those of Marmor.
These results, ~long with those ofMarmor (1978) and Izmailov
and Sokolov (this issue), suggest how visual experience, lan- c·
guage, and innate structure contribute to the mental represen- THE PROBLEM OF COLOR REPRESENTATION
tation of colors.
The hues of the rainbow correspond to a one-dimensional
continuum of physical wavelengths. Yet the red and violet ex-
Colors, subjectively so simple and vivid and yet objectively
tremes of this visible continuum appear more like each other
so difficult to describe or to reduce to physical terms, have long
than either appears like an intermediate hue (e.g., green). Thus,
raised puzzling questions for scientists, philosophers, and lay-
the physically rectilinear continuum of wavelength gives rise to
persons: Do colors exist in the external world, or only in the
a psychologically circular continuum of hues-the color circle
mind of the beholder? How do colors appear to color-blind
described by Newton (1704). Nonmetric multidimensional scal-
individuals? Does your experience of red have the same sub-
ing, which finds the spatial configuration of points such that
jective quality as my experience of red? What ideas about col-
stimuli judged to be more similar correspond to points that are
ors develop in persons who have been totally blind since birth?
closer together (Shepard, 1962a, 1962b; see also Kruskal,
Partly motivated by such questions, during 1969 and 1970,
1964a), can recover this color circle merely from a subject's
we collected similarity data for nine hues around the color circle
similarity ordering of all pairs of n presented hues (Shepard,
from people with normal vision, with three types of color-
1962b, Fig. 13; Shepard, 1980).
blindness, and with complete absence of vision since birth. In
Mixtures of different wavelengths give rise to additional, less
keeping with an ongoing,program of research on "second-order
saturated colors not falling on the circle of pure spectral hues.
isomorphism" between internal representations and the exter-
Indeed, because the amount of light at each wavelength can be
nal things they represent (Shepard, 1975, 1978b; ~hepard &
independently varied, the physical degrees of freedom of light
Chipman, 1970; Shepard, Kilpatric, & Cunningham, 1975), we
are essentially unlimited. Yet, human observers, having only
asked the subjects to judge the similarities of colors when only
three types of color-sensitive retinal cones, can always match
the names of the colors were presented and (except for the blind
the resulting appearance by adjusting no more than three vari-
subjects) when the corresponding colors themselves were pre- ables on a suitable color-mixing apparatus-variables corre-
sented. We summarized our results at a scientific meeting sponding to brightness, hue, and saturation, or, alternatively, to
(Shepard & Cooper, 1975) and in some published overviews of opponent-process dimensions of Iight-versus-dark, blue-versus-
work on mental representations (Finke & Shepard, 1986, pp. yellow, and red-versus-green (Hering, 1878/1964; Hurvich &
37-5-37-7; Shepard, 1975, p. 97). With our ensuing engagement Jameson, 1957). Multidimensional scaling can also recover the
in research on mental transformation (reviewed in Cooper & implied three-dimensional color space of the color normal
Shepard, 1984; Shepard & Cooper, 1982), however, we never (Indow, 1988; Indow & Kanazawa, 1960; Shepard, 1978a), as
published a full report of our now 20-year-old investigation into
well as the differentially compressed or lower dimensional color
the representation of colors. spaces of the color deficient (Carroll & Chang, 1970b; Wish &
In the meantime, two studies that complement our work in Carroll, 1974).
But how are colors represented in individuals who are not
Address correspondence to Roger N. Shepard, Department of Psy- actually perceiving those colors but only imagining them? And
chology, Building 420, Stanford University, Stanford, CA 94305-2130; how are colors represented or imagined in individuals whose
e-mail: [email protected]. visual experience has been impoverished or absent since birth,

VOL. 3, NO.2, MARCH 1992 Copyright © 1992 American Psychological Society 97

Downloaded from pss.sagepub.com at SETON HALL UNIV on April 1, 2015
 PSYCHOLOGICAL SCIENCE

Representation of Colors

as a result of altered or missing retinal receptors? Such ques- color samples in the MUflsell Book of Color (1976) that best
tions bear on epistemological debates between empiricist and match our selected red, orange, gold, yellow, green, turquoise,
nativist philosophers. Even Hume departed from the strict em- blue, violet, and purple Color-aid papers are designated 5R
piricist maxim that all ideas initially enter the mind through the 4/14, lOR 5/14, 8.75YR 7/14, 2.5Y 8.5/12, 7.5G 4/10, lOBG 5/8,
senses when he admitted that a person who had experienced all 6.25PB 3/12, 5P 3/8, and lOP 3/10, respectively. These colors
colors "excepting one particular shade of blue" might never- necessarily were not equated in lightness; the prototypical color
theless be able to imagine that never-experienced color (Hume, corresponding to yellow, for example, is appreciably lighter
1739/1896, p. 6). than the prototypical colors corresponding to red, green, blue,
To investigate the representation of colors in persons with or purple (Berlin & Kay, 1969, p. 119).
widely differing sensory experiences of colors, we employed a We then prepared four decks, each containing 36 (7.6 cm x
method (introduced by Shepard & Chipman, 1970) in which 12.7 cm) white cards corresponding to the 36 pairs of distinct
subjects judge the similarities between stimuli under two con- colors (randomly assigned to the left and right positions on each
ditions: when those stimuli are actually presented and when card). For a Colors-only deck, we mounted on each card (with
those stimuli are only named. At about the same time, Fillen- a separation of2.5 cm) two 3.8-cm squares cut from the corre-
baum and Rapoport (1971; see also Rapoport & Fillenbaum, sponding colored papers. For two Names-only decks, we in-
1972) presented only the names of colors to normally sighted stead typed on each card only the names of the pair of assigned
subjects and, using multidimensional scaling, obtained color colors-in black letters for the deck to be presented to the
circles resembling those obtained by other investigators for sighted subjects and in raised braille characters for the deck to
physically presented hues (Carroll & Chang, 1970b; Helm, be presented to the blind subjects. For a Names + Colors deck,
1964; Indow &, Uchizono, 1960; Shepard, 1962b; also Schnei- we affixed on each card the two assigned color squares and
der, 1972). Our experiment was, however, the first directly to directly under each color, put the corresponding printed name:
compare simila'rity data for color names and for actual colors
selected to correspond one-to-one to those same names, and the
first to include blind and color-blind subjects, along with the Procedure
normally sighted. Each subject was given a shuffled deck and asked to rear-
range the cards until satisfied that all 36 cards were ordered
according to the similarity between the two colors displayed or
METHOD
named on each card. All subjects with any color vision did this
for the three decks: Names-only, Colors-only, and
Subjects Names + Colors, in that order. The monochromat did this for
We recruited 37 subjects, primarily through notices circu- just the Names-only and (subsequently) the Names + Colors
lated at Stanford University and through a local agency provid- decks, and the six blind subjects did this for just the braille
ing aids for the blind. Based on the responses of the sighted Names-only deck. Thus, only the Names-only condition was
subjects to the test plates in Ishihara (1977), we classified the presented to all subjects, and it was always presented first.
subjects into six groups: (a) 14 adults with normal trichromatic (Other studies have not found an appreciable effect of order of
color vision, (b) 7 with color deficiency of the strong deutan presentation of perceptual and names-only conditions; see
type, (c) 4 with color deficiency of the strong protan type (see Shepard & Chipman, 1970; Shepard et aI., 1975.)
Farnsworth, 1947, and Ishihara, 1977, concerning these two
types), (d) 1 with rare monochromatic vision (a female adi.I1t
RESULTS
with rod vision only, referred to us by Barbara Sakitt), (e) 6
with total blindness since immediately after birth (typically as a . The data for each subject in each condition take the form of
result of retrolental fibroplasia), and (0 5 others with miscella- a 9 x 9 triangular half-matrix in which each of the 36 cells
neous weaker color anomalies, who are not included in the contains the rank (between 1 and 36) of the similarity of the
analyses presented here. colors corresponding to the row and column of that cell. The
analyses we present are based on 83 such matrices, correspond-
ing to 25 subjects with color vision x 3 conditions, I mono-
Stimuli chromat x 2 conditions, and 6 blind subjects x 1 condition.
As familiar names of spectral hues, we selected the nine
words red, orange, gold, yellow, green, turquoise, blue, violet, Analyses of Correlations Among
and purple. For each of these names, we then selected the
Subject-Condition Combinations
Color-aid paper that a small preexperimental group of judges
with normal color vision most often chose as the best exemplar For initial analyses of the similarities and differences among
for that name. Moreover, for the six of our names that qualify types of subjects (normal, deutan, protan, monochromat, or
as basic color terms (Berlin & Kay, 1969}--red, orange, yellow, blind) and conditions (Names-only, Colors-only, or
green, blue, and purple-the colors we selected fall within the Names + Colors), we computed an average matrix for each
regions of Munsell color space that other normally sighted na- combination of type of subject and condition (without, how-
tive speakers of American English have accepted as prototyp- ever, the monochromat's Names + Colors condition, which we
ical exemplars of those colors (Berlin & Kay, 1969, p. 119). The added only later). We then computed the product-moment cor-

98 VOL. 3, NO.2, MARCH 1992

Downloaded from pss.sagepub.com at SETON HALL UNIV on April 1, 2015
 PSYCHOLOGICAL SCIENCE

Roger N. Shepard and Lynn A. Cooper

relation between corresponding cells for each pair of the result- the program KYST-see Kruskal, voung, & Seery, 1973; Shep-
ing (3 + 3 + 3 + 1 + I) matrices, yielding an 11 x 11 sym- ard, 1980, note 21), with the already mentioned hierarchical
metric matrix of correlation coefficients. clustering embedded as a set of nested curves. (The very same
Application of the hierarchical clustering method HICLUS clusters emerged from the "maximum" and "minimum" vari-
(Johnson, 1967) to this correlation matrix revealed that the 11 ants of HICLUS-indicating that the obtained clusters are
subject-condition combinations fell into three disjoint groups: strongly determined by the correlation data.) The number as-
one for the subjects having any color vision, one for the mono- sociated with each cluster is the smallest correlation between
chromat, and one for the completely blind subjects. Indeed, any two enclosed combinations of subject type and condition.
regardless of whether names, colors, or names plus colors were The remaining two, increasingly inclusive clusters (not shown)
presented, all 36 correlations between matrices within-the color added, first, the monochromat (with a minimum correlation of
vision group (ranging from .802 to .989) were larger than any of .74) and, finally, the blind subjects (with a minimum correlation
the 19 correlations between matrices divided between any two of .53).
of these three groups (ranging from .534 to .796). This two-dimensional scaling solution is nondegenerate, pro-
Because the similarity rankings fell into these three disjoint vides a good fit to the correlation data (stress = 8.9%), and is
groups, nonmetric multidimensional scaling (Kruskal, 1964a, readily interpretable. For the normal trichromats, the three con-
1964b; Shepard, 1962a, 1962b) yielded a degeneracy (see Shep- ditions (labeled N, C, and NC in ~he upper left) all intercorrelate
ard, 1962b, 1974) in which the 11 subject-condition combina- at least .97 and, hence, are tightly clustered. As in similar stud-
tions collapsed onto the vertices of an equilateral triangle, with ies using other stimuli-including visual shapes, spoken words,
each of the three groups at a different vertex. Any pattern of faces, and odors (e.g., Shepard, 1975, 1978b; Shepard & Chip-
individual differences within the color vision group was lost in man, 1970; Shepard et a!., 1975)-the similarity comparisons
its collapse to a point. We therefore applied non metric multidi- seem to have been based on internal representations that func-
,mensional scaling io just the 9 x 9 correlation matrix remaining tioned equivalent.1Y whether the colors were perceived or only
after deletion of the rows and columns for the monochromat imagined. .,
and for the group of blind subjects. As might be expected from the fact that the deutans and
Figure 1 shows the two-dimensional solution (obtained by protans have a reduced ability to discriminate between reds and
greens, their points for the conditions in which the colors were
actually presented (C and NC) are far removed from the region
of the color-normal subjects. Remarkably, however, when
Normal Deutans Protans these color-deficient subjects were presented with the names
Trichromats only, their points (N) remained relatively closer to the points for
c-o
--
>- the color-normal subjects. These subjects knew that red and
c:
o green appear very dissimilar to most people. Vet the proximity
ctl til
of the points labeled NC to those labeled C (rather than to those
c: C1.I
Q) E labeled N) indicates that when the names were presented along
(f) llJ
....
Q) Z with the colors, the color-deficient subjects based their similar-

-
a.
o
c:
ity judgments on their immediate visual experience of those
colors more than on their knowledge of the relations among

-
those colors for color-normal observers.
o
1Il
"'0 ~
c: o Analyses of Similarity Rankings From Subjects With
o
-
u
C'J
o
u Color Vision
The two-dimensional solutions obtained by separately apply-
c:
o ing multidimensional scaling (KVST) to the similarity iankings
'in from all subjects in the different subject-condition combina-
c: Key
Q)
tions reinforces this interpretation. On the left in Figure 2, we
E ® Names only show the four solutions based on the data from the 14 color-
is © Colors normal subjects (at the far left) and from the 11 color-deficient
® Names + Colors subjects (in the middle), both for the Names-only condition
(above) and for the Colors-only condition (below). In agreement
Dimension 1 type of subject) with our interpretation of Figure 1, Newton's color circle
clearly emerged for the color-normal subjects in both conditions
and (except for a reversal between the two encircled closest
Fig. 1. Nonmetric multidimensional scaling solution, with em-
neighbors, violet and purple) for the color-deficient subjects in
bedded hierarchical clustering, based on the correlations among
the averaged similarity rankings for nine combinations of type the Names-only condition. Despite their perceptual deficit, the
of subject (normal, deutan, or protan) and condition of presen- color-deficient subjects have an accurate conceptual represen-
tation (Names-only, Colors-only, or Names + Colors). The tation of the relations among the nine colors, just as the color-
number inscribed within each cluster is the smallest correlation normal subjects do. In the Colors-only condition, however, the
between any two combinations in that cluster. color circle collapsed for the color-deficient subjects, bringing

VOL. 3, NO.2, MARCH 1992 99

Downloaded from pss.sagepub.com at SETON HALL UNIV on April 1, 2015
 PSYCHOLOGICAL SCIENCE

Representation of Colors

Normal TrichromaLs DeuLans & ProLans Monochromat

Names Only Names Only Names Only

Stress: Stress:
stress:
13.9 % for individuals 17.7% for individuals 9.4% for the
(2.4% for the group) (2.6 % for the group) one individual

Three Blind Subj~ets

(most correlated with sighted)
Names Only
Stress:
17.8 % for individuals
(7.6 % for the group)

Colors Only Colors Only

Stress: Stress:
13.3% for individuals 19.6% for individuals
(0.9 % for the group) (3.4 % for the group)

Fig. 2. Nonmetric multidimensional scaling representations for the nine colors-red (R), orange (0), gold (Go), yellow (Y), green
(Gr), turquoise (T), blue (B), violet (V), and purple (P)-based on average similarity rankings: (a) for the normal trichromats (left)
and the color-deficient subjects (middle) under the Names-only condition (above) and the Colors-only condition (below), (b) for
the monochromat under the Names-only condition (upper right), and (c) for the three blind subjects who correlated most highly
with the sighted subjects under the Names-only condition (lower right). (Stress Formula 1 was minimized in all cases.) The first
stress value given for each group of subjects was computed from the rankings of individual subjects in that group and should be
roughly comparable across the groups of different sizes. The second (parenthetical) stress value for each group was computed from
the more stable similarities averaged over all subjects in the group as a whole and, hence, tends to be smaller for the larger groups.
The smooth curve drawn through the nine points in each solution indicates the degree of approximation to the standard color circle.
Points in permuted orders on the curve are encircled.

the red and green sides. of the circle together. In fact, such a SCAL estimates weights that (when squared) indicate how
C-shaped configuration indicates a close approximation to uni- much of the variance of each input matrix is explained by each
dimensionality (Shepard, 1974) and hence (if we allow for vari- of the orthogonal dimensions of the spatial solution. .
ation of lightness as well as color) to dichromacy. . For the first of our two such analyses, we computed an
average similarity matrix for the six blind subjects, for the sin-
gle monochromat, and for each combination of type of subject
INDSCAL Weights for Different
having color vision (normal, deutan, or protan) and ~ondition
Subject-Condition Combinations (Names-only, Colors-only, or Names + Colors). The two-
To obtain a spatial representation of individual differences dimensional INDSCAL solution accounted for 88.0% of the
that includes the blind subjects and the rod monochromat, we variance in the resulting II averaged matrices. The nine colors
analyzed the similarity matrices for all 32 subjects by the in the group stimulus space formed a color.circle (see lower left
INDSCAL method of multidimensional scaling (Carroll & inset in Fig. 3) resembling the three color circles in the left and
Chang, 1970a). INDSCAL, being a metric method, uses more top center of Figure 2. Moreover, the horizontal and vertical
than the ordinal properties of the data, but is less susceptible to axes-which in an INDSCAL solution should be interpretable
degeneracy than are nonmetric methods. In addition, IND- without rotation (Carroll & Chang, 1970a)-were immediately

100 VOL. 3, NO.2, MARCH 1992

Downloaded from pss.sagepub.com at SETON HALL UNIV on April 1, 2015
 PSYCHOLOGICAL SCIENCE

Roger N. Shepard and Lynn A. Cooper

1.0..-........- - r - - r -........--r--r-"""T--r--r---,- - - - - - -- 1.0.---r--.-r----r-.--r-----.--,·...--r---·-.- - - - - - --

Weight Space Weight Space
0
"""
~ .8
• No color vision . . ' ¢ Color vision
\ ,, """
~ .8
:y/:
:(~:
Blind •.•.•.•
••••
:HH~:
Normal
,
c.> ell
to. \ t.. subjects Monochromat' Trichromats
C'l . C'l
I I
~",~, x 2: ••• 1•••• / ( 1 4 x 3)
r-G-;.~-~p-Str~~i~;-sp-;;;' \ , ~ (6X I) (l
~ .6
I '..9...
1 :6k:!;·,. /
~
t:lormal c.> .6
....,
to.
.(D... \
c ....Trlchromats
....,
t..
", .:::::::::.... .' T
N @5" t0r···. , ~ N ,..,~..
..... :~::H~ MH!:~; .• ~I
• .. :.;::::::::~ ... ~. [).. Protans
t

5.4
til 1:'"0 j y~::::""'~y pl':lT~\~N.... Oeutans
5.4
~~.~~)IJ.II.iij,;:~'
~ 1"}{':4X3:
®.... ~
• V-' til
c N
c
c.> !fij ~"O~ c.>
E
._ .2 .... ~ .=t; c E ~~:~~~ }!.:
'0. •••• Deutans
'~{,. :._.&~g~l" (7X3)
C
'. '" .' u-ClI .- .2
o ......
....®.....
..!95 go :to:c \\'\protans o
Lo -' 1:
:l
(f) \
•
.•••• :i';X<~~ 0 •
••••••~. -(COlors):

'~O -------- • • •• ..l-.,!.. _

.2 .4 .6 .8 1.0 '~O .2 .4 .6 .8'" ., 1.0
Dimension 1 (blue-yellow) Dimension 1 (blue-yellow)

Fig. 3. Weights on i Dimension 2 (interpreted as the red-green Fig. 4. Weights on Dimension 2 (interpreted as the red-green
dimension) plotted against weights on Dimension 1 (interpreted dimension) plotted against weights on Dimension 1 (interpreted
as the blue-yellow dimension) for the 11 combinations of type of as the blue-yellow dimension) for all 83 combinations of indi-
subject and condition, from an INDSCAL analysis of the aver- vidual subject and condition, from an INDSCAL analysis of the
age ranked similarities for those 11 combinations. Subject types ranked similarities for those 83 combinations. Each dotted con-
with and without color vision are represented by white and vex curve encloses all points for the indicated type of subject.
black circles, respectively. The inset in the lower left shows the Multiplication of the number of subjects of that type by the
color circle that emerged in the group stimulus space. number of conditions for that type-the two parenthetically in-
dicated numbers-yields the number of points enclosed in that
identifiable with the blue-yellow and red-green opponent- curve.
process dimensions of human color vision (Hering, 1878/1964;
Hurvich & Jameson, 1957). subject-condition combinations. Individual points are plotted
The two-dimensional weight space for the 11 combinations only for the six blind subjects and for the monochromat's two
of subject type and condition, which constitutes the main part conditions (N and NC). The 76 points for subjects with color
of Figure 3, is also readily interpretable. The three points for the vision are too crowded to be shown separately, but we indicate
normal" trichromats cluster together in the upper right of the the subregions to which they are confined by convex dotted
positive quadrant, where both dimensions are heavily weighted. curves that enclose all points for each type of subject.
The deutans and the protans are displaced, as expected, in the Again, the normal trichromats all have high weights on both
downward direction of lower weights on the red-green dimen- dimensions, While the color-deficient subjects have generally
sion. As in Figure 1, however, the color-deficient subjects re- lower weights on the red-green dimension. The individual
main closer to the color-normal subjects in the Names-only points are, of course, more scattered here than the points based
condition than in the other conditions. on averaged matrices in Figure 3. Although the points in the
As indicated by the slanting broken line, the blind subjects upper portion of the deutan region overlap with the normal
and the monochromat are linearly separable from all points for trichromat region, these points are from the deutans' Names-
subjects with color vision (whether normal or deficient), and are only condition, while the points in the lower, nonoverlapping
closer to the origin (0, 0) at the bottom left corner of the quad- portion of the deutan region are from the conditions in which
rant, where no variance would be accounted for by the color the colors themselves were present. As before, the protan re-
circle. Overall, the variances accounted for were 96% for the gion has less vertical spread than the deutan region. (The wider
normal trichromats, 93% for the deutans, 88% for the protans, separation in Fig. 1 between the colors-present and colors-
71% for the blind subjects, and 68% for the single monochro- absent conditions for the protans may reflect differences on
mat. other dimensions; see Chang & Carroll, 1980.)
Our second INDSCAL analysis was applied to the whole set The points for the subjects having no color vision are much
of 83 individual matrices. The two-dimensional solution ac- more widely scattered. Among the six congenitally blind sub-
counted for 77% of the variance of these more numerous but jects, three fall far to the left, with greatly reduced weights on
individually less reliable matrices. A color circle virtually iden- the blue-yellow dimension, and with variances accounted for of
tical to that shown in the lower left inset in Figure 3 emerged in only 22.9%, 23.0%, and 23.2%; the other three fall within the
the group stimulus space, with the horizontal and vertical di- lower (Colors-only) region of the color-deficient subjects, with
mensions again corresponding to the blue-yellow and red-green variances accounted for of 58.7%, 67.2%, and 70.3%. Even
opponent processes. Figure 4 shows the weight space for the 83 these latter three blind subjects, however, have no overlap with

VOL. 3, NO.2, MARCH 1992 101

Downloaded from pss.sagepub.com at SETON HALL UNIV on April 1, 2015
 PSYCHOLOGICAL SCIENCE

Representation -of Colors

the convex region containing all 42 points for the normal tri- included a larger sample of blind 'subjects and, also, subjects
chromats. who had become blind only later in life. The results for her 16
Among the subjects without color vision, only the rod mono- late-blind subjects were intermediate between those for her 16
chromat overlaps the normal trichromats' region-and this oc- early-blind and her 16 normally sighted subjects. Apparently,
curs only for the monochromat's Names-only condition. The individuals who have become blind, even after having had con-
point for her Names + Colors condition is again far to the left. In siderable visual experience, have not assimilated or have not
the face of her own, different sensory experience, when the retained the similarity relations among colors as fully as indi-
colors were visually in front of her, the monochromat, like the viduals who continue to have sight.
color-deficient subjects, evidently made little use of the con- Marmor's relatively large sample of early-blind subjects in-
ceptual knowledge she had acquired about the relations among cluded some whose multidimensional scaling results approxi-
colors. mated Newton's color circle more closely than the results for
our six blind subjects. (Compare the quasi-circular configura-
Analyses of Similarity Rankings From Subjects tions on the right in Marmor's Fig. 1 and top right in her Fig. 3
Without Color Vision ' with the distorted configuration in the lower right of our Fig. 2.)
Even for Marmor's most successful blind subjects, however,
We applied the multidimensional scaling program (KYST), neighboring hues (such as yellow and gold, or violet and purple)
separately, to the average data (a) for the three blind subjects were sometimes reversed. Overall, moreover, the multidimen-
whose INDSCAL weights overlapped those of the sighted sub- sional scaling results indicate just as marked a difference be-
jects, (b) for the three blind subjects whose weights fell outside tween the blind and sighted subjects in Marmor's study as in
those of the sighted subjects, (c) for the monochromat under the ours:
Names-only co~dition, and (d) for the monochromat under the
Names + Colors condition. The two-dimensional solutions for 1. The variances accounted for by the two-dimensional
the two of these' four cases (viz., b and d) for which the subjects INDSCAL solutions (though estimated in different ways in
most differed from subjects with color vision appeared quite the two studies) were consistently smaller for the early-blind
chaotic and bore little resemblance to Newton's color circle. group than for the sighted group-in our study, 71% for the
Some colors that are very different for the normally sighted blind versus between 95% and 97% for the normally sighted
were close together (e.g., red and green, in case b; gold and (depending on the condition of presentation) and, in Mar-
turquoise, or violet and orange, in case d), while some colors, mor's study, 44% for the early blind versus 79% for the
that are very similar for the normally sighted were far apart normally sighted (and 72% for the late blind).
(e.g., purple and violet, in case b; orange and yellow, in case d).
More interpretable are the two solutions, displayed to the far 2. There was almost no overlap, in the two-dimensional
right in Figure 2, for the two of these four cases (viz.; a and c) INDSCAL weight space, between the weights for the early-
that were most correlated with data from subjects possessing blind and the normally sighted subjects. Specifically, the
color vision. The solution for the monochromat in the Names- weights for all of our 6 blind subjects fell outside the smallest
only condition (upper right) recovers the color circle-but with convex region containing all 42 points for our 14 normally
a nearly degenerate grouping of the nine colors into four groups: sighted subjects under the different conditions (see our Fig.
red, green, turquoise-through-purple, and yellow-through- 4). Similarly, the weights for all but I of Marmor's 16 early-
orange (with yellow in a permuted position within its encircled blind subjects fell outside the smallest convex region con-
group). The solution for the three blind subjects who were most taining all 16 points for her normally sighted subjects (see her
correlated with the sighted subjects (lower right) preserves only Fig. 2).
a quite deformed vestige of the color circle in which the
3. The (stress) measures of departure from good fit of the two-
"warm" colors are on the right, the "cool" colors are on the
dimensional configurations obtained by nonmetric multidi-
left, and the red and green sides of the circle are again close
mensional scaling were consistently higher for the early-
together. Possibly, the red-green opposition, which is weak or
blind group than for the normally sighted group-in our
lacking in the most common (deutan and protan) types of color-
study, 6% versus 2%, respectively, and, in Marmor's, 13%
blindness, is also the most likely to be weakly instantiated in the
versus 2%, respectively.
higher, representational levels of the brain. (See White, Lock-
head, & Evans, 1977, for further support for this possibility.) 4. The multidimensional scaling configurations for some of the
Still, the three solutions for the Names-only condition in the blind subjects bore little resemblance to the color circle.
right and top center of Figure 2 all indicate that some individ- Much as in the solution for our three blind subjects who least
uals whose color vision has been either deficient or absent since correlated with the sighted subjects, in the illustrative solu-
birth can acquire some idea of the relations among colors as tion presented by Marmor (Subject 6 in her Fig. 3), quite
experienced by the normally sighted. similar colors (red and orange, or purple and violet) were far
apart, while very dissimilar colors (yellow and purple) were
DISCUSSION AND CONCLUSIONS close together.
We are not saying that color relations could not be learned,
Comparison With the Results of Marmor in the absence of visual experience, by an individual with suf-
Marmor's (1978) follow-up study, although limited to blind ficient motivation and linguistic input about colors. Our mono-
and normally sighted subjects and to the Names-only condition, chromat, having vision though no color vision, clearly took a

102 VOL. 3, NO.2, MARCH 1992

Downloaded from pss.sagepub.com at SETON HALL UNIV on April 1, 2015
 PSYCHOLOGICAL SCIENCE

Roger N. Shepard and Lynn A. Cooper

greater interest in colors than did any of our totally blind sub- virtue of an inborn opponent-process system (and associated
jects. (She even claimed she could discriminate colors, though hue circle) for representing color (in much the way proposed for
any such discriminations were presumably based on lightnesses sighted individuals; see Miller & Johnson-Laird, 1976, pp. 344-
or contextual cues.) Over the years, her interest in vision may 346).
have led her to pick up, through language, the similarity rela- If so, such an understanding need not be purely conceptual,
tions inherent in the color circle. Some of Marmor's 16 early- however. Suggestively, one of our strong protan subjects told
blind subjects may have had a similarly augmented interest in or us that his mental images of red and green were much more
exposure to the semantics of color terms. Nevertheless, our vivid and different from each other than any color sensations he
results, as well as Marmor's, suggest that the relations among experiences when actually looking at red or green objects. Sug-
colors are not easily, completely, or precisely mastered in the gestively, too, the spinning black-and-white pattern on a Ben-
absence of any direct acquaintance with colors. ham disk, which produces illusory sensations of desaturated
red, yellow, green, and blue in normal observers, has led color-
blind observers to say such things as "That line'is definitely
Comparison With the Results of Izmailov and Sokolov colored, but I don't know what to call it" and "Those lines are
greener than grass" (White et aI., 1977, p. 523).
Once an internal representation of colors has been estab-
The possibility remains, however, that the various degrees
lished, however, one need not construct a new mental structure
of understanding of color relations demonstrated by our blind
from scratch in order to make similarity judgments but can
and color-blind subjects arose primarily from their years of
simply reinterpret and use the already existing mental structure.
learning the pairwise relations between colors through lan-
Such a possibility is suggested, for example, by an early exper-
guage. Yet, despite a lifetime of linguistic exposure, individuals
iment by Phillips (1958). English-speaking adults first learned
who have been totally blind since birth seem not to develop,
arbitrary pairings between five Turkish words and five Munsell
grays varying only in lightness. Then, following the pairing of a fully, the remarkably.consistent representation that is charac-
teristic of all sighted individuals. Surely, the neural circuitry for
loud tone with the word associated with the darkest shade of
the normal representation of colors is to some extent innate, but
gray, the strength of the conditioned galvanic skin' response
the development of that circuitry and of its connections with
'(GSR) was found to be greater for those words that had previ-
linguistic input may be reduced in the total absence of any
'ously been associated with more similar (Le., darker) shades of
sensory grounding of that circuitry in the external world.
gray.
Izmailov and Sokolov (this issue) provide a fuller and more
quantitative demonstration, using judged similarity (rather than Acknowledgments-We thank J.P. Cunningham and R.S. Fish for
generalized GSR), a much greater variety of colors, and multi- assistance with many of the data analyses, B. SakiU for providing
access to the rod monochromat, and all 37 subjects for their interest
dimensional scaling. After having their Russian subjects.learn and cooperation. This research was supported by National Science
arbitrary pairings between 20 colors and 20 artificial color Foundation Grants GS-2283 and BNS 90-21684 to the first author
names, they presented each pair of the artificial color names and by a National Science Foundation Predoctoral Fellowship to
only and had the subjects rate the degree of difference between the second author.
the two corresponding colors. Following a small amount of
color-word learning, similarity ratings revealed a crude group-
ing of the words into four clusters-somewhat, perhaps, as the
solution for our monochromat (top right of our Fig. 2) revealed REFERENCES
four clusters. Then, following more extensive color-word learn-
Berlin. B.• & Kay. P. (1969). Basic color terms. Berkeley: University of California
ing, the full color circle emerged. A significant advantage of Press.
using words that had been arbitrarily associated with colors was Carroll;J.D., & Chang, J.-J. (t970a). Analysis of individual differences in multi-
that positions around the color circle could be filled in between dimensional scaling via an N-way generalization of "Eckart-Young" de-
composition. Psychometrika. 35.283-319.
hues with existing names (e.g., between the nine hues corre- Carroll. J.D., & Chang. J.-J. (l970b). Reanalysis of some color data of Helm's by
sponding to our nine familiar English color names). INDSCAL procedure for individual differences multidimensional scaling. In
Proceedings of the 78th Annual Convention of the American Psychological
Association (pp. 137-138). Washington. DC: American Psychological As-
sociation.
Conclusions Chang. J.-J., & Carroll, J.D. (1980). Three are not enough: An INDSCAL analysis
suggesting that color space has seven (± I) dimensions. COLOR Research
The perception and representation of colors does not depend and Application. 5. 193-206.
only on color-sensitive retinal receptors. It also depends on the Cooper. L.A., & Shepard, R.N. (1984). Turning something over in the mind.
Scientific American. 251. 106-114. .
higher level neural networks required to achieve an internal Farnsworth. D. (1947). The Farnsworth dichotomous test for color blindness.
representation of colors that captures what has been most color New York: Psychological Corp.
constant and biologically significant in the light scattered to our Fillenbaum. S., & Rapoport. A. (1971). Structure in the subjecti>'e lexicon. New
York: Academic Press.
eyes from external objects over evolutionary history (Shepard,
Finke, R.A., & Shepard. R.N. (1986). Visual functions of mental imagery. In K.R.
1991). An understanding of the relations among surface colors Boff. L. Kaufman. & J.P. Thomas (Eds.), Handbook of perception and
(as experienced by color-normal observers) tends to develop human performance (pp. 37.1-37-55). New York: Wiley.
even in individuals who have never directly experienced the full Helm. C.D. (1964). A multidimensional ratio scaling analysis of perceived color
relations. Journal of the Optical Society of America. 54.25&-262.
range of colors, as in our protans, deutans, and monochromat. Hering. E. (1964). Outlines of a theory of the light sense. (Trans.). Cambridge.
Such individuals may achieve this understanding, in part, by MA: Harvard University Press. (Original work published 1878)

VOL. 3, NO.2, MARCH 1992 103

Downloaded from pss.sagepub.com at SETON HALL UNIV on April 1, 2015
 PSYCHOLOGICAL SCIENCE

Representation of Colors

Hume, D. (1896). A treatise of human nature. (L.A. Selby-Bigge, Ed.). Oxford: Shepard, R.N. (l962a). The analysis ofproximities: Multidimensional scaling with
Clarendon Press. (Original work published 1739) an unknown distance function, I. Psychometrika, 27, 125-140.
Hurvich, L.M., & Jameson, D. (1957). An opponent-process theory of color Shepard, R.N. (1962b). The analysis ofproximities: Multidimensional scaling with
vision. Psychological Review, 64, 384-404. an unknown distance function, II. Psychometrika, 27, 219-246.
Indow, T. (1988). Multidimensional studies of the Munsell color solid. Psycho- Shepard, R.N. (1974). Representation of structure in similarity data: Problems
logical Review, 95, 456-470. and prospects. Psychometrika, 39, 373-421.
Indow, T., & Kanazawa, K. (1960). Multidimensional mapping of Munsell colors
Shepard, R.N. (1975). Form, formation, and transformation of internal represen-
varying in hue, chroma, and value. Journal ofExperimental Psychology, 59,
tations. In R. Solso (Ed.), Information processing and cognition: The Loy·
330-336.
ola Symposium (pp. 87-122). Hillsdale, NJ: Erlbaum.
Indow, T., & Uchizono, T. (1960). Multidimensional mapping of Munsell colors
varying in hue and chroma. Journal of Experimental Psychology, 59, 321- Shepard, R.N. (l978a). The circumplex and related topological manifolds in the
329. study of perception. In S. Shye (Ed.), Theory construction and data anal·
Ishihara, S. (1977). Tests for colour-blindness (24 plates ed.). Tpkyo: Kanehara ysis in the behavioral sciences (pp. 29-80). San Francisco: Jossey-Bass.
Shuppan. Shepard, R.N. (l978b). The mental image. American Psychologist, 33, 125-137.
Johnson, S.C. (1967). Hierarchical clustering schemes. Psychometrika, 32, 241- Shepard, R.N. (1980). Multidimensional scaling, tree-Iitting, and clustering. Sci·
254. ence, 210, 390-398.
Kruskal, J.B. (l964a). Multidimensional scaling by optimizing goodness of lit to a Shepard, R.N. (1992). The perceptual organization of colors: An adaptation to
nonmetric hypothesis. Psychometrika, 29, 1-27. regularities in the terrestrial world? In J. Barkow, L. Cosmides, & J. Tooby
Kruskal, J.B. (I%4b). Nonmetric multidimensional scaling: A numerical method. (Eds.), The adapted mind: Evolutionary psychology and the generation of
Psychometrika, 29, 28-42. culture. Oxford: Oxford University Press.
Kruskal, J.B., Young, F.W., & Seery, J.B. (1973). How to use KYST: A l'Uy
Shepard, R.N., & Chipman, S. (1970). Second-order isomorphism of internal
flexible program to do multidimensional scaling and unfolding. Unpub- representations: Shapes of states. Cognitive Psychology, 1. 1-17.
lished manuscript. (Available from the AT&T Bell Laboratories, Murray
Hill, NJ) Shepard, R.N., & Cooper, L.A. (1975, September). Representation of colors in
Marmor, G.S. (1978). Age at onset of blindness and the development of the _ normal, blind, and color blind subjects. Paper presented in the symposium
semantics of color names. Journal of Experimental Child Psychology, 25, Applications of Multidimensional Scaling, at the joint meeting of Division 5
267-278. i
of the American Psychological Association and tlie Psychometric Society,
Miller, G.A., & Johnson-Laird, P.N. (1976). Language and perception. Cam- Chicago.
bridge, MA: Harvard University Press. Shepard, R.N., & Cooper, L.A. (1982). Mental images and their transformations.
Munsell Book of Color (Glossy Finish Collection in two binders). (1976). Balti- Cambridge, MA: MIT Press/Brandford Books.
more: Macbeth Division of Kollmorgen Corp. Shepard, R.N., KiJpatHc, D.W., & Cunningham: I.P. (1975). The internal repre·
Ney.10n, I. (1704). Opticks (Book 3). London: S. Smith and B. Walford. sentation of numbers. Cognitive Psychology, 7, 82-138.
Phillips, L.W. (1958). Mediated verbal similarity as a determinant of the gener- White, C.W., Lockhead, G.R., & Evans, N.J. (1977). Multidimensional scaling of
alization of a conditioned GSR. Journal of Experimental Psychology, 55, subjective colors by color-blind observers. Perception & Psychophysics, 21,
56-{j2. 522-526.
Rapoport, A., & Fillenbaum, S. (1972). An experimental study of semantic struc- Wish, M., & Carroll, J.D. (1974). Application of individual differences scaling to
ture. In A.K. Romney, R.N. Shepard, & S.B. Nerlove (Eds.), Multidimen- studies of human perception and judgment. In E.C. Carterette & M.P.
sional scaling: Theory and applications in the behavioral sciences: Vol 2; Friedman (Eds.), Handbook of perception (Vol. 11, pp. 449-491). New
Applications (pp. 93-131). New York: Seminar Press. York: Academic Press.
Schneider, B. (1972). Multidimensional scaling of color difference in the pigeon.
Perception & Psychophysics, 12, 373-378. (RECEIVED 7/29/91; ACCEPTED 9/3/91)

104 VOL. 3, NO.2, MARCH 1992

Downloaded from pss.sagepub.com at SETON HALL UNIV on April 1, 2015