Feature Highlighting Enhances Learning of a ComplexNatural-Science Category

Toshiya Miyatsu and Reshma GouravajhalaWashington University in St. Louis

Robert M. NosofskyIndiana University

Mark A. McDanielWashington University in St. Louis

Learning naturalistic categories, which tend to have fuzzy boundaries and vary on many dimensions, canoften be harder than learning well defined categories. One method for facilitating the category learningof naturalistic stimuli may be to provide explicit feature descriptions that highlight the characteristicfeatures of each category. Although this method is commonly used in textbooks and classrooms,theoretically it remains uncertain whether feature descriptions should advantage learning complexnatural-science categories. In three experiments, participants were trained on 12 categories of rocks,either without or with a brief description highlighting key features of each category. After training, theywere tested on their ability to categorize both old and new rocks from each of the categories. Providingfeature descriptions as a caption under a rock image failed to improve category learning relative toproviding only the rock image with its category label (Experiment 1). However, when these same featuredescriptions were presented such that they were explicitly linked to the relevant parts of the rock image(feature highlighting), participants showed significantly higher performance on both immediate gener-alization to new rocks (Experiment 2) and generalization after a 2-day delay (Experiment 3). Theoreticaland practical implications are discussed.

Keywords: category learning, educational application, geology, multimedia learning

The cognitive psychology literature on how humans acquire andrepresent categories is quite substantial (for reviews, see Ashby &Maddox, 2005; Murphy, 2002; Pothos & Wills, 2011). Yet, nearlyall of the research on category learning in the past 50 years hasbeen conducted with artificial stimuli, such as geometric shapesand simplified drawings. Consequently, the extent to which theprinciples discovered in this research apply to learning of naturalcategories is unclear. One point of departure between the work onartificial stimuli and natural categories that has been extensivelynoted is that artificial categories have often been constructed suchthat they have necessary and sufficient features that preciselydefine the category (Bourne, 1974; Bower & Trabasso, 1963;Shepard, Hovland, & Jenkins, 1961). By contrast, natural catego-

ries are less well structured—they cannot be defined by necessaryand sufficient features, and category boundaries are often fuzzy(Murphy, 2002; Rosch, 1973; Rosch & Mervis, 1975). The pres-ence of defining features in classical artificial categories posi-tioned feature learning as a central focus in this work (e.g., Bower& Trabasso, 1963; Levine, 1975), as well as significantly impactedthe paradigms used to investigate category learning. Specifically,one paradigm involves providing learners with the critical featuresthat defined the category (e.g., “red,” “square”) to facilitate cate-gory learning and to isolate the processes involved in learningbidimensional rules that relate the features to define the category(e.g., conjunctive, disjunctive, conditional, and biconditional;Choi, McDaniel, & Busemeyer, 1993; Salatas & Bourne, 1974).

To our knowledge, unlike research with artificial categories,none of the experimental work with naturalistic categories hasimplemented paradigms in which key features are explicitly iden-tified for the learner. More important, no research has examinedthe consequences of providing explicit feature information forlearning of natural categories—see for instance the paradigmsdeveloped to examine learning of different types of snakes (Noh,Yan, Vendetti, Castel, & Bjork, 2014), categories of paintings(Hartley & Homa, 1981; Kang & Pashler, 2012; Kornell & Bjork,2008), types of birds (Jacoby, Wahlheim, & Coane, 2010; Wahl-heim & DeSoto, 2017; Wahlheim, Finn, & Jacoby, 2012), types ofbutterflies (Birnbaum, Kornell, Bjork, & Bjork, 2013), and typesof rocks (Nosofsky, Sanders, Gerdom, Douglas, & McDaniel,2017; Nosofsky, Sanders, & McDaniel, 2017). In this article, we

Toshiya Miyatsu and Reshma Gouravajhala, Department of Psycholog-ical and Brain Sciences, Washington University in St. Louis; Robert M.Nosofsky, Department of Psychological and Brain Sciences, Indiana Uni-versity; Mark A. McDaniel, Department of Psychological and BrainSciences, Washington University in St. Louis.

We thank Bruce Douglas for providing the feature descriptions for therock categories used in this study. This research was supported by Grant1534014 from the National Science Foundation.

Correspondence concerning this article should be addressed to ToshiyaMiyatsu, Department of Psychological and Brain Sciences, WashingtonUniversity in St. Louis, 1 Brookings Drive, Campus Box 1125, St. Louis,MO 63130. E-mail: tmiyatsu@wustl.edu

Thi

sdo

cum

ent

isco

pyri

ghte

dby

the

Am

eric

anPs

ycho

logi

cal

Ass

ocia

tion

oron

eof

itsal

lied

publ

ishe

rs.

Thi

sar

ticle

isin

tend

edso

lely

for

the

pers

onal

use

ofth

ein

divi

dual

user

and

isno

tto

bedi

ssem

inat

edbr

oadl

y.

Journal of Experimental Psychology:Learning, Memory, and Cognition

© 2018 American Psychological Association 2018, Vol. 0, No. 999, 0000278-7393/18/$12.00 http://dx.doi.org/10.1037/xlm0000538

1

present three novel experiments that provide an initial investiga-tion of the effects of providing explicit feature information tolearners who are attempting to learn to classify rock types. As wedevelop below, this issue is a theoretically interesting one fornaturalistic-category learning, with uncertain outcomes. On theone hand, there are reasons to expect that providing features willbenefit naturalistic category learning; on the other hand, there arealso plausible reasons to expect that providing features will notbenefit, and may even interfere with, generalization.

Beyond the theoretical interest, explicitly identifying key fea-tures of to-be-learned natural categories aligns with educationalpractice. Presentations of examples of to-be-learned categories inclassrooms and other educational situations are almost alwaysaccompanied by verbal descriptions of some key features of thecategory that have been prepared by experts in the field. For theinstruction of rock categories, an instructor may show her studentsan example of the rock obsidian while pointing out that this typeof rock tends to be black, have a glassy texture, and come withshell-like fractures. Textbooks often do the same in that a pictureof a particular type of object is often accompanied by some verbaldescription of the key features of the given category—for anexample involving rock categories, see Figure 1, which is ex-tracted from a popular geology lab manual (Busch & Tasa, 2009).Clearly, the educators’ intuitions are that such feature descriptionswill facilitate students’ learning of these natural categories. Yet,there is no experimental evidence that informs this intuition. Be-low we develop the theoretical reasons for why providing featuredescriptions could facilitate category learning, and other theoreti-cal reasons for why feature descriptions may not facilitate category

learning. Although we draw our reasoning from research using avariety of category structures, we believe that the processes weilluminate in these sections are potentially present in the learningof novel geological rock categories. Then, we present three exper-iments directly contrasting learning of rock categories when fea-ture descriptions are provided relative to when they are not.

The Case for Highlighting Critical Features

There is preliminary evidence from studies involving rule-basedlaboratory categories that feature provision facilitates learning.Salatas and Bourne (1974) had participants study a series ofgeometric shapes differing in four dimensions: color (blue, yellow,or red), size (large, medium, or small), form (hexagon, square, ortriangle), and number of figures (one, two, or three). Although theydid not directly manipulate whether they gave feature descriptions,their participants were told which features were relevant (onefeature from each of two dimensions was always relevant; e.g., thecolor “blue” and “large” size). Compared with other studies thatused similarly structured materials, Salatas and Bourne’s partici-pants showed accelerated learning.

More recently, Eglington and Kang (2017) provided evidencethat highlighting key features facilitates acquisition of a rule-basedscience category. Participants learned various chemical categoriesby studying diagrams of organic compounds. Highlighting diag-nostic features of each chemical category in red (e.g., the presenceof phosphorus or chlorine in the example diagrams belonging toorganophosphate and organochloride categories respectively; seeEglington & Kang, 2017, for examples of highlighted diagrams)

Figure 1. Illustrations of example pictures accompanied by feature descriptions from a popular geologylaboratory manual (Busch & Tasa, 2009, p. 101, 98). See the online article for the color version of this figure.

Thi

sdo

cum

ent

isco

pyri

ghte

dby

the

Am

eric

anPs

ycho

logi

cal

Ass

ocia

tion

oron

eof

itsal

lied

publ

ishe

rs.

Thi

sar

ticle

isin

tend

edso

lely

for

the

pers

onal

use

ofth

ein

divi

dual

user

and

isno

tto

bedi

ssem

inat

edbr

oadl

y.

2 MIYATSU, GOURAVAJHALA, NOSOFSKY, AND MCDANIEL

substantially improved learning (comparing across Experiments 2and 3).

Rock categories (and natural categories in general) bring addi-tional complexities that might lead one to expect a similar positiveeffect of highlighting features. In complex natural category learn-ing—as opposed to rule-based laboratory categories like thosefrom Salatas and Bourne (1974) or rule-based science categorieslike those in Eglington and Kang (2017)—learners have to firstidentify dimensions that characterize the corpus of instances thatare delineated by different categories (and that presumably con-tribute to category differentiation). One challenge for the learner isthat some of these dimensions may not be immediately apparent(Schyns & Rodet, 1997). This challenge contrasts with the vastmajority of artificial categories in which the dimensions are readilyavailable (e.g., common dimensions are size, color, shape, andnumber of stimuli). In the case of rocks, there are many dimen-sions on which rocks vary, such as color, texture, and granularity,to name a few (cf. Nosofsky, Sanders, Meagher, et al., 2017).However, learners may not encode all the dimensions on theirown. For example, when first encountering several rock examples,learners may notice only more salient dimensions and overlookother dimensions. If the learners persist in attending to just salientdimensions, it is possible that they may fail to encode less salientbut relevant dimensions for categorization. It is even theoreticallypossible that some dimensions are ambiguous and must be per-ceptually constructed (not directly extracted), as demonstrated inthe Schyns and Rodet (1997) work with artificial categories. Ac-cordingly, explicitly highlighting the relevant dimensions and spe-cific feature values for particular categories would support learn-ing by ensuring that the relevant dimensions are attendedregardless of its salience.

Moreover, after identifying the dimensions, learners have toobserve the variation in each dimension and identify the appropri-ate range of values that characterize the examples from a givencategory. In the case of the rock categories, after identifying color,texture, and granularity as diagnostic dimensions (for instance),learners have to identify the values of each dimension that char-acterize particular categories. For example, obsidian tends to beglassy black whereas marble is light-colored. Highlighting criticalfeatures could give learners an advantage in identifying the char-acteristic features of each category. As an example, consider theprocesses assumed in one particular model of category learning,the RULEX model (Nosofsky, Palmeri, & McKinley, 1994). Ac-cording to RULEX, category learners generate simple rules (i.e.,hypotheses about which features are critical for determining thecategories) that can encompass as many instances as possible andthen memorize the exceptions to these rules for each category.Highlighting feature descriptions would presumably speed theadoption of reasonably accurate simple rules, thereby allowinglearners to more quickly focus on learning the exceptions. Indeed,in one way or another, a process of identifying critical dimensionsfor categorization is assumed in almost all of prominent models ofcategory learning, such as prototype models (e.g., Reed, 1972),rule-learning models (e.g., Bower & Trabasso, 1963; Levine,1975; Nosofsky et al., 1994), exemplar models (e.g., Kruschke,1992; Nosofsky, 1984), and clustering models (e.g., Love, Medin,& Gureckis, 2004) and, thus, it might be broadly expected thathighlighting features for learners would facilitate category learn-ing.

Finally, unlike the vast majority of artificial category structuresin which variation in the same dimension(s) forms the basis of allcategories to be learned, characteristic features of rock categories,at least, may be associated with different dimensions across cate-gories. For instance, whereas the characteristic features of graniteinclude color (light color or reddish) and granularity (interlockingcoarse-grained crystals), the primary characteristic feature of pum-ice involves texture (rough textured volcanic glass with smallholes). This form of variation may create additional complexity forlearners. Consequently, highlighting critical features for each cat-egory could enhance learning by focusing learners’ attention to therelevant dimensions that can be especially helpful in this type ofcomplex category structure.

The Case Against Highlighting Critical Features

The literature on category learning and perceptual expertise alsoraises the possibility that highlighting features may not enhanceand may even interfere with natural category learning. First, ex-perts and novices differ in the dimensions to which they attend, aswell as the precision with which they can detect the differenceswithin a given dimension (Goldstone, 2003; Sowden, Davies, &Roling, 2000; see Palmeri, Wong, & Gauthier, 2004, for review).For example, when medical X-ray images were shown to expertsand novices, experts had higher sensitivity to low-contrast dots, adimension that is critical in detecting abnormalities (Sowden et al.,2000). Thus, feature characteristics that experts use to categorizeobjects may not be immediately useful for novices, because nov-ices cannot identify these dimensions or lack the acuity necessaryto detect the changes in the given dimension associated with eachcategory.

Second, highlighting features may circumvent processes neces-sary for natural category learning, given that natural categories canhave many dimensions. Pertinent to the present focus on rockcategory learning, Nosofsky, Sanders, Meagher, et al. (2017) con-ducted multidimensional scaling on similarity rating data of 12different examples from each of 30 common rock categories. Theresulting psychological feature space for rocks’ visual featuresinvolved at least eight complex dimensions, more than is typicallyfound in artificial categories where it is not uncommon to generatestimuli that vary on only two dimensions (e.g., Ashby & Maddox,1992; Nosofsky, 1986).

More important, allocating significant attention to all of thesedimensions may be critical in successful learning of the rockcategories. Consistent with this possibility, Nosofsky, Sanders, andMcDaniel, (2017) had participants learn some of these rock cate-gories (from which the multidimensional scaling study was based)and fitted several models (e.g., prototype model, exemplar model)to the test classification data. The exemplar model was the mostsuccessful in accounting for the observed data (the generalizedcontext model; Nosofsky, 1986, 2011). Importantly for presentpurposes, the successful instantiation of the exemplar model as-sumed that learners gave significant attention to all dimensions(rather than weighing some dimensions dramatically more thanothers). Providing selected feature descriptions would clearly beexpected to bias learners’ attention toward these selected dimen-sions at the outset of learning. Doing so, however, could conceiv-ably distract learners from adequately encoding the training in-stances along the entire feature space. Thus, if the encoding of the

Thi

sdo

cum

ent

isco

pyri

ghte

dby

the

Am

eric

anPs

ycho

logi

cal

Ass

ocia

tion

oron

eof

itsal

lied

publ

ishe

rs.

Thi

sar

ticle

isin

tend

edso

lely

for

the

pers

onal

use

ofth

ein

divi

dual

user

and

isno

tto

bedi

ssem

inat

edbr

oadl

y.

3FEATURE PROVISION IN NATURAL CATEGORY LEARNING

entire feature space is necessary for successful learning of rockcategories, it is possible that provision of selective features mayhinder learning.

Lastly, the findings from studies using artificial categories thatfavor providing defining features to learners (Salatas & Bourne,1974) may not be directly applicable to more complex categorieslike rocks because of the numerous differences between artificialand natural categories. Unlike many artificial laboratory categoriesthat vary in a few dimensions (e.g., color, size, and form) with onlya few discrete features (e.g., blue, yellow, or red in the colordimension), as noted above rock categories have a greater numberof dimensions (Nosofsky, Sanders, Meagher, & Douglas, 2017)and the feature variation within a dimension is continuous. More-over, natural categories tend to have high variability within dimen-sions (Murphy, 2002), and the rock categories in particular appearto have rather complex structures (Nosofsky, Sanders, Gerdom, etal., 2017). For example, in some cases, an instance from onecategory looks more similar to instances from other categories thanto the instances from the category to which it belongs. Thus, thefacilitative effect of providing features for clearly defined rule-based categories (Salatas & Bourne, 1974) may not extend tolearning rock categories.

Experiment 1

One straightforward way to provide feature descriptions is toplace them right below the rock images used for category training,as is often done in textbooks (see Figure 1). Thus, in this firstexperiment, during training, some participants were presented withrock images and an associated feature description below eachimage (along with the category name), whereas other participantsstudied the same rock images with only the category name shown.

We also manipulated training-exemplar variability. Increasingthe number of unique training exemplars as opposed to repeatingthe same training exemplars has been shown to enhance classifi-cation performance of new examples (Wahlheim & DeSoto, 2017;Wahlheim et al., 2012). However, it is possible that participantsmay not need as many unique training exemplars if feature de-scriptions are provided. One potential mechanism through whichtraining-exemplar variability helps learning is that it allows learn-ers to observe the variation associated with each dimension moreprecisely. This enhancement in precision can lead to a betteridentification of unique features of each category in a givendimension. To explore this possibility in Experiment 1, half theparticipants studied the categories of rocks in a condition with sixunique training exemplars per category, with each individual ex-emplar repeated twice; whereas the other half studied with twounique training exemplars per category, with each individual ex-emplar repeated six times.

Following training, we evaluated participants’ ability to classifythe old training items and new transfer items in a test phase. Inaddition, we assessed participants’ learning strategy with a post-experimental questionnaire that asked participants to rate whetherthey took a more abstraction-based or memorization-based ap-proach. Much debate in the category learning literature focused onhow learned categories are represented (Murphy, 2002). Accord-ing to an abstraction-based approach, learners develop some sort ofa summary representation of the learned categories. Specifically,learners may develop a set of rules that determines categorization

(rule-based) or prototypes of the learned categories (prototype-based), and the rules or the prototypes are applied to classifyexemplars that are subsequently encountered (e.g., Allen &Brooks, 1991; Erickson & Kruschke, 1998; Posner & Keele, 1968,1970; Regehr & Brooks, 1993; Trabasso & Bower, 1968). On theother hand, a memorization-based (or exemplar-based) approachassumes that learners store exemplars from each category and thesubsequent categorization judgment is made according to the sim-ilarity of the newly encountered exemplars to the stored exemplars(e.g., Brooks, 1978; Kruschke, 1992; Medin & Schaffer, 1978;Nosofsky, 1986; Nosofsky & Kruschke, 1992). Based on theartificial category learning literature, we thought it possible thatlearners would vary in terms of adopting a more abstraction-basedor memorization-based approach within a single categorizationtask (Bourne, Healy, Kole, & Graham, 2006; Little & McDaniel,2015a; Wahlheim, McDaniel, & Little, 2016; see also Little &McDaniel, 2015b).

Method

Participants and design. There were 204 Amazon Mechan-ical Turk workers who participated in exchange for monetarycompensation. To qualify for participation in the experiment,Mechanical Turk workers had to have met the following threequalifications: A HIT (Human Intelligence Task) approval rate forall requesters’ HITs greater than or equal to 95%; the number ofHITs approved was greater than or equal to 500; and must residein the United States. The mean age of the sample was 36.5 yearsof age, and 62% had obtained an Associate’s degree or higher.Participants who failed to complete the experiment or who re-ported a high prior knowledge of the material on a postexperimentquestionnaire (n � 28) were excluded from the analyses (final N �176). A 2 � 2 between-subjects factorial design was used, withfeature highlighting (feature-description present, absent) and num-ber of unique training exemplars (2, 6) as independent variables.

Materials. Pictures of rocks from 12 categories (four catego-ries from each of the three higher-order categories of igneous,metamorphic, and sedimentary) were assembled. The images werescaled to be similar sizes and presented against a white back-ground. The materials were taken from a larger set created byNosofsky, Sanders, and Meagher, et al. (2017). The present ma-terials consisted of 12 rocks from each of the following 12 cate-gories: granite, obsidian, pegmatite, pumice, amphibolite, gneiss,marble, slate, breccia, conglomerate, rock gypsum, and sandstone.For each of these categories, a geological science faculty memberat Indiana University provided short feature descriptions that werepresented in conjunction with the rocks in the highlighted featureconditions. Note that because these are characteristic features (i.e.,features that are typically but not always present in categorymembers), some of the rock tokens were missing one or more ofthe described category-relevant features. The pictures of rocksused in all the experiments reported in the current article can befound at Open Science Framework (https://osf.io/9vg8m/#); Ap-pendix A provides the feature descriptions for each category. Allexperiments reported in the current article were programmed inCollector (http://github.com/gikeymarcia/Collector), a PHP-basedopen-source experiment program designed to conduct psycholog-ical experiments through Web-browsers.

Thi

sdo

cum

ent

isco

pyri

ghte

dby

the

Am

eric

anPs

ycho

logi

cal

Ass

ocia

tion

oron

eof

itsal

lied

publ

ishe

rs.

Thi

sar

ticle

isin

tend

edso

lely

for

the

pers

onal

use

ofth

ein

divi

dual

user

and

isno

tto

bedi

ssem

inat

edbr

oadl

y.

4 MIYATSU, GOURAVAJHALA, NOSOFSKY, AND MCDANIEL

Procedure. Participants first completed an observation phase,where they were presented with a randomly chosen rock from eachof the 12 categories and its corresponding category label for 6 s(feature descriptions were not provided). No feedback was pro-vided in this phase.

Then, participants completed 12 blocks of feedback learning.Each feedback learning block contained 12 trials. For each block,across trials a training exemplar from each of the 12 categories waspresented. For a trial, the training exemplar was presented for 6 s,during which time participants clicked one of the 12 categorynames (using the mouse) to indicate a response. Feedback was thenpresented for 4 s. If participants failed to make a response withinthe 6 s, the program nevertheless progressed to the feedbackscreen. For the feature description-absent conditions (FD-absent),the feedback was the picture of the just-tested rock with the correctcategory label. For the feature description-present (FD-present)conditions, the feedback in addition presented a feature descriptionbelow the picture.

For each participant, the appropriate number (2 or 6) oftraining exemplars were randomly selected from the 12 exem-plars in each category, and the presentation sequence of thesetraining exemplars was randomized. In the 2-unique-training-exemplar conditions, the feedback learning blocks contained sixrepetitions of the same two randomly chosen training exemplarsfrom each of the 12 categories. In the 6-unique-training-exemplar conditions, the feedback learning blocks containedtwo repetitions of these six training exemplars. In all condi-tions, the first feedback learning block contained the same rocksas those presented in the observation phase. The repetition oftraining exemplars was spaced by block; in other words, in the6-unique-training-exemplar conditions, for example, an itemwas presented once in the first six blocks and then repeatedonce in the last six blocks.

After participants completed the 12 feedback learning blocks,they completed a distractor task (Tetris) for three minutes.Following this distractor task, participants then completed thefinal test phase. In the final test phase, participants were pre-sented with 72 test items, containing two rocks from each of the12 categories they had been trained on (old training items) andfour randomly chosen rocks that they had not been trained on(new transfer items) from each of the 12 categories. The orderof presentation was randomized for each participant. The par-ticipants were instructed to select, at their own pace, whichcategory each presented item belonged to by clicking one of the12 category names using the mouse. No feedback was providedduring the final test phase.

Following the final test phase, participants completed a strategyquestionnaire (adopted from Wahlheim et al., 2016) where theyreported the strategy used during the training phase. Participantswere instructed to think back to the training phase and rememberwhether they were more focused on developing a rule for eachcategory, or on memorizing rocks and their corresponding cate-gory labels. They made their ratings on a Likert scale ranging from1 (memorization strategy) to 7 (rule-establishing strategy). Partic-ipants were instructed to give an extreme rating only if they usedthat strategy exclusively throughout training, but to otherwise usethe full range of the scale to indicate the degree to which theypreferred one strategy over the other. Participants were also told to

give a rating of 4 if they used both strategies equally or if theywere unsure about their strategy preferences.1

Results

Training phase performance. Figure 2 shows the learningcurves for each condition across the 12 feedback learning blocks.A 2 (FD present or FD absent) � 2 (number of unique trainingexemplars: 2 or 6) � 12 (feedback learning blocks) mixed analysisof variance (ANOVA), with the presence or absence of featuredescriptions and the number of unique training exemplars as thebetween-subjects variables and the feedback learning blocks as thewithin-subjects variable, was conducted on these data.

There was no main effect of the presence or absence of featuredescriptions, F(1, 172) � 1. There was a significant main effect ofthe number of unique training exemplars, F(1, 172) � 32.11, p �.001, MSE � 11.70, �p

2 � 0.16, indicating that the performanceduring the training phase was higher when participants were pro-vided 2 unique training exemplars (M � .69, SD � .18) than whenthey were provided 6 unique training exemplars (M � .54, SD �.18). There was a significant main effect of learning block, F(11,172) � 121.44, p � .001, MSE � 1.72, �p

2 � 0.41, such thataccuracy increased over the course of the 12 feedback learningblocks.

The interaction between number of unique training exemplarsand the presence or absence of feature descriptions was not sig-nificant, F(1, 172) � 1. There was a significant interaction be-tween feedback learning block and the number of unique trainingexemplars, F(11, 172) � 19.39, p � .001, MSE � 0.28, �p

2 � 0.10.Inspection of Figure 2 suggests that accuracy for those in the2-unique-training-exemplar condition increased more from thefirst half (M � 0.59, SD � .18) of training to the second half (M �0.78, SD � 0.19) than it did for those in the 6-unique-training-exemplar condition (first half: M � 0.48, SD � 0.17; second half:M � 0.60, SD � 0.19).

There was also a significant interaction between feedback learn-ing block and the presence or absence of feature descriptions,F(11, 172) � 3.38, p � .001, MSE � 0.05, �p

2 � 0.02. As seen inthe lower panel of Figure 2, the pattern of training performanceacross the 12 learning blocks does not seem to reveal any system-atic differences between the FD-present and FD-absent conditions.We will further examine this interaction if it turns out to be reliable(i.e., also emerges in Experiments 2 and 3). The three-way inter-action between feedback learning block, the number of uniquetraining exemplars, and the presence or absence of feature descrip-tions was not significant, F(11, 172) � 1.04, p � .05, MSE � 0.02,�p

2 � 0.01.Final test performance. Figure 3 shows participants’ final

test performance as a function of conditions and test type. A 2 (testtype: old training or new transfer) � 2 (number of unique trainingexemplars: 2 or 6) � 2 (FD present or FD absent) mixed ANOVA,with the test type as the within-subjects variable and the number ofunique training exemplars and the presence or absence of featuredescription as the between-subjects variables, was conducted onthese data.

1 The strategy questionnaire results did not reveal anything directlypertinent to the main questions of the current study; however, for com-pleteness these results are reported in Appendix B.

Thi

sdo

cum

ent

isco

pyri

ghte

dby

the

Am

eric

anPs

ycho

logi

cal

Ass

ocia

tion

oron

eof

itsal

lied

publ

ishe

rs.

Thi

sar

ticle

isin

tend

edso

lely

for

the

pers

onal

use

ofth

ein

divi

dual

user

and

isno

tto

bedi

ssem

inat

edbr

oadl

y.

5FEATURE PROVISION IN NATURAL CATEGORY LEARNING

Most importantly, there was no hint of an effect of the presenceor absence of feature descriptions on test performance, F(1,172) � 1. Performance on the old training items was significantlyhigher (M � .70, SD � .20) than on the new transfer items (M �.59, SD � .16), F(1, 172) � 117.49, p � .001, MSE � 1.20, �p

2 �0.41. There was also a significant main effect of the number ofunique training exemplars, F(1, 172) � 10.37, p � .01, MSE �0.62, �p

2 � 0.06, revealing better overall test performance whenparticipants were provided 2 unique training exemplars (M � .69,SD � .18) than when they were provided 6 unique trainingexemplars (M � .60, SD � .18). This effect was qualified by asignificant interaction between test type and the number of uniquetraining exemplars, F(1, 172) � 60.84, p � .001, MSE � 0.62,�p

2 � 0.26, such that the advantage of the 2-unique-exemplarcondition was limited to the old training items. Post hocindependent-sample t tests confirmed that performance on the oldtraining items was significantly better in the 2-unique-exemplarcondition (M � .79, SD � .22) than in the 6-unique-exemplarcondition (M � .62, SD � .21), t(174) � 5.24, p � .05, d � 0.79.Performance on the new transfer items was identical across the twoconditions (2-unique exemplar conditions: M � .59, SD � .15;6-unique exemplar conditions: M � .59, SD � .16).

The two-way interactions between the presence or absence offeature descriptions and number of unique training exemplars andthe presence or absence of feature descriptions and test type

were not significant (largest F(1, 172) � 1.60, MSE � 0.02,�p

2 � 0.01, for the latter interaction), nor was the three-wayinteraction (F � 1).

Discussion

For present purposes the most important finding was that pro-viding feature descriptions did not enhance learning. One interpre-tation is that this pattern aligns with the theoretical perspectivesoutlined in the introduction that anticipated no benefits of high-lighting features for learning these natural rock categories. Analternative, however, is that the presentation of the feature descrip-tions did not function in its intended manner for several potentialreasons. First, the expert-like terminology that was included in thefeature descriptions (e.g., extrusive, vesicles) may not be familiarto laypeople (MTurk workers) and, thus, what the descriptionswere referring to may have been unclear. Moreover, even if theparticipants could understand what the terms meant, it is possiblethat they could not tie the terms to the corresponding perceptualdimensions in the rock pictures. For example, when a participantsaw a picture of amphibolite with a description that read “darkcolored and weakly grained structure,” she might not understandwhat perceptual features reflected “grain.” Lastly, even if theparticipants could understand the meaning of the terms and iden-tify the described dimension, they might not be able to tie thedescription to specific perceptual features that characterized thevalues within the dimension (e.g., what perceptual featurescounted toward characterizing the grains as “weak”).

Closely related to the above point, it appears in retrospect thatthe features were not presented in a way that supported optimalprocessing of the information. Multimedia learning principles(Mayer, 2002, 2005) assume that to maximize the benefits ofpresenting words and pictures together, extraneous processing(i.e., any processing that does not contribute to learning and drainslimited cognitive capacity) needs to be minimized, so that learnerscan allocate more cognitive resources to essential processing (i.e.,selecting relevant information and organizing them in workingmemory). In Experiment 1, gaze shifts from the caption below thepicture to corresponding parts of the picture, as well as the act oflooking for the parts of the picture described in the feature de-scriptions, could be considered extraneous processing. Reducingthese aspects of extraneous processing may be crucial in drawing

Figure 3. Participants’ final test performance as a function of conditionsand test type in Experiment 1.

Figure 2. Participants’ performance across the training blocks as a func-tion of feature description and number of unique training exemplars train-ing in Experiment 1.

Thi

sdo

cum

ent

isco

pyri

ghte

dby

the

Am

eric

anPs

ycho

logi

cal

Ass

ocia

tion

oron

eof

itsal

lied

publ

ishe

rs.

Thi

sar

ticle

isin

tend

edso

lely

for

the

pers

onal

use

ofth

ein

divi

dual

user

and

isno

tto

bedi

ssem

inat

edbr

oadl

y.

6 MIYATSU, GOURAVAJHALA, NOSOFSKY, AND MCDANIEL

out the benefits of feature highlighting, especially given the limitedtime participants had to process the feature feedback (4 s).

One surprising finding was that training with a larger sample ofdifferent exemplars did not enhance generalization to new transferitems. From the perspective of prototype theory, more completecategory learning requires appreciating the allowable variationfrom a prototype that category members can exhibit (e.g., Posner& Keele, 1968). In line with this perspective, training with a largerset of different exemplars has been shown to enhance learning ofother natural categories, that of birds (Wahlheim et al., 2012). It isbeyond the scope of the present research, however, to pinpointwhy these results with rock categories versus bird categories arediscrepant.

Experiment 2

In Experiment 2 the same feature descriptions used in Experi-ment 1 were embedded within the rock pictures, and the corre-sponding parts of the image that were reflected in the feature

description were circled (see Figure 4 for examples; for conve-nience we refer to this particular feature-description implementa-tion as feature highlighting). We reasoned that these modificationsshould clarify the perceptual aspects of the rock to which thefeatures referred (i.e., signaling; Meyer, 1975) as well as facilitatethe scanning needed to integrate reading the described features andprocessing the images. Specifically, these modifications follow thespatial contiguity principle of the multimedia principle (Mayer,2002, 2005) that allows learners to minimize extraneous process-ing and maximize essential and generative processing. Supportingthis reasoning, Moreno and Mayer (1999) had participants studyhow lightning forms with an animation and words. Participants’performance on a transfer test was higher when the words wereembedded in the animation next to the specific parts being de-scribed compared with when the words were placed below theanimation as a caption. Similarly, Eglington and Kang’s (2017)successful implementation of feature highlighting was embeddedwithin the training pictures of organic-compound diagrams.

Figure 4. Examples of rock pictures with embedded feature descriptions used in Experiment 2 and 3. See theonline article for the color version of this figure.

Thi

sdo

cum

ent

isco

pyri

ghte

dby

the

Am

eric

anPs

ycho

logi

cal

Ass

ocia

tion

oron

eof

itsal

lied

publ

ishe

rs.

Thi

sar

ticle

isin

tend

edso

lely

for

the

pers

onal

use

ofth

ein

divi

dual

user

and

isno

tto

bedi

ssem

inat

edbr

oadl

y.

7FEATURE PROVISION IN NATURAL CATEGORY LEARNING

If the participants in Experiment 1 failed to benefit from thefeature provision because of their inability to understand or locatethe perceptual aspects described by the feature descriptions, theirinability to overcome the burden posed by extraneous processing,or both, then the participants in the present feature highlightingcondition should outperform those in the control condition. Alter-natively, if the participants in Experiment 1 failed to benefit fromfeature provision because of other reasons (as outlined in theintroduction), Experiment 2 should not show the advantage offeature highlighting.

Method

Participants and design. There were 80 Washington Univer-sity in St. Louis undergraduate students who participated in thisexperiment in exchange for course credit or monetary compensa-tion. There were two between-subjects conditions: FD present orabsent. All participants were trained on 6 unique training exem-plars, each repeated twice across the 12 feedback learning blocks.

The materials and procedure were identical to Experiment 1,except the feature descriptions in the FD-present condition wereembedded in the rock pictures and the parts being described werecircled as shown in Figure 4. Specifically, each feature descriptionwas split into shorter fragments that referred solely to one partic-ular dimension (e.g., color, pattern, texture, and grain size). Fea-tures were highlighted using a colored circle and the correspondingfeature description fragment was placed directly adjacent to thecircle along with the category name.

Results and Discussion

Training phase performance. Figure 5 shows the learningcurves for each condition across the 12 feedback learning blocks.A 2 (FD present or FD absent) � 12 (feedback learning blocks)mixed ANOVA, with feature-description condition as thebetween-subjects variable and the feedback learning blocks as thewithin-subjects variable, was conducted on these data. There wasno significant main effect of feature description, F(1, 78) � 1.There was a significant main effect of learning block, F(11, 78) �52.05, p � .001, MSE � 0.80, �p

2 � 0.40, such that performanceincreased as training progressed.

There was a significant interaction between feature-descriptioncondition and learning block, F(11, 78) � 6.79, p � .001, MSE �0.10, �p

2 � 0.08. To try to simplify the source of this interactionbased on our observations from the Experiment 1 training resultswe collapsed performance across the first and second halves of thefeedback learning blocks. We found no significant interactionbetween the first and second halves of feedback learning andcondition, however, F(1, 78) � 2.14, p � .05, MSE � 0.01, �p

2 �0.03. Therefore, we conducted simple comparisons, and found thatthe interaction appears to have been mostly driven by a significantadvantage in learning in the FD-absent condition (M � 0.81, SD �0.18) relative to the FD-present condition (M � 0.63, SD � 0.20)in Block 7, t(78) � 4.23, p � .001, d � 0.95. Therefore, althoughthere was a significant interaction in both Experiments 1 and 2,there appears to be no systematicity in the pattern. In summary, aswas the case in Experiment 1, providing feature descriptions didnot generally affect categorization performance during the trainingphase.

Final test performance. Figure 6 shows participants’ finaltest performance as a function of conditions and test type. A 2 (FDpresent or FD absent) � 2 (test type: old training or new transfer)mixed ANOVA, with the feature description condition as thebetween-subjects variable and the test type as the within-subjectsvariable, was conducted on these data. In general, performance onthe old training items was significantly better (M � .85, SD � .10)than on the new transfer items (M � .80, SD � .08), F(1, 78) �46.39, p � .001, MSE � 0.13, �p

2 � 0.37.More important, there was a significant main effect of feature

description, F(1, 78) � 25.49, p � .001, MSE � 0.74, �p2 � 0.25,

such that those in the FD-present condition performed significantlybetter on the final test (M � .89, SD � .12) than did those in theFD-absent condition (M � .76, SD � .12). This effect was qual-ified by a significant interaction with test type, F(1, 78), � 132.39,p � .001, MSE � 0.37, �p

2 � 0.63. Post hoc independent-samplet tests showed that new transfer performance was substantiallyadvantaged when features were provided during training relative towhen they were not provided (FD present M � .91, SD � .10; FDabsent M � .68, SD � .12), t(78) � 9.38, p � .001, d � 2.08. Bycontrast, performance on old training items did not significantlydiffer between the groups (FD present: M � .87, SD � .13 vs. FDabsent: M � .83, SD � .15), t(78) � 1.28, p � .05, d � 0.28,consistent with the patterns during the training phase.

In summary, the results supported the possibility that the Ex-periment 1 participants failed to benefit from feature provisionbecause the feature description did not function in its intendedmanner. The present results suggest that learners can benefit fromfeature descriptions if those descriptions are presented such thatlearners can locate and identify the described perceptual aspects.Also, the placement of the feature descriptions may have supportedoptimal processing of the verbal descriptions in conjunction withthe perceptual input (according to the spatial contiguity principle).An interesting find was that feature highlighting selectively en-hanced performance on the new transfer items; it did not enhanceperformance on the old training items or initial training accuracy.This pattern indicates that providing features did not facilitatelearning particular training instances, but it did facilitate learningthe categories—that is, aspects that supported generalization. Itshould be noted, however, that the interaction may have emergedbecause of a ceiling effect given the high performance on both old

Figure 5. Participants’ performance across the training blocks as a func-tion of conditions in Experiment 2.

Thi

sdo

cum

ent

isco

pyri

ghte

dby

the

Am

eric

anPs

ycho

logi

cal

Ass

ocia

tion

oron

eof

itsal

lied

publ

ishe

rs.

Thi

sar

ticle

isin

tend

edso

lely

for

the

pers

onal

use

ofth

ein

divi

dual

user

and

isno

tto

bedi

ssem

inat

edbr

oadl

y.

8 MIYATSU, GOURAVAJHALA, NOSOFSKY, AND MCDANIEL

training (M � .87) and new transfer items (M � .91) in theFD-present condition. Before discussing the theoretical implica-tions of this result, we first report another experiment to establishthe reliability of the benefit of highlighting features for learningrock categories.

Experiment 3

Experiment 3 was conducted to extend the experiment 2 find-ings to a delayed final test (a 2-day delay).

Method

Participants, design, and procedure. There were 59 under-graduate students at Washington University in St. Louis whoparticipated in this experiment in exchange for course credit ormonetary compensation. There were 2 between-subjects condi-tions: FD present (n � 29) or FD absent (n � 30). The procedurewas identical to Experiment 2, except for the introduction of a48-hr delay between the training and the final test phases.

Results

Training phase performance. Figure 7 shows the learningcurves for each condition across the 12 feedback learning blocks.The performance increase observed in the feedback learning Block

7 was because of the fact that that was the first time the partici-pants saw the same training exemplars for the second time (sixtraining exemplars per category were repeated twice). A 2 (FDpresent or FD absent) � 12 (feedback learning blocks) mixedANOVA, with the feature description condition as the between-subjects variable and the feedback-learning blocks as the within-subjects variable, was conducted on these data. Converging withthe previous experiments, there was no significant main effect offeature description, F(1, 57) � 3.07, p � .09, MSE � 0.36, �p

2 �0.05. There was a significant main effect of feedback learningblock, F(11, 57) � 42.59, p � .001, MSE � 0.66, �p

2 � 0.43, suchthat performance increased over the course of the 12 feedbacklearning blocks. The interaction between feature description andfeedback learning block was not significant, F(11, 57) � 1.16, p �.05, MSE � 0.02, �p

2 � 0.02.Final test performance. Figure 8 shows participants’ final

test performance as a function of conditions and test type. A 2 (FDpresent or FD absent) � 2 (test type: old training or new transfer)mixed ANOVA, with the feature-description condition as thebetween-subjects variable and test type as the within-subjectsvariable, was conducted on these data. There was a significantmain effect of test type, F(1, 57) � 104.13, p � .001, MSE � 0.46,�p

2 � 0.65, such that performance on the old training items (M �.84, SD � .12) was significantly better than on the new transferitems (M � .72, SD � .11).

More important, as in Experiment 2, providing highlightedfeatures during training produced more accurate test performance(M � .81, SD � .10) compared with not providing features duringtraining (M � .75, SD � .10), F(1, 57) � 5.58, p � .05, MSE �0.11, �p

2 � 0.09. A marginally significant interaction with test typesuggested that the advantage of providing critical features waslimited to new transfer items, F(1, 57) � 3.60, p � .06, MSE �0.02, �p

2 � 0.06. Post hoc independent-sample t tests confirmedthis suggestion. On new transfer items, the participants providedwith feature descriptions during training (M � .76, SD � .10)significantly outperformed those who were not provided withfeature descriptions (M � .68, SD � .11), t(57) � 3.12, p � .05,d � 0.76. On old training items, the groups did not statisticallydiffer (FD present: M � .86, SD � .10 vs. FD absent: M � .83,SD � .13), t(57) � 1.27, p � .05, d � 0.26.

In short, Experiment 3 extended the main finding in Experiment2 to a final test that was administered 2 days after training. Feature

Figure 6. Participants’ final test performance as a function of conditionsand test type in Experiment 2.

Figure 7. Participants’ performance across the training blocks as a func-tion of conditions in Experiment 3.

Figure 8. Participants’ final test performance as a function of conditionsand test type in Experiment 3.

Thi

sdo

cum

ent

isco

pyri

ghte

dby

the

Am

eric

anPs

ycho

logi

cal

Ass

ocia

tion

oron

eof

itsal

lied

publ

ishe

rs.

Thi

sar

ticle

isin

tend

edso

lely

for

the

pers

onal

use

ofth

ein

divi

dual

user

and

isno

tto

bedi

ssem

inat

edbr

oadl

y.

9FEATURE PROVISION IN NATURAL CATEGORY LEARNING

provision during training selectively enhanced the performance onnew transfer items but not on old training items.

General Discussion

Providing feature descriptions of rock categories as a captionbelow images of training instances failed to improve learning ofthe rock categories relative to training without the feature descrip-tions (Experiment 1). However, when these verbal feature descrip-tions were displayed so that the described feature was explicitlylinked to particular perceptual aspects of the rock image (that wetermed feature highlighting), learning of rock categories was sig-nificantly enhanced as evidenced on an immediate generalizationtest (Experiment 2) and a 2-day-delayed generalization test (Ex-periment 3). As reviewed in our introduction, classic previouswork has shown that providing explicit information concerningdefining features facilitates learning of artificial rule-based cate-gories (Salatas & Bourne, 1974). The present work extends suchfindings in a significant manner by showing that parallel resultshold for the learning of natural categories. Given that naturalcategories have more complex category structures than laboratoryrule-based categories (e.g., unlikely to have defining features), aricher array of perceptual dimensions, continuous dimensionalvalues (unlike many laboratory categories), and high variationwithin dimensions, it was uncertain a priori that feature provisionwould enhance natural-category learning. We first discuss theseresults in terms of their theoretical implications and then highlightthe important practical implications.

One theoretical idea suggested at the outset was that providingnovice learners with feature descriptions generated by an expertmight not be helpful for novices because novices may not be ableto interpret the verbal descriptions or may not have the ability toperceive and extract the visual aspects to which the descriptionsrefer (unlike for artificial categories in which feature descriptionscan be readily interpretable, such as “blue” and “square”). Con-sider that a feature provided for pumice was “rough texturedvolcanic glass.” A novice may have a poor idea of what “volcanic”glass is. As another example, for conglomerate a provided featurewas “rounded fragments cemented together.” Though a novicemay be able to understand the terms, it is not certain that the novicewould be able to identify the perceptual qualities of a rock image thatcounts as “fragments” and “cemented together.” The results from thestudy as a whole support a more refined specification of this initialtheoretical view, at least for rock category learning. With a standardpresentation of features often found in educational settings, inwhich the features were presented separately but concurrently witha training instance (e.g., the feature description was placed belowan image), feature descriptions failed to improve learning (Exper-iment 1). That is, an expert’s summary of characteristic features ofto-be-learned categories does not necessarily support novices’learning of those categories. Because at least some of the featuresare themselves likely constructed as part and parcel of expertise(Schyns & Rodet, 1997), the features must be highlighted so thatnovices can understand the feature descriptions and their percep-tual referents. When these conditions are met, then providinglearners with those characteristic features improves rock categorylearning (Experiments 2 and 3). We assume that the same principlewould hold for learning of many natural science categories basedon visual features. We hasten to acknowledge, however, that the

different populations from which participants were sampled inExperiment 1 versus Experiments 2 and 3 could possibly qualifythe above interpretation. For instance, perhaps with the highlyselective sample of college students from Experiments 2 and 3,separate presentation of features (rather than highlighted presen-tation) could still enhance rock category learning.

A second theoretical idea was that providing features might notbenefit rock category learning because encouraging learners tofocus on one or two dimensions might restrict the encoding of arich array of perceptual information from rocks, which in turncould limit category learning. Previous work has shown that whennovice learners are given no feature information and asked to ratesimilarity of rock images, their ratings reflect a complex featurespace of at least eight perceptual dimensions (as reported inNosofsky, Sanders, Meagher, & Douglas’, 2017 multidimensionalscaling study). Moreover, additional studies found that rock clas-sification performances after training (without feature descrip-tions) were well accommodated by an exemplar-based model thatassumed equivalent attention to all eight dimensions in represent-ing the exemplars (Nosofsky, Sanders, & McDaniel, 2017). Itappears, however, that highlighting features did not result in func-tionally more impoverished encodings of the training items thanwhen features were not provided. This conclusion is based onseveral observations. First, when features were highlighted, clas-sification performance for trained items was high (over 80%correct) and virtually equivalent (even nominally better) to thatfound when features were not highlighted (see the old trainingperformances in Figures 6 and 8). Second, when the test wasdelayed for 2 days after training, performance for trained items didnot decline relative to when the test was immediate, and thispattern held when features were provided. Thus, classificationperformance for trained items was highly accurate and resistant toforgetting even in the feature-provision condition, suggesting thatproviding features did not produce impoverished encodings of thetraining items, at least under the present training conditions.

Theoretical Explanation of the FeatureHighlighting Benefit

One general hypothesis is that providing features speeds severalaspects of category learning. One idea developed was that featureprovision could help learners more quickly identify and orient todimensions along which rocks differ. If so, then the expectationwould be that learners in the feature-provided condition wouldhave a Head Start at encoding and differentiating rocks, whichmight benefit training initially (and possibly throughout). In op-position to this idea, the learning curves were not significantlysteeper or systematically better in the early stages of training forthe conditions in which features were provided than for conditionsin which features were not provided (see Figures 5 and 7).

Another idea was that providing features would facilitate theprocesses by which learners identify characteristic features of eachrock category. One such process might be hypothesis-testing.Within one successful rule learning model of ill-defined categories(like rocks), learners search for relatively simple rules that classifymany of the instances and then learn exceptions (e.g., RULEX;Nosofsky et al., 1994). The feature provision condition essentiallyidentified simple rules for the learners (e.g., “obsidian is glassyblack and has shell-like fracture”) and, thus, learning would be

Thi

sdo

cum

ent

isco

pyri

ghte

dby

the

Am

eric

anPs

ycho

logi

cal

Ass

ocia

tion

oron

eof

itsal

lied

publ

ishe

rs.

Thi

sar

ticle

isin

tend

edso

lely

for

the

pers

onal

use

ofth

ein

divi

dual

user

and

isno

tto

bedi

ssem

inat

edbr

oadl

y.

10 MIYATSU, GOURAVAJHALA, NOSOFSKY, AND MCDANIEL

expected to benefit, at least early on in training and possiblythroughout training, because the need for trial and error hypothesistesting would be minimized. For instance, other studies haveshown that training conditions that facilitate learning a useful (butnot perfect) rule produced steeper learning curves (Mathy & Feld-man, 2009). Clearly, such a pattern was not associated with thecategory learning benefit of feature highlighting: Highlightingfeatures did not change the rate of learning or the learning asymp-tote during training. Yet, providing features did produce betterperformance on a generalization test for untrained instances (Ex-periments 2 and 3), suggesting that providing features producedmore accurate acquisition of category-level information.

Another potential process by which characteristic features gainascendance is through learning appropriate attention weights to thecategory-relevant dimensions (e.g., ALCOVE; Kruschke, 1992).Providing features would be expected to facilitate setting effectiveattention-weights (e.g., in a connectionist implementation of anexemplar model; see Kruschke, 1992) or to allow the system topreset the weights (cf. Choi et al., 1993, for knowledge-basedpresetting of ALCOVE parameters). In particular, if observers givegreater attention to category-level-relevant features, then suchinformation would tend to be highly beneficial in generalizing tothe novel transfer instances. The reason is that the novel instancesare likely to vary in numerous idiosyncratic respects from the oldtraining items; instead, by definition, what they would share withthe training items are the invariant category-level-relevant fea-tures.

Without feature highlighting, by contrast, instead of biasingtheir attention toward category-relevant features, learners presum-ably focused on more idiosyncratic features associated with indi-vidual training instances. In favor of this claim, in rock categorylearning experiments in which features were not highlighted, anexemplar model with unbiased attention weights across eight di-mensions successfully accounted for rock category learning andgeneralization (Nosofsky, Sanders, & McDaniel, 2017). Idiosyn-cratic features would be less useful for supporting generalization,consistent with the worse test performance on untrained instancesfor conditions with no feature descriptions relative to conditionswith feature highlighting.

It is a more open question whether greater attention to charac-teristic features of the category (stimulated by feature highlighting)would lead to enhancements in performance on the old traininginstances themselves relative to no feature highlighting. On the onehand, attending to the category-level-relevant features might alsobe expected to benefit training performance, because those featuresare also useful for discriminating between the training examples ofthe contrasting categories. On the other hand, for these complexobjects, there are numerous idiosyncratic features associated withindividual instances that an observer could also use to supportinitial learning. For example, an observer may remember that aparticular rock with a red color patch in the upper-right locationwas an example of breccia, and this information would supportperformance on the training item. The detailed predictions fromalternative models (selective attention to characteristic features vs.attention to a range of idiosyncratic features as well) would dependon factors such as the precise configuration of objects in themultidimensional similarity space and the extent to which attentioncan be shared across the multiple features that compose the ob-jects.

The obtained patterns might also be consistent with the idea thathighlighting characteristic features during training promoted amore substantial qualitative difference in category representationsrelative to when these features were not provided. Specifically, thecurrent patterns share similarities with findings from the functionlearning domain in which learners show equivalent performancesduring training, but diverge in terms of accuracy in generalization(extrapolation; McDaniel, Cahill, Robbins, & Wiener, 2014; seealso Hoffmann, von Helversen, & Rieskamp, 2014). The theoret-ical account for those findings is that some learners acquiredexemplar representations during training, whereas others acquiredan abstraction of the function rule; both representations supportedgood training performance but the abstract representation wasmore useful for transfer. A speculative possibility is that whenfeatures are provided, the rock category representations (perhapsbased on exemplar storage) are enriched with some abstract sum-mary information (e.g., possibly rule-like as in RULEX; Nosofskyet al., 1994; or prototype-like as when training involves substantialnumber of instances; Homa, Sterling, & Trepel, 1981; see alsoRegehr & Brooks, 1993); whereas, when features are not provided,a solely exemplar-based representation is favored (but see Appen-dix B for the learning-strategy self-reports which did not substan-tially diverge as a function of whether features were provided).Specification of the nature of the category representations formedwhen features are provided relative to when features are notprovided clearly awaits further work.

A final theoretical observation is that, if our interpretationregarding the role of feature highlighting on selective attention tocategory-level information is correct, then our results have impli-cations for numerous formal models of category-learning. In par-ticular, for a wide variety of such models, the construct of selectiveattention plays a central role (e.g., differential weighting of com-ponent dimensions in computing distances to prototypes: Reed,1972; selective attention stretching and shrinking of the psycho-logical similarity space in which the exemplars are embedded:Kruschke, 1992; Nosofsky, 1986). In a similar vein, clusteringmodels (e.g., Love et al., 2004) and rule-based models (e.g., Bower& Trabasso, 1963; Levine, 1975; Nosofsky et al., 1994; Shepard etal., 1961) also place emphasis on selective attention.

As currently formalized, the key factors that are presumed toinfluence the patterns of dimensional selective attention in suchmodels include the inherent salience of the component dimensionsas well as the structure of the categories to be learned (e.g.,attention-optimization in an exemplar-based model, Nosofsky,1986, and trial-by-trial learning of optimal weights in a connec-tionist architecture, Kruschke, 1992; in rule-learning models ob-servers learn to selectively attend to the dimensions that allow forthe development of the most effective rules).The current articlepoints to another crucial mechanism that can influence patterns ofselective attention that has been somewhat neglected in theseformalized models: explicit instructions and clues provided by theteacher. To provide a complete account of category learning, themanner in which such explicit and illustrated clues are integratedwith other attention-learning mechanisms would need to be spelledout. Thus, the present work motivates an important direction forfurther development of a wide class of formal models of categorylearning.

Thi

sdo

cum

ent

isco

pyri

ghte

dby

the

Am

eric

anPs

ycho

logi

cal

Ass

ocia

tion

oron

eof

itsal

lied

publ

ishe

rs.

Thi

sar

ticle

isin

tend

edso

lely

for

the

pers

onal

use

ofth

ein

divi

dual

user

and

isno

tto

bedi

ssem

inat

edbr

oadl

y.

11FEATURE PROVISION IN NATURAL CATEGORY LEARNING

Practical Implications

The current findings are in line with the multimedia principlesof signaling and spatial contiguity (Mayer, 2002, 2005), and ex-tend their reach to learning of perceptually based natural catego-ries. These principles have previously been demonstrated in sci-ence concept learning (how brakes work; how planes fly). In thepresent Experiments 2 and 3, but not Experiment 1, the verbalfeature descriptions were positioned to be spatially contiguous tothe exact aspect of the rock image to which the descriptionsreferred; furthermore, those aspects were explicitly signaled (byencircling part of the rock image). Relative to Experiment 1, thespatial contiguity of the feature description presumably reducedextraneous processing (e.g., scanning the rock images to try to findthe feature) so that learners could devote more time to processingthe pertinent parts of the rock (e.g., Mayer, 1989). Signaling theexact parts that were described in the feature description wouldhelp learners appreciate the particular perceptual qualities of therelevant dimension (e.g., Mautone & Mayer, 2001), unlike inExperiment 1 where learners could not take advantage of thefeature descriptions placed below the images. In general, ourresults in conjunction with the literature on multimedia principlesprovide guidance for effective communication and placement offeature descriptions to enhance learning of natural perceptuallybased categories.

More specifically, our results have possible implications forinstruction of natural categories in educational settings. First, asillustrated in Figure 1 feature descriptions (at least for rocks) intextbooks are provided as captions placed adjacent to the image(e.g., above or below). Our results (Experiment 1) suggest that thisstandard practice might be revisited in further work. Such workcould take steps to achieve closer alignment (than did the presentexperiments) with actual textbook practice, such as using fewerexample images and perhaps familiarizing learners with terms inthe feature descriptions before learners viewing example rockimages. Second, to the extent that instructors fail to signal relevantaspects of the images when providing descriptions of characteristicfeatures of the categories, learning may not be benefitted. Instruc-tors’ intuitions that providing feature descriptions will enhancelearning of natural science categories—seem well founded, pro-vided that those features are explicitly signaled in the exampleimages or tokens used during instruction.

References

Allen, S., & Brooks, L. (1991). Specializing the operation of an explicitrule. Journal of Experimental Psychology: General, 120, 3–19. http://dx.doi.org/10.1037/0096-3445.120.1.3

Ashby, F. G., & Maddox, W. T. (1992). Complex decision rules incategorization: Contrasting novice and experienced performance. Jour-nal of Experimental Psychology: Human Perception and Performance,18, 50.

Ashby, F. G., & Maddox, W. T. (2005). Human category learning. AnnualReview of Psychology, 56, 149–178. http://dx.doi.org/10.1146/annurev.psych.56.091103.070217

Birnbaum, M. S., Kornell, N., Bjork, E. L., & Bjork, R. A. (2013). Whyinterleaving enhances inductive learning: The roles of discriminationand retrieval. Memory & Cognition, 41, 392–402. http://dx.doi.org/10.3758/s13421-012-0272-7

Bourne, L. E. (1974). An inference model for conceptual rule learning. In

R. L. Solso (Ed.), Theories in cognitive psychology: The Loyola Sym-posium. Oxford, England: Lawrence Erlbaum.

Bourne, L. E., Jr., Healy, A. F., Kole, J. A., & Graham, S. M. (2006).Strategy shifts in classification skill acquisition: Does memory retrievaldominate rule use? Memory & Cognition, 34, 903–913. http://dx.doi.org/10.3758/BF03193436

Bower, G., & Trabasso, T. (1963). Reversals prior to solution in conceptidentification. Journal of Experimental Psychology, 66, 409–418. http://dx.doi.org/10.1037/h0044972

Brooks, L. R. (1978). Nonanalytic concept formation and memory forinstances. In E. Rosch & B. Lloyd (Eds.), Cognition and categorization(pp. 169–211). Hillsdale, NJ: Erlbaum.

Busch, R., & Tasa, D. (2009). Laboratory manual in physical geology (8thed.). Upper Saddle River, NJ: Pearson.

Choi, S., McDaniel, M. A., & Busemeyer, J. R. (1993). Incorporating priorbiases in network models of conceptual rule learning. Memory & Cog-nition, 21, 413–423.

Eglington, L. G., & Kang, S. H. K. (2017). Interleaved PresentationBenefits Science Category Learning. Journal of Applied Research inMemory and Cognition, 6, 475–485. http://dx.doi.org/10.1016/j.jarmac.2017.07.005

Erickson, M. A., & Kruschke, J. K. (1998). Rules and exemplars incategory learning. Journal of Experimental Psychology: General, 127,107–140. http://dx.doi.org/10.1037/0096-3445.127.2.107

Goldstone, R. L. (2003). Learning to perceive while perceiving to learn. InR. Kimuchi, M. Behrmann, & C. Olson (Eds.), Perceptual organizationin vision: Behavioral and neural perspectives (pp. 233–278). Mahwah,NJ: Erlbaum.

Hartley, J., & Homa, D. (1981). Abstraction of stylistic concepts. Journalof Experimental Psychology: Human Learning and Memory, 7, 33–46.http://dx.doi.org/10.1037/0278-7393.7.1.33

Hoffmann, J. A., von Helversen, B., & Rieskamp, J. (2014). Pillars ofjudgment: How memory abilities affect performance in rule-based andexemplar-based judgments. Journal of Experimental Psychology: Gen-eral, 143, 2242–2261. http://dx.doi.org/10.1037/a0037989

Homa, D., Sterling, S., & Trepel, L. (1981). Limitations of exemplar-basedgeneralization and the abstraction of categorical information. Journal ofExperimental Psychology: Human Learning and Memory, 7, 418–439.http://dx.doi.org/10.1037/0278-7393.7.6.418

Jacoby, L. L., Wahlheim, C. N., & Coane, J. H. (2010). Test-enhanced learningof natural concepts: Effects on recognition memory, classification, andmetacognition. Journal of Experimental Psychology: Learning, Memory,and Cognition, 36, 1441–1451. http://dx.doi.org/10.1037/a0020636

Kang, S. H., & Pashler, H. (2012). Learning painting styles: Spacing isadvantageous when it promotes discriminative contrast. Applied Cogni-tive Psychology, 26, 97–103. http://dx.doi.org/10.1002/acp.1801

Kornell, N., & Bjork, R. A. (2008). Learning concepts and categories: Isspacing the “enemy of induction”? Psychological Science, 19, 585–592.http://dx.doi.org/10.1111/j.1467-9280.2008.02127.x

Kruschke, J. K. (1992). ALCOVE: An exemplar-based connectionistmodel of category learning. Psychological Review, 99, 22–44. http://dx.doi.org/10.1037/0033-295X.99.1.22

Levine, M. (1975). A cognitive theory of learning: Research on hypothesistesting. Oxford, England: Erlbaum.

Little, J. L., & McDaniel, M. A. (2015a). Individual differences in categorylearning: Memorization versus rule abstraction. Memory & Cognition,43, 283–297. http://dx.doi.org/10.3758/s13421-014-0475-1

Little, J. L., & McDaniel, M. A. (2015b). Some learners abstract, othersmemorize examples: Implications for education. Translational Issues inPsychological Science, 1, 158 –169. http://dx.doi.org/10.1037/tps0000031

Love, B. C., Medin, D. L., & Gureckis, T. M. (2004). SUSTAIN: Anetwork model of category learning. Psychological Review, 111, 309–332. http://dx.doi.org/10.1037/0033-295X.111.2.309

Thi

sdo

cum

ent

isco

pyri

ghte

dby

the

Am

eric

anPs

ycho

logi

cal

Ass

ocia

tion

oron

eof

itsal

lied

publ

ishe

rs.

Thi

sar

ticle

isin

tend

edso

lely

for

the

pers

onal

use

ofth

ein

divi

dual

user

and

isno

tto

bedi

ssem

inat

edbr

oadl

y.

12 MIYATSU, GOURAVAJHALA, NOSOFSKY, AND MCDANIEL

Mathy, F., & Feldman, J. (2009). A rule-based presentation order facilitatescategory learning. Psychonomic Bulletin & Review, 16, 1050–1057.http://dx.doi.org/10.3758/PBR.16.6.1050

Mautone, P. D., & Mayer, R. E. (2001). Signaling as a cognitive guide inmultimedia learning. Journal of Educational Psychology, 93, 377–389.http://dx.doi.org/10.1037/0022-0663.93.2.377

Mayer, R. E. (1989). Systematic thinking fostered by illustrations inscientific text. Journal of Educational Psychology, 81, 240–246. http://dx.doi.org/10.1037/0022-0663.81.2.240

Mayer, R. E. (2002). Multimedia learning. Psychology of Learning andMotivation, 41, 85–139. http://dx.doi.org/10.1016/S0079-7421(02)80005-6

Mayer, R. E. (Ed.). (2005). The Cambridge handbook of multimedialearning. New York, NY: Cambridge University Press. http://dx.doi.org/10.1017/CBO9780511816819

McDaniel, M. A., Cahill, M. J., Robbins, M., & Wiener, C. (2014). Individualdifferences in learning and transfer: Stable tendencies for learning exem-plars versus abstracting rules. Journal of Experimental Psychology: Gen-eral, 143, 668–693. http://dx.doi.org/10.1037/a0032963

Medin, D. L., & Schaffer, M. M. (1978). Context theory of classificationlearning. Psychological Review, 85, 207–238. http://dx.doi.org/10.1037/0033-295X.85.3.207

Meyer, B. J. F. (1975). The organization of prose and its effects onmemory. New York, NY: Elsevier.

Moreno, R., & Mayer, R. E. (1999). Cognitive principles of multimedialearning: The role of modality and contiguity. Journal of EducationalPsychology, 91, 358–368. http://dx.doi.org/10.1037/0022-0663.91.2.358

Murphy, G. L. (2002). The big book of concepts. Cambridge, MA: TheMIT Press.

Noh, S. M., Yan, V. X., Vendetti, M. S., Castel, A. D., & Bjork, R. A.(2014). Multilevel induction of categories: Venomous snakes hijack thelearning of lower category levels. Psychological Science, 25, 1592–1599. http://dx.doi.org/10.1177/0956797614535938

Nosofsky, R. M. (1984). Choice, similarity, and the context theory ofclassification. Journal of Experimental Psychology: Learning, Memory,and Cognition, 10, 104–114.

Nosofsky, R. M. (1986). Attention, similarity, and the identification-categorization relationship. Journal of Experimental Psychology: Gen-eral, 115, 39–57. http://dx.doi.org/10.1037/0096-3445.115.1.39

Nosofsky, R. M. (2011). The generalized context model: An exemplarmodel of classification. In M. Pothos & A. Wills (Eds.), Formal ap-proaches in categorization (pp. 18–39). New York, NY: CambridgeUniversity Press.

Nosofsky, R. M., & Kruschke, J. K. (1992). Investigations of an exemplar-based connectionist model of category learning. In G. H. Bower (Ed.),The psychology of learning and motivation (Vol. 28, pp. 207–250). SanDiego, CA: Academic Press. http://dx.doi.org/10.1016/S0079-7421(08)60491-0

Nosofsky, R. M., Palmeri, T. J., & McKinley, S. C. (1994). Rule-plus-exception model of classification learning. Psychological Review, 101,53–79. http://dx.doi.org/10.1037/0033-295X.101.1.53

Nosofsky, R. M., Sanders, C. A., Gerdom, A., Douglas, B. J., & McDaniel,M. A. (2017). On learning natural-science categories that violate thefamily-resemblance principle. Psychological Science, 28, 104–114.http://dx.doi.org/10.1177/0956797616675636

Nosofsky, R. M., Sanders, C. A., & McDaniel, M. A. (2017). Tests of anExemplar-Memory Model of Classification Learning in a High-Dimensional Natural-Science Category Domain. Journal of Experimen-

tal Psychology: General. Advance online publication. http://dx.doi.org/10.1037/xge0000369

Nosofsky, R. M., Sanders, C. A., Meagher, B. J., & Douglas, B. J. (2017).Toward the development of a feature-space representation for a complexnatural category domain. Behavior Research Methods. Advance onlinepublication. http://dx.doi.org/10.3758/s13428-017-0884-8

Palmeri, T. J., Wong, A. C., & Gauthier, I. (2004). Computational ap-proaches to the development of perceptual expertise. Trends in CognitiveSciences, 8, 378–386. http://dx.doi.org/10.1016/j.tics.2004.06.001

Posner, M. I., & Keele, S. W. (1968). On the genesis of abstract ideas.Journal of Experimental Psychology, 77, 353–363. http://dx.doi.org/10.1037/h0025953

Posner, M. I., & Keele, S. W. (1970). Retention of abstract ideas. Journalof Experimental Psychology, 83, 304–308. http://dx.doi.org/10.1037/h0028558

Pothos, E. M., & Wills, A. J. (Eds.). (2011). Formal approaches incategorization. New York, NY: Cambridge University Press. http://dx.doi.org/10.1017/CBO9780511921322

Reed, S. K. (1972). Pattern recognition and categorization. CognitivePsychology, 3, 382– 407. http://dx.doi.org/10.1016/0010-0285(72)90014-X

Regehr, G., & Brooks, L. R. (1993). Perceptual manifestations of ananalytic structure: The priority of holistic individuation. Journal ofExperimental Psychology: General, 122, 92–114. http://dx.doi.org/10.1037/0096-3445.122.1.92

Rosch, E. H. (1973). Natural categories. Cognitive Psychology, 4, 328–350. http://dx.doi.org/10.1016/0010-0285(73)90017-0

Rosch, E., & Mervis, C. B. (1975). Family resemblances: Studies in theinternal structure of categories. Cognitive Psychology, 7, 573–605.http://dx.doi.org/10.1016/0010-0285(75)90024-9

Salatas, H., & Bourne, L. E. (1974). Learning conceptual rules: III. Pro-cesses contributing to rule difficulty. Memory & Cognition, 2, 549–553.http://dx.doi.org/10.3758/BF03196919

Schyns, P. G., & Rodet, L. (1997). Categorization creates functionalfeatures. Journal of Experimental Psychology: Learning, Memory, andCognition, 23, 681–696. http://dx.doi.org/10.1037/0278-7393.23.3.681

Shepard, R. N., Hovland, C. I., & Jenkins, H. M. (1961). Learning andmemorization of classifications. Psychological Monographs, 75, 1–42.http://dx.doi.org/10.1037/h0093825

Sowden, P. T., Davies, I. R., & Roling, P. (2000). Perceptual learning ofthe detection of features in X-ray images: A functional role for improve-ments in adults’ visual sensitivity? Journal of Experimental Psychology:Human Perception and Performance, 26, 379–390. http://dx.doi.org/10.1037/0096-1523.26.1.379

Trabasso, T., & Bower, G. H. (1968). Attention in learning: Theory andresearch. New York, NY: Wiley.

Wahlheim, C. N., & DeSoto, K. A. (2017). Study preferences for exemplarvariability in self-regulated category learning. Memory, 25, 231–243.http://dx.doi.org/10.1080/09658211.2016.1152378

Wahlheim, C. N., Finn, B., & Jacoby, L. L. (2012). Metacognitive judg-ments of repetition and variability effects in natural concept learning:Evidence for variability neglect. Memory & Cognition, 40, 703–716.http://dx.doi.org/10.3758/s13421-011-0180-2

Wahlheim, C. N., McDaniel, M. A., & Little, J. L. (2016). Categorylearning strategies in younger and older adults: Rule abstraction andmemorization. Psychology and Aging, 31, 346–357. http://dx.doi.org/10.1037/pag0000083

(Appendices follow)

Thi

sdo

cum

ent

isco

pyri

ghte

dby

the

Am

eric

anPs

ycho

logi

cal

Ass

ocia

tion

oron

eof

itsal

lied

publ

ishe

rs.

Thi

sar

ticle

isin

tend

edso

lely

for

the

pers

onal

use

ofth

ein

divi

dual

user

and

isno

tto

bedi

ssem

inat

edbr

oadl

y.

13FEATURE PROVISION IN NATURAL CATEGORY LEARNING

Appendix A

Material Overview

Feature descriptions for each category as well as the rock pictures used in the current experiments. Visit Open Science Frame Work(https://osf.io/9vg8m/#) for high resolution version of these images and the ones with feature highlighting that were used in Experiments2 and 3.

Amphibolite – dark colored and weakly grained structure

Breccia – broken angular fragments cemented together

Conglomerate – rounded fragments cemented together

Gneiss – alternating dark and light colored bands with oriented crystals

Granite – light colored or reddish, interlocking coarse-grained crystals

Marble – gray to white crystalline material with darker swirls and veins

Obsidian – glassy black, scallop shell-like fracture

Pegmatite – uneven texture with large-size interlocking crystal components, can be greenish

(Appendices continue)

Thi

sdo

cum

ent

isco

pyri

ghte

dby

the

Am

eric

anPs

ycho

logi

cal

Ass

ocia

tion

oron

eof

itsal

lied

publ

ishe

rs.

Thi

sar

ticle

isin

tend

edso

lely

for

the

pers

onal

use

ofth

ein

divi

dual

user

and

isno

tto

bedi

ssem

inat

edbr

oadl

y.

14 MIYATSU, GOURAVAJHALA, NOSOFSKY, AND MCDANIEL

Pumice – rough textured volcanic glass with small holes

Rock Gypsum – silvery and chalky

Sandstone – sand-sized grains, often prominently layered

Slate – fine-grained, layered, homogeneous

Appendix B

Learning-Strategy Questionnaire Results

Results from the learning-strategy questionnaire responses. Thereported learning strategy did not modulate the learning perfor-mance and the final test performance.

Experiment 1

We computed average strategy ratings for the two feature de-scription (FD)-present conditions grouped together and the twoFD-absent conditions grouped together. There was no significantdifference in the strategy use ratings between the FD-presentconditions (M � 4.13, SD � 1.67) and the FD-absent conditions(M � 3.94, SD � 1.82), t(172) � 0.69, p � .05, d � 0.11. FigureB1 shows the number of participants who reported each rating asa function of presence or absence of feature description.

Experiment 2

There was no significant difference in the strategy use ratingsbetween the FD-present condition (M � 5.18, SD � 1.50) and theFD-absent condition (M � 5.10, SD � 0.98), t(78) � 0.27, p �.05, d � 0.06. Despite the similarity in the tasks between Exper-iments 1 and 2, participants’ self-reported strategy seemed to bemore rule-based in Experiment 2. This is evident when comparingthe extreme ratings (21% reported relying on memorization and

26% reported relying on rules in Experiment 1 versus 5 and 43%,respectively, in Experiment 2, see Figures B1 vs. B2) and whencomparing the average ratings (M � 4.03, SD � 1.75 in Experi-ment 1 vs. M � 5.14, SD � 1.26 in Experiment 2, t(252) � 5.07,p � .001, d � 0.73). Figure B2 shows the number of participantswho reported each rating as a function of presence or absence offeature description. We tentatively suggest that the strategy differ-ences could reflect the differential populations from which partic-ipants were sampled in Experiments 1 (Amazon Mechanical Turk)and 2 (students from a highly selective private university). Thekeyfinding remains that providing features did not appear tochange learners’ self-reported strategies relative to not providingfeatures.

Experiment 3

Participants’ reported more rule-based strategies than in the priorexperiments, and there were no extreme memorization-based strate-gies reported (see Figure B3) In interpreting this pattern, we cautionthat participants’ memory of how they tried to learn 2 days beforeproviding the self-report is likely less accurate than in the previousexperiments. Accordingly, perhaps participants relied primarily onhow they tried to classify the items on the just given test task.

(Appendices continue)

Thi

sdo

cum

ent

isco

pyri

ghte

dby

the

Am

eric

anPs

ycho

logi

cal

Ass

ocia

tion

oron

eof

itsal

lied

publ

ishe

rs.

Thi

sar

ticle

isin

tend

edso

lely

for

the

pers

onal

use

ofth

ein

divi

dual

user

and

isno

tto

bedi

ssem

inat

edbr

oadl

y.

15FEATURE PROVISION IN NATURAL CATEGORY LEARNING

Received July 24, 2017Revision received November 27, 2017

Accepted December 2, 2017 �

Figure B1. The number of participants who reported each rating (where 1 indicates a strong preference for amemorization strategy and 7 indicates a strong preference for a rule-abstraction strategy) as a function ofpresence or absence of feature description in Experiment 1 postexperimental questionnaire.

Figure B2. The number of participants who reported each rating (where 1indicates a strong preference for a memorization strategy and 7 indicates astrong preference for a rule-abstraction strategy) as a function of presence orabsence of feature description in Experiment 2 postexperimental questionnaire.

Figure B3. The number of participants who reported each rating (where 1indicates a strong preference for a memorization strategy and 7 indicates astrong preference for a rule-abstraction strategy) as a function of presence orabsence of feature description in Experiment 3 postexperimental questionnaire.

Thi

sdo

cum

ent

isco

pyri

ghte

dby

the

Am

eric

anPs

ycho

logi

cal

Ass

ocia

tion

oron

eof

itsal

lied

publ

ishe

rs.

Thi

sar

ticle

isin

tend

edso

lely

for

the

pers

onal

use

ofth

ein

divi

dual

user

and

isno

tto

bedi

ssem

inat

edbr

oadl

y.

16 MIYATSU, GOURAVAJHALA, NOSOFSKY, AND MCDANIEL

View publication statsView publication stats