Content uploaded by Rubi Hammer
Author content
All content in this area was uploaded by Rubi Hammer on Oct 10, 2017
Content may be subject to copyright.
PAPER
Developmental changes between childhood and adulthood in
passive observational and interactive feedback-based
categorization rule learning
Rubi Hammer,
1,2
Jim Kloet
1
and James R. Booth
1,2,3
1
1. Department of Communication Sciences and Disorders, Northwestern University, USA
2. Interdepartmental Neuroscience Program, Northwestern University, USA
3. Department of Communication Sciences and Disorders, The University of Texas at Austin, USA
Abstract
As children start attending school they are more likely to face situations where they have to autonomously learn about novel
object categories (e.g. by reading a picture book with descriptions of novel animals). Such autonomous observational category
learning (OCL) gradually complements interactive feedback-based category learning (FBCL), where a child hypothesizes
about the nature of a novel object, acts based on his prediction, and then receives feedback indicating the correctness of his
prediction. Here we tested OCL and FBCL skills of elementary school children and adults. In both conditions, participants
performed complex rule-based categorization tasks that required associating novel objects with novel category-labels. We
expected children to perform better in FBCL tasks than in OCL tasks, whereas adults to be skilled in both tasks. As
hypothesized, in early-phase learning children performed better in FBCL tasks than in OCL tasks. Unexpectedly, adults
performed somewhat better in OCL tasks. Early-phase FBCL performance in the two age groups was matched, but the OCL
performance of adults was higher than that of children. In late-phase learning there was only an age group main effect
(adults >children). Moreover, performance in post-learning categorization tasks that did not require label recollection
indicated that in FBCL tasks children were likely to directly learn the associations between an object and a category label,
whereas in the OCL tasks they were likely to first learn which feature-dimensions were relevant. These findings shed light on
developmental changes in cognitive control and learning mechanisms. Implications for educational settings are discussed.
Research highlights
•Observational and feedback-based categorization
rule learning skills of children and adults were tested.
•Children performed best in feedback-based category
learning tasks.
•Children showed inferior learning skills, as compared
to those of adults, primarily in observational cate-
gory learning tasks.
•In feedback-based category learning tasks. children
were likely to directly learn the associations between
an object and a category label.
•In the observational category learning tasks, children
were likely to learn which feature-dimensions were
relevant for categorization prior to learning object–
label associations.
Introduction
A primary challenge for the developing child is to render
the constant flow of sensory information into a coherent
mental representation of the surrounding environment.
A key cognitive process underlying this capacity is
Category Learning (CL), which enables the child to
categorize objects or events in a meaningful way by
inferring which attributes are important and which are
not (Hanania & Smith, 2010; McColeman & Blair, 2013;
Address for correspondence: Rubi Hammer, Developmental Cognitive Neuroscience Lab, Department of Communication Sciences and Disorders,
2240 Campus Drive, Northwestern University, Evanston, IL 60208-2952, USA; e-mails: rubi.hammer@northwestern.edu; rubihammer@gmail.com
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
©2015 John Wiley & Sons Ltd
Developmental Science (2015), pp 1–15 DOI: 10.1111/desc.12338
D E S C 12338
Dispatch: 1.8.15 CE: Wiley
Journal Code Manuscript No.
No. of pages: 15 PE: Dhatchayani S
Sloutsky, 2010). Being a fundamental cognitive skill, CL
starts developing very early (Best, Yim & Sloutsky, 2013;
Graham & Diesendruck, 2010; Feldman, Myers, White,
Griffiths & Morgan, 2013; Needham, Goldstone &
Wiesen, 2014). Early in development, CL often relies
on active exploration that involves trial-and-error or
feedback-based learning. Here the child acts based on
limited knowledge and uses the sensory input that
becomes available following her actions for updating
and expanding her knowledge (Bechtel, Jeschonek &
Pauen, 2013; Markman & Ross, 2003). For example, a
child can learn about distinct subcategories of citrus
fruits by eating some oranges and finding them to be
sweet, and then contrasting these experiences with the
bitterness of grapefruit. Such experiences ultimately
shape the child’s expectations and behavior, enabling
her to instantly categorize citrus fruits based on their
appearance.
CL does not always require the child to experiment by
herself. By observing the actions of others and the
following outcomes, a child may infer which of her own
actions are likely to be rewarding (Charman, Baron-
Cohen, Swettenham, Baird, Cox et al., 2000; Meltzoff,
1988a, 1988b; Sommerville, Woodward & Needham,
2005; Walker & Gopnik, 2014). Currently it is debatable
whether in such scenarios the child uses distinct learning
mechanisms for interpreting other people’s actions, or
uses generic learning mechanisms similar to those used
for learning from her own actions (Korman, Voiklis &
Malle, 2015). It is agreed, however, that whenever a child
watches the outcomes of other people’s actions, she may
rely on her feedback and reward processing skills
(Alvarez & Booth, 2014; Bohlmann & Fenson, 2005;
Meyer, Bekkering, Janssen, de Bruijn & Hunnius, 2014),
or her causal inference skills (Booth & Alvarez, 2015;
Cohen, Rundell, Spellman & Cashon, 1999; Griffiths &
Tenenbaum, 2009; Smith, 2013).
Conversely, in some scenarios children are expected to
learn autonomously, without immediate corrective feed-
back, or without observing other people’s actions as a
reference. For example, a child can be asked to learn
about new categories of animals, novel objects or distant
places by reading a book with labeled pictures (e.g. ‘this
is an antelope’), or pictures supplemented with a short
written description highlighting distinct category char-
acteristics (e.g. ‘unlike the deer, the antelope has bony
horns’). As children start attending school, this form of
learning becomes more significant (Andre & Thieman,
1988; Butler, Godbole & Marsh, 2013; Forbus, Riesbeck,
Birnbaum, Livingston, Sharma et al., 2007; Robinson &
Best, 2012). In order to be effective, such autonomous
observational CL requires executive skills enabling the
child to be engaged in an effortful learning process
without receiving immediate confirmation for the cor-
rectness of her insights.
Here we investigated developmental changes in active
feedback-based CL (FBCL) versus autonomous obser-
vational CL (OCL) skills in order to test the degree to
which elementary school children rely on an interactive
learning environment, and how this reliance changes at
adulthood. We used complex rule-based CL tasks
(conjunction rule) where two binary feature-dimensions
were relevant for categorizing creature-like stimuli into
four categories (not the same as, but similar to Edmunds,
Milton & Wills, 2015; McColeman, Barnes, Chen, Meier,
Walshe et al., 2014). Such rule-based CL tasks require
effective differentiation between task-relevant and task-
irrelevant feature-dimensions, and they are considered to
be declarative learning tasks that rely on inference skills
and executive control (Huang-Pollock, Maddox & Kar-
alunas, 2011; Huang-Pollock, Maddox & Tam, 2014).
Elementary school age was found to be a critical period,
where children start exhibiting adult-like rule-based CL
skills (Hammer, Diesendruck, Weinshall & Hochstein,
2009a; Rabi & Minda, 2014).
FBCL operationally involves perceiving an object,
categorizing the object based on (current) limited sub-
jective knowledge, and then using supervisory informa-
tion (feedback) that becomes available following the
categorization decision, for updating knowledge. In
contrast, OCL involves perceiving an object while being
provided with its category identity prior to being
required to categorize the object. Earlier studies with
adults showed that OCL and FBCL are based on distinct
neurocognitive mechanisms. In one study, Ashby, Mad-
dox and Bohil (2002) found better performance in
FBCL, as compared with OCL, in information-integra-
tion tasks (where categories were determined by two
interdependent feature-dimensions). This was contrasted
with no differences between OCL and FBCL tasks that
required learning a simple rule (considered as an easier
task). Consistent with these findings, FBCL, but not
OCL, was found to be impaired in elderly people with
Parkinson’s disease. This suggests that OCL and FBCL
rely on distinct neurocognitive mechanisms with distinct
susceptibility to aging-related disorders (Reber & Squire,
1999; Shohamy, Myers, Grossman, Sage, Gluck et al.,
2004). Others showed that FBCL results in better
performance than OCL in both rule-based and informa-
tion-integration tasks (Edmunds et al., 2015). These
studies (see also Ramscar, Yarlett, Dye, Denny &
Thorpe, 2010) suggest that performance in OCL can
match FBCL in some scenarios, but in most scenarios it
would yield lower performance (at least in healthy
adults). Other recent findings suggest that FBCL may
bias participants to identify features that differentiate
©2015 John Wiley & Sons Ltd
2 Rubi Hammer et al.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
between categories, whereas OCL is more effective for
learning the internal structure of categories, such as
correlations between feature-dimensions within a cate-
gory (Carvalho & Goldstone, 2014; Levering & Kurtz,
2014).
Clearly, much is still unknown about the differences
between OCL and FBCL, and even less is known about
how these two learning skills develop (there are no
developmental studies addressing this question). Since
cognitive skills required for autonomous OCL are more
frequently used starting at elementary school age, after
the child starts reading, we hypothesized that this form
of learning would be less developed than FBCL in
elementary school children. Consequently, we expected
children to perform best when they are ‘forced’to
generate a hypothesis regarding a novel object category
by explicitly categorizing it, followed by receiving feed-
back that indicates whether their categorization decision
was right or wrong. That is, we expected children to
perform better in FBCL than in OCL, and to find
smaller differences between the two CL skills in adults.
To test this hypothesis we compared rule-based CL
skills of adults with those of elementary school children
(second to fifth grade). We tested participants from both
age groups in FBCL and OCL tasks. We kept the two
types of CL tasks identical in stimuli statistics, the
objective information that could become available to the
participants during learning, and the motor responses
the participants had to produce. Assessing the contribu-
tion of an interactive learning environment required an
experimental design that deviated from those used in
earlier studies (see Discussion section for possible
implications). Here, FBCL trials started with the
presentation of a novel object (creature) and a novel
label. The participant was then required to press a key
indicating whether she thinks that the creature-label
association is correct or incorrect (see Maddox, Bohil &
Ing, 2004, for a discussion regarding the usability of
such a procedure in rule-based CL tasks). The partic-
ipant key-press was followed by corrective feedback,
indicating whether the decision was right (smiling face)
or wrong (sad face). If the participant did not make any
decision, she was not provided with informative feed-
back, and instead was only informed about not pressing
any key on time (neutral face). That is, in FBCL tasks,
failure to take action resulted in not receiving informa-
tion essential for learning. In OCL trials, the presen-
tation of the novel creature and label was followed by
the presentation of a cue indicating the correctness of
the creature–label association (smiling face =correct
association, and sad face =incorrect association). Only
then was the participant required to press a key,
‘approving’the answer provided by the cue (essentially,
the participant had to repeat what she had just been
told). That is, in OCL tasks constructive information
was made available regardless of the participant’s
actions.
Methods
Participants
Forty adults (Mean age =23.9 STD =5.9 years old;
minimum age =18.0, maximum age =38.0; 22 females)
and 40 elementary school children (9.8 1.0; 7.7 to
11.7; 23 females), with normal or corrected to normal
vision and no known history of psychiatric or cognitive
disorders participated in the study. Seven additional
children were excluded from the analysis due to poor
performance (near chance) in both experimental condi-
tions. A sample size of 40 participants in each age group
enabled the proper investigation of differences between
the two CL conditions in each group, and between-group
differences in each condition (Friston, Holmes & Wors-
ley, 1999). Participants were recruited from the Chicago
metropolitan area and gave their informed consent
(parental consent for children) in accordance with the
policies of the Institutional Review Board (IRB) at
Northwestern University. Participants received 15 US
dollars for each hour of participation (total of $45 to
$60).
Materials and settings
We used a mixed experimental design where each
participant completed two experimental CL tasks in
each of the two experimental conditions (FBCL and
OCL). For each CL task a distinct stimuli set was used.
We used a total of nine distinct sets of creature-like
stimuli (rendered using Autodesk 3D Studio Max R8
â
and Adobe Photoshop CS2
â
). Three stimuli sets that
were used in an earlier study (Hammer at al., 2009a)
were used here exclusively for training the participants.
Six additional stimuli sets that were created for the
current study were used in the experimental tasks (for
each participant, four out of these six stimuli sets were
used). Each of the nine stimuli sets included 16 distinct
creatures that differed from one another in four binary
feature-dimensions (e.g. the shape of the creature’s feet
or body). Each creature was rendered in two distinct
points of view, adding task-irrelevant perceptual vari-
ability to the stimuli. In each task, the four feature-
dimensions by which the stimuli varied were perceptually
salient. Using multiple stimuli sets ensured that the
overall pattern of performance would not be biased due
©2015 John Wiley & Sons Ltd
Development of autonomous category learning skills 3
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
to specific stimuli characteristics and feature saliency
(Diesendruck, Hammer & Catz, 2003; Hammer &
Diesendruck, 2005; Hammer, Sloutsky & Grill-Spector,
2015; Hammer, 2015). In each CL task, participants had
to learn which two feature-dimensions are relevant for
categorizing the stimuli into four distinct categories or
‘subspecies’, and to ignore the two irrelevant feature-
dimensions (Figure 1a). The assignment of learning
condition to each stimuli set and the order of learning
conditions were counterbalanced for each participant.
For each creature set we created a corresponding set of
written category labels. Each label set included four
distinct category labels, one for each creature subspecies.
Labels were composed of two syllables, each three letters
long. Each syllable corresponded to one of the two task-
relevant feature-dimensions, describing a specific cate-
gory feature (e.g. flat feet vs. two-toe feet, and large head
vs. small head creatures). For each label we used both an
uppercase (e.g. ‘A POKYED’) and a lowercase (‘a
pokyed’) version. This added irrelevant perceptual vari-
ability, forcing participants to read the labels (rather than
using labels as holistic perceptual feature). For each
stimuli set labels were created using a distinct font
(Figure 1b).
Each creature category included four creatures that
were identical in the two task-relevant feature-dimen-
sions (e.g. feet and head) but could differ in the two
irrelevant feature-dimensions (e.g. body and wings/tail)
and in point of view. The first syllable of the written label
corresponded to one relevant feature-dimension (e.g.
‘POK/pok’=large-head and ‘LUX/lux’=small-head)
and the second syllable corresponded to the other
relevant feature-dimension (e.g. ‘MAG/mag’=two-toe
feet and ‘YED/yed’=flat feet). The four creature
categories differed from one another in at least one of
the two task-relevant feature-dimensions (Figure 1c).
Figure 1 Stimuli. (a) Two representative orthogonal creature stimuli from each one of the nine stimuli sets. In each set stimuli
varied in four binary feature-dimensions, out of which two were task relevant (lower row are stimuli from the sets used for training).
(b) Two representative labels from each one of the nine stimuli sets. Each label was composed of two syllables where each syllable
corresponded with one of the two task-relevant creature feature-dimensions. (c) Composition of the four creature categories in one
stimuli set. Categories differed from one another in at least one of the two task-relevant feature-dimensions (e.g. heads and feet).
Within each category, creatures differed in the two irrelevant feature-dimensions (e.g. bodies and wings/tails) and in the point of
view (see Appendix S1 in the Supporting Information).
COLOR
©2015 John Wiley & Sons Ltd
4 Rubi Hammer et al.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
E-Prime
â
2.0 (Psychology Software Tools, Inc.) was
used for stimuli presentation and for recording partici-
pants’responses. Stimuli were presented on a 19ʺ
1280 91024 pixels computer display. The creature
stimulus occupied 600 9600 pixels at the center of the
display. The label stimulus was displayed above the
creature stimulus. Stimuli were presented on a neutral-
gray background. Participants were seated so their heads
were within 50–80 centimeters from the computer
display, and they made their responses using the ‘1’
(incorrect creature-label association; using left hand
index finger) and ‘3’(correct creature-label association;
using right hand index finger) keys on the keyboard
number pad.
Procedure overview
After providing informed consent, the participant com-
pleted a standardized testing session, assessing cognitive
skills relevant to the experimental tasks. Within a few
days of completing the standardized testing session, the
participant came to the first experimental session and
performed two experimental CL tasks (following a
practice task), where both tasks were in the same
experimental condition (either FBCL or OCL). A few
days after the first experimental session, the participant
came to the second session where she performed CL
tasks in the second experimental condition.
Standardized testing session
In the first session participants completed a battery of
standardized tests, administered by a trained research
person in a quiet room. Standardized testing consisted of
seven tests from the Woodcock & Johnson Tests of
Achievement and Tests of Cognitive Abilities (version
III; Woodcock, McGrew & Mather, 2001), used to assess
the participant ability in reading, visual processing, and
rule-based learning tasks. The entire testing battery took
approximately 45 minutes for adults, and approximately
75 minutes for children (see Appendix S2 in the Sup-
porting Information for more details and the partici-
pants’scores).
Experimental sessions
Instructing participants and practice tasks
In each experimental session, participants completed one
practice task and two experimental tasks all in the same
experimental conditions. That is, the participant com-
pleted only OCL tasks in one session and only FBCL in
the other. Order of experimental conditions was coun-
terbalanced across participants. At the beginning of the
first experimental session the participant was first
introduced to a simple categorization task using more
familiar stimuli (subtypes of dogs). The participant had
to verbally specify which attribute was important for
differentiating between the two subtypes of dogs and
which was not (i.e. fur length or fur color). After this
introduction the participant completed a short example
of the CL tasks, where the duration of each trial was
longer than the trial duration in the actual experimental
tasks. During this short task the experimenter clarified
the task requirements. Following this introduction, the
participant performed a full-length practice CL task
simulating the experimental tasks (see details below),
with little intervention by the experimenter. This enabled
the experimenter to confirm that the participant under-
stood the task requirements and that she was capable of
performing the task. Data from the practice tasks were
not used in the reported data analysis.
Experimental CL tasks
In each experimental session (experimental condition)
the participant performed two CL tasks each based on a
distinct stimuli set. Each CL task included nine test
blocks that alternated with eight learning blocks. Each
block included eight trials. The duration of a trial in the
test blocks was 4.5 seconds, and the duration of a trial in
a learning block was 5.5 seconds (with an additional
0.5 second inter-trial interval). At the end of each test
block the participant was presented for 4 seconds with
her overall performance score for that block (a score
between 0/8 and 8/8 correct responses) as well as with her
on-time response rate (also ranging between 0/8 and 8/8).
The overall duration of each CL task was 13 minutes,
and it had to be completed without any breaks (see
Figure 2 for illustrations of trials and tasks composi-
tion). After completing the first CL task, the experi-
menter asked the participant if she would like to take a
short break before continuing to the second CL task.
OCL tasks and FBCL tasks differed only in the
composition of trials in the learning blocks. An obser-
vational learning trial (Figure 2a) started with the
presentation of a creature stimulus (0.5 seconds) fol-
lowed by a simultaneous presentation of the label and
the creature stimulus (3 seconds). Next, for 1 second, a
cue was presented in the center of the display, indicating
whether the creature–label association was correct (smil-
ing face) or incorrect (sad face). The trial was concluded
with the presentation of a go-signal (green circled-cross)
presented for 1 second, during which the participant had
to confirm the correctness of the creature–label associ-
ation by pressing either the right (‘3’on the keyboard’s
©2015 John Wiley & Sons Ltd
Development of autonomous category learning skills 5
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
number pad) key (correct association) or left (‘1’on the
keyboard’s number pad) key (incorrect association). In
feedback-based learning trials (Figure 2b), following the
stimuli presentation, the go-signal was presented for
1 second. Here the participant had to decide about the
correctness of creature–label association prior to being
provided with any indication about the association
correctness. Following the participant’s decision, feed-
back was presented for 1 second, indicating to the
participant whether her categorization decision was
correct (smiling face) or incorrect (sad face). Not
pressing any key on-time resulted in the presentation of
a neutral face. Trials in the test blocks were identical in
both experimental conditions. As in the learning blocks,
participants were presented first with a creature stimulus
(0.5 seconds), followed by simultaneous creature and
label presentation (3 seconds), followed by the go-signal
(1 second) during which participants responded (Fig-
ure 1c).
The composition of each learning trial maximized its
information content, ensuring that all learning trials
contribute to the inference of the categorization rule (see
Hammer, Bar-Hillel, Hertz, Weinshall & Hochstein,
2008; Hammer et al., 2009a; Hammer et al., 2015). In
trials with correct creature-label association the two
relevant features matched the two syllables. For example,
referring to Figure 1c, a ‘POKMAG’always had a large-
head and two-toe feet, but across learning trials different
POKMAGs differed in their body and wing/tail compo-
sition. In a given learning trial with incorrect creature–
label association, there was a mismatch between only one
feature and one syllable. For example, in an OCL trial,
the participant was told that the association between a
large-head and two-toe creature with the label
‘POKYED’is incorrect. In a later trial the participant
was told that the association between such a creature and
the label ‘LUXMAG’is also incorrect. This enabled
deducing, across multiple trials, the specific feature–
syllable association. This could not have been deduced if
trials with a simultaneous mismatch between the two
relevant features and the two syllables had been used (e.g.
being told that a large-head, two-toe creature is not a
‘LUXYED’). In FBCL trials, the correctness of a
creature–label (or feature–syllable) association could be
deduced from the feedback that followed the participant
response, indicating the response correctness.
Similarly, in test trials with correct creature–label
association the two relevant features matched the two
syllables. In a test trial with incorrect creature–label
association, there was a mismatch between only one
feature and one syllable. Out of the eight trials in each
test block, in two trials there was an incorrect associ-
ation between the first syllable and its corresponding
feature, and in two other trials there was an incorrect
association between the second syllable and its corre-
sponding feature. The four other trials were trials with
correct creature–label association. That is, the partici-
pant could perform the task perfectly only by learning
the correct association between both syllables and their
corresponding features. The two irrelevant feature-
dimensions corresponded with the labels in a similar
pattern, preventing learning which feature-dimensions
Figure 2 Trial and task composition. (a) Observational learning trials started with the stimuli presentation, continued with a cue
indicating whether the creature–label association was correct or incorrect, and concluded with a response interval where the
participant responded based on what was indicated by the cue (target stimuli presentation, followed by informative cue, followed by
an action). (b) Feedback-based learning trials started with the stimuli presentation, continued with a response interval, and
concluded with feedback presentation indicating whether the categorization decision was correct or incorrect (target stimuli
presentation, followed by an action, followed by corrective feedback). (c) Test trials included only stimuli presentation and a
response interval. (d–e) Each CL task included nine test blocks (T1 to T9) that alternated with eight learning blocks (L1 to L8). Each
test or learning block included eight trials.
COLOR
©2015 John Wiley & Sons Ltd
6 Rubi Hammer et al.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
are task relevant based on the stimuli presentation
statistics.
During the experimental tasks the experimenter sat
behind the participant (outside the participant’s field of
view).
Post-CL categorization tasks
In each one of the two experimental sessions, after
concluding the CL tasks, the participant underwent
three short categorization tasks that were based on the
same three stimuli sets on which she was tested in the CL
tasks (again, one was used for practice and two as
experimental tasks). Here, in each trial the participant
had to decide whether the two creatures presented
simultaneously on the display were from the same
category (pressing ‘3’) or different categories (pressing
‘1’), based on the categorization rule she had learned
earlier in the corresponding CL task. In these post-CL
categorization tasks category labels were not used, and
thus correct categorization only indicated the partici-
pant’s capacity to identify which feature-dimensions
were relevant for categorization. Each one of these tasks
was 32 trials long, where each trial was 3.5 seconds long
(2.5 seconds for stimuli presentation, 1 second for
response) with a 0.5 second inter-trial interval. In
same-category trials (half of the trials) the two creatures
presented together were identical in the two task-
relevant feature-dimensions and differed in the two
irrelevant feature-dimensions. In different-categories tri-
als the two creatures differed in one of the two relevant
feature-dimensions and one of the two irrelevant feature-
dimensions, and were identical in the other two feature-
dimensions. These tasks enabled the assessment of
participants’categorization rule learning independent
of their capacity to recall the category labels or creature–
label associations.
Post-CL labels recollection tasks
Finally, after concluding the post-CL categorization
tasks, the participant completed three short label recol-
lection tasks (one practice and two experimental) using
the labels presented during the CL tasks. In each trial the
participant was presented only with a label that could be
either identical to one of the category labels used earlier
in the session in a CL task (familiar label), or a label in
which the letters in one of the two syllables were reversed
in their order (altered label). The participant had to
decide whether the label was used earlier as a category
label (pressing ‘3’) or not (pressing ‘1’). Each one of
these tasks was 32 trials long, and each trial was
3.5 seconds long (2.5 seconds for stimuli presentation,
1 second for response and 0.5 second inter-trial inter-
val). These tasks enabled the assessment of participants’
label recollection capacities independently of their CL
skills.
Performance measures
For CL tasks, we defined a ‘Hit’as correctly deciding
that a creature–label association was correct, and a
‘False-Alarm’as incorrectly deciding that a creature-
label association was correct. Based on the Hit and
False-Alarm rate in each test block (or learning block)
we calculated the categorization accuracy using the non-
parametric measure A-prime (Grier, 1971; Stanislaw &
Todorov, 1999). A-prime =0.5 indicates chance-level
performance and A-prime =1.0 indicates perfect per-
formance. A-prime scores close to 0.0 indicate that a
participant is sensitive to the categorization rule but
reversed her responses. Hit rate and False-Alarm rate are
defined as:
H¼Hit rate ¼Hits
Hits þMisses
F¼False Alarm rate ¼False Alarms
False Alarms þCorrect Rejections
A-prime is defined as:
A0¼0:5þsignðHFÞ ðHFÞ2þjHFj
4maxðH;FÞ4HF
The equation term ‘sign’returns ‘+1’if ‘H>F’,‘1’
if ‘F>H’,and‘0’if ‘F=H’. In rare cases where the
denominator in any of the above equations was zero, a
value of 0.5 (chance-level) was assigned to the equation
product. For each participant we calculated the A-
prime score separately for each block in each experi-
mental task. The A-prime score in the first, pre-
learning, test block (T1) was used as the participant
performance baseline, and the A-prime scores in the
following test blocks (T2 to T9) and learning blocks
(L1 to L8) were rescaled, respectively, to reflect
performance change from pre-learning. After rescaling
the A-prime scores, due to the relatively small number
of trials (eight) in each test block, we smoothed the
data across the eight post-learning test blocks using a
running average with a window of three test blocks (see
similar analyses in Cohen & Schneidman, 2013, and
Hammer, Sloutsky & Grill-Spector, 2012). This reduced
the impact of intermediate trends in the learning
trajectories, while preserving generic characteristics of
the trajectories.
Data from the learning blocks (see Appendix S3 in the
Supporting Information) were not used for the primary
©2015 John Wiley & Sons Ltd
Development of autonomous category learning skills 7
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
analysis since data from the observational learning trials
did not reflect the participants’knowledge.
Results
A two-way ANOVA, with the mean A-prime score in
test blocks T2 to T9 as a dependent variable, shows a
significant age group by learning condition interaction,
F(1, 78) =5.96, p<.02, g2
p=0.071. There was a
significant age group main effect, F(1, 78) =29.74,
p<.0001, g2
p=0.276, but no condition main effect, F
(1, 78) =0.05 (p-values are Geisser-Greenhouse cor-
rected). Post-hoc analysis shows a significant perfor-
mance difference between the two experimental
conditions in the adults group, t(39) =2.11, p<.05,
d=0.676 (OCL >FBCL; Figure 3a). There were no
significant differences in performance between the two
experimental conditions in children, t(39) =1.43,
p=.161; though the trend was in the hypothesized
direction (FBCL >OCL), and it contributed to the
significance of the interaction (Figure 3b). There were
significant differences between the two age groups both
in the OCL condition, t(78) =5.75, p<.0001, d=
1.302 (equal variance not assumed) and in the FBCL
condition, t(39) =2.86, p<.005, d=0.648 (adults >
children in both conditions).
Having multiple learning blocks enabled participants
to adjust to the challenging learning tasks by switching
between learning strategies during the task. In such
conditions, it is likely that earlier phase performance
reflected cognitive skills employed by participants in
everyday life scenarios, whereas later phase performance
may have reflected their capacity to adapt to the
experimental task. This is supported by a trend toward
a significant age group by learning condition by test
block (T2 to T9) quadratic contrast, F(1, 78) =3.74,
p=.057, g2
p=0.046, showing that differences in perfor-
mance between conditions in the two age groups were
not consistent throughout the CL tasks (i.e. were
minimal midway through). Accordingly, we conducted
separate analyses for the early (T2 to T4) and late (T7 to
T9) CL phases. A two-way ANOVA, with the mean A-
prime in test blocks T2 to T4 (early-phase) as a
dependent variable, shows a significant age group by
learning condition interaction, F(1, 78) =7.63, p<.01,
g2
p=0.089 (Figure 3c), a significant age group main
effect, F(1, 78) =9.59, p<.005, g2
p=0.109, but no
condition main effect, F(1, 78) =0.10. Post-hoc analysis
shows a trend toward a significant difference between the
two experimental conditions in adults, t(39) =1.75,
p=.088, d=0.560 (observational >feedback-based).
Consistent with our hypothesis, here we found a signif-
icant difference between the two experimental conditions
in children, t(39) =2.15, p<.04, d=0.686 (FBCL >
OCL). Also consistent with our hypothesis, we found
significant differences between the two age groups in
the OCL condition, t(78) =4.26, p<.0001, d=0.965
(adults >children), but not in the FBCL condition, t
(78) =0.54.
A two-way ANOVA, with mean A-prime in test blocks
T7 to T9 (late-phase) as a dependent variable, shows only
a trend toward a significant age group by learning
condition interaction, F(1, 78) =3.11, p=.082, g2
p=
0.038 (Figure 3d), a significant age group main effect, F
(1, 78) =35.51, p<.0001, g2
p=0.313, but no condition
main effect, F(1, 78) =0.22. Post-hoc analyses showed a
trend toward a significant difference between the two
experimental conditions in adults, t(39) =1.70, p=.097,
d=0.544, but no significant differences in children, t
(39) =0.86. Here we found significantly higher perfor-
mance in adults than in children in both the OCL
condition, t(78) =5.58, p<.0001, d=1.264 (equal
variance not assumed) and in the FBCL condition, t
(78) =3.45, p<.002, d=0.781 (equal variance not
assumed).
Figure 3 Analysis of variance. In all panel error bars are the standard errors of the mean. (a) Adult group learning trajectories (based
on mean performance in each test block). (b) Children group learning trajectories. (c) Mean performance in the early learning phase
(T2 to T4). (d) Mean performance in the late learning phase (T7 to T9). See Appendix S4 in the Supporting Information for bootstrap
test and a confidence interval report.
COLOR
©2015 John Wiley & Sons Ltd
8 Rubi Hammer et al.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
In some of the above analyses we found large differences
in performance distributions (violations of sphericity or
unequal variances) in the two CL conditions and the two
age groups. Consequently, we conducted non-parametric
post-hoc tests, complementing the above ANOVAs. Find-
ings from these non-parametric post-hoc analyses (Sup-
porting Information, Appendix S5) are consistent with
the post-hoc t-tests reported above.
There were no significant Pearson correlations between
performance in the OCL and FBCL tasks in either the
early (T2 to T4) or late (T7 to T9) learning phases, in either
age group. All correlation coefficients ranged between
0.03 <r(40) <0.33, with all p>.15 (Bonferroni corrected
for four tests; note that if not correcting for multiple tests,
in children late-phase performance in the two experimen-
tal conditions were significantly correlated, r(40) =0.32,
p<.05). This indicates that being skilled in one CL
condition did not guarantee being skilled in the other.
Post-CL categorization and label recollection tasks
Thirty-eight children and 38 adults completed the post-
CL categorization and label recollection tasks. In the
post-CL categorization tasks we found an age group
main effect, F(1, 74) =17.74, p<.0001, g2
p=0.193 (with
better performance for adults), but no age group by
learning condition interaction, F(1, 74) =0.00, and no
condition main effect, F(1, 74) =1.80, p=.184 (Fig-
ure 4a). This pattern is similar to the one we found in the
late-phase CL tasks.
Similarly, in the labels recollection tasks there was a
significant age group main effect, F(1, 74) =23.33,
p<.0001, g2
p=0.240, but no age group by condition
interaction, F(1, 74) =0.38, and no condition main
effect, F(1, 74) =0.24. Adults’mean performance in the
labels recollection tasks was near ceiling, and children’s
performance was also very high (mean A-prime 0.9;
Figure 4b). This indicates that the processing of labels
did not constrain the observed CL performance.
The post-CL categorization tasks required knowing
only which feature-dimensions were task relevant, but on
the other hand, the CL tasks required learning the task-
relevant features and properly associating each feature to
a corresponding syllable. We expected these two capac-
ities to be highly correlated. However, we found that
performance in the CL tasks and post-CL categorization
tasks were not always significantly correlated. In adults
we found significant Pearson correlations between the
late CL performance and post-CL categorization per-
formance in both the OCL, r(38) =0.67, p<.001, and
FBCL, r(38) =0.48, p<.008, conditions (Figure 4c).
Besides both correlations being significant, the strength
of the two correlations did not significantly differ in
adults, z(38) =1.20, p>.20 (two-tailed test). In con-
trast, in children we found a significant correlation in the
FBCL condition, r(38) =0.68, p<.001, but not in the
OCL condition, r(38) =0.33, p>.15 (Bonferroni cor-
rected for four tests; uncorrected, p<.05). In children,
the two correlations significantly differed, z(38) =2.05,
p<.05 (two-tailed test; Figure 4d).
In children, we also found a significant correlation
between the early-phase CL performance and the post-
CL categorization performance in the FBCL condition, r
(38) =0.46, p<.015, but not in the OCL condition, r
(38) =0.07, p>.90 (Bonferroni corrected for four tests).
Here there was a trend toward a significant difference
between the two correlations, z(38) =1.82, p=.069
(two-tailed test). There were no significant correlations
between the early-phase CL performance and post-CL
categorization performance in adults, with r(38) =0.18,
p>.50 in the OCL condition, and r(38) =0.28, p>.15
in the FBCL condition. This indicates consistency in CL
performance in children, but only in the FBCL condition
(little difference between the early-phase and late-phase
performance, where both were significantly correlated
with the post-CL categorization performance). Children’s
Figure 4 Performance in the post-CL categorization tasks and
label recollection tests. (a) Mean performance (error bars are
standard errors of the means) in the post-CL categorization
tasks (same-category/different-categories decision, without
labels). (b) Performance in the label recollection tasks (correct/
altered label decision). (c) Correlation between late-phase CL
performance and post-CL categorization performance in
adults. (d) Correlation between late-phase CL performance and
post-CL categorization performance in children.
COLOR
©2015 John Wiley & Sons Ltd
Development of autonomous category learning skills 9
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
OCL performance pattern indicates higher likelihood for
strategy switching.
Discussion
This study is the first to show developmental differences,
taking place between childhood and adulthood, in
autonomous OCL and interactive FBCL. Our findings
show that children perform best in scenarios where they
are required to actively make predictions, and infer a
categorization rule from corrective feedback that follows
their actions. In contrast, adults perform somewhat
better in passive observational learning scenarios where
useful information (e.g. an indication of the correctness
of an object–label association) becomes available regard-
less of their actions. Specifically, we found that adults’
performance exceeded that of children primarily in OCL
tasks, whereas the difference between the two age groups
was smaller in FBCL tasks. This was mostly evident at
early-phase CL, where we found a significant difference
between the two age groups in OCL, but not in FBCL
(Figure 3a and 3b). We also found that children’s early-
phase FBCL was better than their OCL, whereas in
adults we found a reverse trend (Figure 3c). In late-phase
learning, children’s performance in OCL matched their
performance in FBCL (Figure 3d). Moreover, evidence
(see Appendix S3B in the Supporting Information)
suggests that children could have done even better (and
more similar to adults) in FBCL if they had exploited the
information in the feedback-based learning trials by
avoiding missed responses (timeouts).
Finding that adults’OCL may be better than their
FBCL suggests that OCL is frequently used in adulthood
(at least when testing university students), resulting in
improved OCL skills. Alternatively, this may be the
outcome of unique characteristics of the current exper-
imental design, or it may reflect more generic develop-
mental changes in executive control that are not
necessarily associated with academic competence. We
discuss these alternatives in the following paragraphs.
Better performance in OCL in adults contradicts
earlier findings showing that OCL in healthy adults at
best equals their FBCL (Ashby et al., 2002; Edmunds
et al., 2015; Ramscar et al., 2010). This may stem from
unique characteristics of the current experimental
design. In the OCL tasks in the current study, partici-
pants were introduced with exemplars associated with
correct category labels (as is the case in other studies) but
also with incorrect creature–label associations. This
enabled participants to integrate insights from trials that
required complementary inferences –for example, par-
ticipants were directly informed about what makes a
creature a POKMAG, and also directly about what is
sufficient for a creature not to be a POKMAG, in both
CL conditions. In standard OCL procedures, partici-
pants are introduced only to correct stimulus–labels
associations, and thus learning is errorless (Middleton &
Schwartz, 2012; Warmington & Hitch, 2014; Warming-
ton, Hitch & Gathercole, 2013). On the other hand, in
standard FBCL tasks learning is error driven, where
participants are presented with both correct and incor-
rect stimulus–label associations, and by using trial and
error the participant identifies false associations (Kluger
& DeNisi, 1996; Levering & Kurtz, 2014; Ramscar, Dye
& McCauley, 2013). Admittedly, the unique composition
of the OCL trials in the current study may compromise
the comparison between the current findings and earlier
ones. However, we argue that our design offers a more
straightforward comparison between the two experimen-
tal conditions by keeping the stimuli statistics in both
conditions identical, and by enabling participants to use
similar inferences. Performance differences between the
two conditions in the current study cannot be attributed
to differences in the objective information that was made
available to the participants, and instead are likely to
reflect differences in how effectively information was
extracted and used.
The above implies that the current design is more
likely to produce similar performance in the two exper-
imental conditions, and it does not explain the better
performance of adults in OCL tasks. A possible expla-
nation for adults’better OCL performance is that in the
observational learning trials, the cue indicating the
correctness of the creature–label association was pre-
sented shortly after the creature–label presentation, and
prior to the participant being required to respond. This
may have enabled adults to produce an implicit hypoth-
esis regarding the correctness of creature–label associa-
tions (before the creature and label stimuli were
presented), and then to use the cue to confirm the
correctness of their implicit hypothesis. That is, in the
OCL tasks adults may have used the cue as ‘corrective
feedback’for an implicit mental categorization decision
they made during the stimuli presentation. If this was the
case, in the OCL tasks there was a shorter time interval
between the production of a mental categorization
decision and the following ‘feedback’(cue), as compared
with the time interval between the stimuli presentation
and feedback in the FBCL tasks. Some suggest that such
delaying of feedback in rule-based CL tasks is likely to
compromise learning (Ell, Ing & Maddox, 2009). How-
ever, note that here differences in time lag between the
stimuli presentation and the cue (OCL) versus the
feedback (FBCL) were smaller (1 second) than dif-
ferences in time lag tested by Ell and colleagues
©2015 John Wiley & Sons Ltd
10 Rubi Hammer et al.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
(4.5 seconds). Moreover, evidence suggests that in some
tasks delayed feedback forces prolonged circulation of
information in working memory, which may ultimately
result in better retention and generalization (Smith &
Kimball, 2010; Roediger & Butler, 2011).
Changing the composition of the OCL trials, so that
the cue indicating the correctness of the creature–label
association was presented before, or together with, the
creature–label stimuli could prevent participants from
producing an implicit hypothesis in OCL. Earlier studies
provide mixed evidence regarding how such modifica-
tions may alter CL in adults (Ashby et al., 2002;
Levering & Kurtz, 2014), and future studies will have
to account for how these may affect CL in children.
Conversely, such modifications may have even greater
disadvantages (e.g. the simultaneous presentation of a
creature, a label and an association-correctness cue
would increase processing load during stimuli presenta-
tion in OCL tasks).
The evidently better performance in FBCL in children
is consistent with our hypothesis. We expected children
to be more skilled in interactive FBCL because this form
of learning resembles learning scenarios characteristic of
early development, where the child actively explores her
environment by interacting with objects and people. In
contrast, in OCL the child’s actions have no impact on
the information that is made available to her. Here, in
both experimental conditions the amount of social
interaction was identical, and in both conditions the
child had to complete the 13 minute long demanding
tasks by making use of the same quality and quantity of
information. In the OCL tasks, a smiling face always
indicated that the creature–label association was correct,
and a sad face always indicated an incorrect association.
In the FBCL tasks, a smiling face always indicated that
the participant categorization decision was correct, and a
sad face always indicated an incorrect decision. Thus, it
is unlikely that the better FBCL performance of children
resulted from a disposition to rely on a credible source
(Brosseau-Liard & Poulin-Dubois, 2014; Poulin-Dubois,
Brooker & Polonia, 2011; Tummeltshammer, Wu, Sobel
& Kirkham, 2014). However, we cannot completely
exclude the possibility that children interpreted the OCL
setting as a less reliable setting where they are first told
one thing about a creature category (when presented
with a creature-label pair), and then half of the time they
are later informed that what they were told earlier was
false. In the FBCL tasks it was perhaps natural for
children to accept that they might initially generate false
associations.
As discussed above, CL could have been accomplished
by complementary inferences (what is essential for a
creature to be a member of a given category, and what is
sufficient for excluding it from a category) in both
conditions. It is possible that children do not use these
two types of inference as effectively as adults. For
example, 6- to 10-year-old children were found to be as
capable as adults in inferring a categorization rule by
comparing objects from the same category, but they
failed in making inferences by contrasting examples from
different categories (Hammer et al., 2009a). The current
design does not enable us to assess the differential utility
of each kind of inference (intermixed trials with correct
vs. incorrect creature–label pairing). We suggest that it is
unlikely that this is a primary explanation for perfor-
mance differences between the two CL conditions in
children because here the two kinds of inferences were
similarly usable in both conditions. Moreover, we found
no differences between children and adults in their
respective tendency to make Misses versus False-Alarms
in the learning trials, in either condition (see
Appendix S3A in the Supporting Information). How-
ever, we cannot exclude the possibility that children had
greater difficulty in processing incorrect examples in
OCL. This is analogous to a scenario where a child is
presented with a dog and told, ‘This is not a cat’. While
contrasting examples can be most informative when
learning to differentiate between similar categories
(Hammer et al., 2008), children may find such observa-
tions confusing.
In most natural feedback-based learning scenarios,
receiving constructive information is conditioned on
taking an action. A correct action would most often
result in a rewarding outcome (e.g. smiling face), and an
incorrect action would result in an aversive or less-
rewarding outcome (e.g. sad face). However, some
studies involve FBCL tasks where the participants are
provided with a correct category label following either a
correct or incorrect categorization decision, or even
when the participant was not producing an on-time
decision. That is, the same information became available
to participants regardless of their actions (see Maddox
et al., 2004, for a related discussion). Arguably, always
providing the correct category label as feedback makes
an FBCL task somewhat more similar to an OCL
scenario, where retrieving constructive information is not
conditioned by the learner’s actions. This resembles
scenarios such as a mother telling her child, ‘That was a
dog’whenever passing by a dog, and ‘That was a
cat’whenever passing by a cat. Our findings raise
questions regarding how well children learn in such
scenarios, even when the timing between the processing
of the target object and the presentation of a category
label may enable the child to generate an implicit
prediction. Future studies should systematically
account for such apparently minor differences in the
©2015 John Wiley & Sons Ltd
Development of autonomous category learning skills 11
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
operational definitions of OCL and FBCL, and their
implementation in experimental designs.
In the current study, we employed several measures for
keeping participants engaged with the tasks. This
included presenting the overall performance score at
the end of each test block, making the two experimental
conditions more similar to one another as compared
with what would be expected in more ‘natural’settings
(e.g. OCL where a child reads a book without any
intervention). We suggest that children’s performance in
OCL tasks would have been even lower if the partici-
pants had not been informed about their overall scores at
the end of the test blocks (making the OCL conditions
totally non-interactive). Such a scenario would have left
participants with no external incentive and no external
indicator for the correctness of their performance in the
OCL tasks. This is expected to impact mostly individuals
with less-developed executive and self-regulating skills,
and thus would mostly impact children (Graham, 1997;
Logan & Gordon, 2001). On the other hand, in the
feedback-based tasks, the trial-by-trial feedback served
both as a motivator and as an external confirmation for
the correctness of each categorization decision, as in
classic trial-and-error-based learning scenarios.
Adults’performance in the post-CL categorization
tasks, which did not require label recollection (same-
category/different-categories decisions), were highly cor-
related with their late-phase CL performance in both
conditions (Figure 4c). On the other hand, among
children who performed poorly in the late-phase of the
OCL tasks, it was not uncommon to find cases with high
performance in the post-CL categorization tasks (Fig-
ure 4d). That is, for children a more likely ‘cognitive
bottleneck’in OCL tasks was learning the specific feature–
syllable associations, rather than inferring the task-
relevant feature-dimensions. In contrast, in the FBCL
tasks children were more likely to directly learn the
feature–syllable associations. For example (referring to
Figure 1c), participants could start by learning that
‘POK’=large head, ‘LUX’=small head, ‘MAG’=
two-toe feet, and ‘YED’=flat feet. This form of knowl-
edge is essential for perfecting performance in the CL
tasks. In the post-CL categorization tasks, this knowledge
could ultimately be translated into a correct same/different
decision (e.g. making same-category decisions only when
paired creatures were identical in both heads and feet).
In OCL tasks many children learned the task-relevant
feature-dimensions better than they learned the syllable–
feature associations. This may have resulted from chil-
dren using the whole label as a proxy for associating or
dissociating between creatures across trials, using tran-
sitivity, which imposes equivalence constraints between
distinct creature stimuli (Hammer, Hertz, Hochstein &
Weinshall, 2009b). For example, if Creature-A (large-
head, purple body, tail, flat feet) =‘POKYED’, and
Creature-B (large-head, orange body, wings, flat
feet) =‘POKYED’, then it can be deduced that Crea-
ture-A =Creature-B, and thus it can be concluded that
only the creatures’heads and feet are likely to be relevant
for categorization. This enables correct same/different
categorization decisions to be made, despite not knowing
the specific feature–syllable associations (e.g. the partic-
ipant may think that ‘POK’indicates either large head
or flat feet). Using such a strategy for eliminating
irrelevant feature-dimensions could enable in later stages
learning the specific feature–syllable associations while
focusing attention only on the task-relevant features.
Based on the current findings, we suggest that in early-
phase learning children faced greater difficulties in
learning the feature–syllable association in OCL tasks.
This forced children to switch to another, possibly more
explicit, learning strategy that involved inferring first the
task-relevant feature-dimensions, and only later learning
the specific feature–syllable associations. This resulted in
lower early-phase performance, but it enabled reaching
in the late-phase of the OCL tasks performance similar
to that in the late-phase FBCL tasks. Nevertheless, even
the late-phase OCL performance was often lower than
the corresponding post-CL categorization performance.
This indicates that learning the feature–syllable associ-
ation was consistently falling behind what the child
learned about the task-relevant feature-dimensions. On
the other hand, in FBCL there was little difference
between the early-phase and late-phase performance,
where both were significantly correlated with the post-
CL categorization performance. This indicates a stronger
relation between the learning of feature–syllable associ-
ations and the learning of the task-relevant feature-
dimensions in FBCL tasks.
The above ideas resonate with earlier findings showing
that classification tasks and inference tasks may have a
differential impact on later categorization performance
and item/exemplar recollection. Classification tasks are
conceptually similar to learning creature–label associa-
tions, whereas inference tasks are conceptually similar to
inferring the task-relevant feature-dimensions using
equivalence constraints (Deng & Sloutsky, 2015; Mark-
man & Ross, 2003; Rips, Smith & Medin, 2012). The
composition of learning trials in the current study, in
both conditions, permitted participants either to learn
directly the creature–label associations (by learning
feature–syllable associations) or to use labels for infer-
ring the task-relevant feature-dimensions by associating/
dissociating between creatures across trials (using tran-
sitivity). Our data suggest that in FBCL tasks children
are more likely to learn creature–label associations,
©2015 John Wiley & Sons Ltd
12 Rubi Hammer et al.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
whereas in OCL tasks they are more likely to start by
inferring the categorization rule. The transferring of
knowledge from one task to another may depend on the
availability of executive resources and cognitive control
(see Miles, Matsuki & Minda, 2014; Rabi, Miles &
Minda, 2015, for related findings).
Our data do not enable us to make similar inferences
regarding adults’default learning strategy in either CL
condition. The significant correlations between the late-
phase CL tasks and post-CL categorization tasks, in both
conditions, may indicate that adults learned the creature–
label associations, as we interpreted children’s perfor-
mance in the FBCL task. However, the high late-phase
CL performance in adults does not enable us to exclude
the possibility that adults first inferred the task-relevant
feature-dimensions (see Chin-Parker & Ross, 2002;
Hoffman & Rehder, 2010; Johansen & Palmeri, 2002;
Palmeri & Flanery, 1999; Sakamoto & Love, 2006;
Yamauchi & Markman, 2000, for studies showing that
adults may use such a strategy). Unlike children, adults
may have inferred the relevant feature-dimensions well
before the end of the CL tasks, leaving them with
sufficient opportunity to perfect their performance by
eventually learning the feature–syllable associations.
Future studies, in which performance in same/different
categorization tasks (without labels) is tested in interme-
diate CL phases, may enable us to better determine
whether adults are likely to learn the relevant feature-
dimensions prior to learning exemplar–label associations.
Our findings suggest that elementary school children
are more skilled in interactive FBCL than in passive
OCL, and this difference in CL skills is eliminated, and
possibly even reversed, by adulthood. We found that
FBCL was likely to result in children effectively associ-
ating between exemplars and category labels, whereas
OCL was more likely to result in children inferring which
feature-dimensions were task relevant. Our experimental
design had some unique characteristics: we kept the
stimuli presentation phase identical in the two CL
conditions, enabling participants to allocate similar
cognitive resources when processing the stimuli in the
OCL and FBCL tasks; moreover, we presented partic-
ipants with both correct and incorrect creature–label
associations in the OCL tasks. Although such a design
makes it harder to compare the current findings with
some published studies, it enables a better comparison
between OCL and FBCL, in children and adults.
Acknowledgements
This study was funded by the Northwestern University
Human Cognition T32-NS047987 NIH training grant to
R. Hammer. We thank Jacqueline Hirschey, Julia Del-
guidice and Palak Patel for their help in data collection.
References
Alvarez, A.L., & Booth, A.E. (2014). Motivated by meaning:
testing the effect of knowledge-infused rewards on preschool-
ers’persistence. Child Development,85 (2), 783–791.
Andre, T., & Thieman, A. (1988). Level of adjunct question,
type of feedback, and learning concepts by reading. Con-
temporary Educational Psychology,13 (3), 296–307.
Ashby, F.G., Maddox, W.T., & Bohil, C.J. (2002). Observa-
tional versus feedback training in rule-based and informa-
tion-integration category learning. Memory and Cognition,
30 (5), 666–677.
Bechtel, S., Jeschonek, S., & Pauen, S. (2013). How 24-month-
olds form and transfer knowledge about tools: the role of
perceptual, functional, causal, and feedback information.
Journal of Experimental Child Psychology,115 (1), 163–179.
Best, C.A., Yim, H., & Sloutsky, V.M. (2013). The cost of
selective attention in category learning: developmental
differences between adults and infants. Journal of Experi-
mental Child Psychology,116 (2), 105–119.
Bohlmann, N.L., & Fenson, L. (2005). The effects of feedback
on perseverative errors in preschool aged children. Journal of
Cognition and Development,6(1), 119–131.
Booth, A.E., & Alvarez, A. (2015). Developmental changes in
causal supports for word learning. Language Learning and
Development,11 (1), 80–92.
Brosseau-Liard, P.E., & Poulin-Dubois, D. (2014). Sensitivity
to confidence cues increases during the second year of life.
Infancy,19 (5), 461–475.
Butler, A.C., Godbole, N., & Marsh, E.J. (2013). Explanation
feedback is better than correct answer feedback for promot-
ing transfer of learning. Journal of Educational Psychology,
105 (2), 290–298.
Carvalho, P.F., & Goldstone, R.L. (2014). The benefits of
interleaved and blocked study: different tasks benefit from
different schedules of study. Psychonomic Bulletin & Review,
22 (1), 281–288.
Charman, T., Baron-Cohen, S., Swettenham, J., Baird, G., Cox,
A. et al. (2000). Testing joint attention, imitation, and play as
infancy precursors to language and theory of mind. Cognitive
Development,15 (4), 481–498.
Chin-Parker, S., & Ross, B.H. (2002). The effect of category
learning on sensitivity to within-category correlations. Me-
mory and Cognition,3, 353–362.
Cohen, L.B., Rundell, L.J., Spellman, B.A., & Cashon, C.H.
(1999). Infants’perception of causal chains. Psychological
Science,10 (5), 412–418.
Cohen, Y., & Schneidman, E. (2013). High-order feature-based
mixture models of classification learning predict individual
learning curves and enable personalized teaching. Proceedings
of the National Academy of Sciences, USA,110 (2), 684–689.
Deng, W.S., & Sloutsky, V.M. (2015). The development of
categorization: effects of classification and inference training
©2015 John Wiley & Sons Ltd
Development of autonomous category learning skills 13
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
on category representation. Developmental Psychology,51
(3), 392–405.
Diesendruck, G., Hammer, R., & Catz, O. (2003). Mapping the
similarity space of children and adults’artifact categories.
Cognitive Development,18 (2), 217–231.
Edmunds, C.E.R., Milton, F., & Wills, A.J. (2015). Feedback
can be superior to observational training for both rule-
based and information-integration category structures.
Quarterly Journal of Experimental Psychology,68 (6),
1203–1222.
Ell, S.W., Ing, A.D., & Maddox, W.T. (2009). Criterial noise
effects on rule-based category learning: the impact of delayed
feedback. Attention, Perception, & Psychophysics,71 (6),
1263–1275.
Feldman, N.H., Myers, E.B., White, K.S., Griffiths, T.L., &
Morgan, J.L. (2013). Word-level information influences
phonetic learning in adults and infants. Cognition,127 (3),
427–438.
Forbus, K.D., Riesbeck, C., Birnbaum, L., Livingston, K., &
Sharma, A. et al. (2007, July). Integrating natural language,
knowledge representation and reasoning, and analogical
processing to learn by reading. In Proceedings of the
National Conference on Artificial Inelegance (22, 2, 1542).
Menlo Park, CA; Cambridge, MA; London: AAAI Press;
MIT Press.
Friston, K.J., Holmes, A.P., & Worsley, K.J. (1999). How many
subjects constitute a study? NeuroImage,10 (1), 1–5.
Graham, S. (1997). Executive control in the revising of students
with learning and writing difficulties. Journal of Educational
Psychology,89 (2), 223–234.
Graham, S.A., & Diesendruck, G. (2010). Fifteen-month-old
infants attend to shape over other perceptual properties in an
induction task. Cognitive Development,25 (2), 111–123.
Grier, J.B. (1971). Nonparametric indexes for sensitivity and
bias: computing formulas. Psychological Bulletin,75 (6), 424–
429.
Griffiths, T.L., & Tenenbaum, J.B. (2009). Theory-based causal
induction. Psychological Review,116 (4), 661–716.
Hammer, R. (2015). Impact of feature saliency on visual
category learning. Frontiers in Psychology,6, 451.
Hammer, R., Bar-Hillel, A., Hertz, T., Weinshall, D., &
Hochstein, S. (2008). Comparison processes in category
learning: from theory to behavior. Brain Research,1225 (15),
102–118.
Hammer, R., & Diesendruck, G. (2005). The role of dimen-
sional distinctiveness in children’s and adults’artifact
categorization. Psychological Science,16 (2), 137–144.
Hammer, R., Diesendruck, G., Weinshall, D., & Hochstein,
S. (2009a). The development of category learning strate-
gies: what makes the difference? Cognition,112 (1), 105–
119.
Hammer, R., Hertz, T., Hochstein, S., & Weinshall, D. (2009b).
Category learning from equivalence constraints. Cognitive
Processing,10 (3), 211–232.
Hammer, R., Sloutsky, V., & Grill-Spector, K. (2012). The
interplay between feature-saliency and feedback information
in visual category learning tasks. In Annual Conference of the
Cognitive Science Society (Vol. 2012, pp. 420–425). NIH
Public Access.
Hammer, R., Sloutsky, V., & Grill-Spector, K. (2015). Feature
saliency and feedback information interactively impact visual
category learning. Frontiers in Psychology,6, 74.
Hanania, R., & Smith, L.B. (2010). Selective attention and
attention switching: towards a unified developmental
approach. Developmental Science,13 (4), 622–635.
Hoffman, A.B., & Rehder, B. (2010). The costs of supervised
classification: the effect of learning task on conceptual
flexibility. Journal of Experimental Psychology: General,
139, 319–340.
Huang-Pollock, C.L., Maddox, W.T., & Karalunas, S.L. (2011).
Development of implicit and explicit category learning.
Journal of Experimental Child Psychology,109 (3), 321–335.
Huang-Pollock, C.L., Maddox, W.T., & Tam, H. (2014). Rule-
based and information-integration perceptual category learn-
ing in children with attention-deficit/hyperactivity disorder.
Neuropsychology,28 (4), 594–604.
Johansen, M.K., & Palmeri, T.J. (2002). Are there representa-
tional shifts during category learning? Cognitive Psychology,
45 (4), 482–553.
Kluger, A.N., & DeNisi, A. (1996). The effects of feedback
interventions on performance: a historical review, a meta-
analysis, and a preliminary feedback intervention theory.
Psychological Bulletin,119 (2), 254–284.
Korman, J., Voiklis, J., & Malle, B.F. (2015). The social life of
cognition. Cognition,135,30–35.
Levering, K.R., & Kurtz, K.J. (2014). Observation versus
classification in supervised category learning. Memory and
Cognition,43, 266–282.
Logan, G.D., & Gordon, R.D. (2001). Executive control of
visual attention in dual-task situations. Psychological Review,
108 (2), 393–434.
McColeman, C.M., Barnes, J.I., Chen, L., Meier, K.M.,
Walshe, R.C. et al. (2014). Learning-induced changes in
attentional allocation during categorization: a sizable catalog
of attention change as measured by eye movements. PloS
ONE,9(1), e83302.
McColeman, C.M., & Blair, M.R. (2013). The relationship
between saccade velocity, fixation duration, and salience in
category learning. Visual Cognition,21 (6), 701–703.
Maddox, W.T., Bohil, C.J., & Ing, A.D. (2004). Evidence for a
procedural-learning-based system in perceptual category
learning. Psychonomic Bulletin & Review,11 (5), 945–952.
Markman, A.B., & Ross, B.H. (2003). Category use and
category learning. Psychological Bulletin,129 (4), 592–613.
Meltzoff, A.N. (1988a). Infant imitation after a 1-week delay:
long-term memory for novel acts and multiple stimuli.
Developmental Psychology,24 (4), 470–476.
Meltzoff, A.N. (1988b). Infant imitation and memory: nine-
month-olds in immediate and deferred tests. Child Develop-
ment,59 (1), 217–225.
Meyer, M., Bekkering, H., Janssen, D.J., de Bruijn, E.R., &
Hunnius, S. (2014). Neural correlates of feedback processing
in toddlers. Journal of Cognitive Neuroscience,26 (7), 1519–
1527.
©2015 John Wiley & Sons Ltd
14 Rubi Hammer et al.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
Middleton, E.L., & Schwartz, M.F. (2012). Errorless learning
in cognitive rehabilitation: a critical review. Neuropsycholog-
ical Rehabilitation,22 (2), 138–168.
Miles, S.J., Matsuki, K., & Minda, J.P. (2014). Continuous
executive function disruption interferes with application of
an information integration categorization strategy. Attention,
Perception, & Psychophysics,76 (5), 1318–1334.
Needham, A., Goldstone, R.L., & Wiesen, S.E. (2014).
Learning visual units after brief experience in 10-month-
old infants. Cognitive Science,38 (7), 1508–1519.
Palmeri, T.J., & Flanery, M.A. (1999). Learning about
categories in the absence of training: profound amnesia
and the relationship between perceptual categorization and
recognition memory. Psychological Science,10 (6), 526–
530.
Poulin-Dubois, D., Brooker, I., & Polonia, A. (2011). Infants
prefer to imitate a reliable person. Infant Behavior and
Development,34 (2), 303–309.
Rabi, R., Miles, S.J., & Minda, J.P. (2015). Learning
categories via rules and similarity: comparing adults and
children. Journal of Experimental Child Psychology,131,
149–169.
Rabi, R., & Minda, J.P. (2014). Rule-based category learning in
children: the role of age and executive functioning. PloS
ONE,9(1), e85316.
Ramscar, M., Dye, M., & McCauley, S.M. (2013). Error and
expectation in language learning: the curious absence of
mouses in adult speech. Language,89 (4), 760–793.
Ramscar, M., Yarlett, D., Dye, M., Denny, K., & Thorpe, K.
(2010). The effects of feature-label-order and their implica-
tions for symbolic learning. Cognitive Science,34 (6), 909–
957.
Reber, P.J., & Squire, L.R. (1999). Intact learning of artificial
grammars and intact category learning by patients with
Parkinson’s disease. Behavioral Neuroscience,113 (2), 235–
242.
Rips, L.J., Smith, E.E., & Medin, D.L. (2012). Concepts and
categories: memory, meaning, and metaphysics. In K.J.
Holyoak & R.G. Morrison (Eds.), The Oxford Handbook of
Thinking and Reasoning (pp. 177–209). New York: Oxford
University Press.
Robinson, C.W., & Best, C.A. (2012). The role of words in
cognitive tasks: what, when, and how? Frontiers in Psychol-
ogy,3, 95.
Roediger, H.L., & Butler, A.C. (2011). The critical role of
retrieval practice in long-term retention. Trends in Cognitive
Sciences,15 (1), 20–27.
Sakamoto, Y., & Love, B.C. (2006). Vancouver, Toronto,
Montreal, Austin: Enhanced oddball memory through dif-
ferentiation, not isolation. Psychonomic Bulletin & Review,13
(3), 474–479.
Shohamy, D., Myers, C.E., Grossman, S., Sage, J., Gluck, M.A.
et al. (2004). Cortico-striatal contributions to feedback-based
learning: converging data from neuroimaging and neuropsy-
chology. Brain,127 (4), 851–859.
Sloutsky, V.M. (2010). From perceptual categories to concepts:
what develops? Cognitive Science,34 (7), 1244–1286.
Smith, L.B. (2013). It’s all connected: pathways in visual object
recognition and early noun learning. American Psychologist,
68 (8), 618–629.
Smith, T.A., & Kimball, D.R. (2010). Learning from feedback:
spacing and the delay-retention effect. Journal of Experi-
mental Psychology: Learning, Memory, and Cognition,36 (1),
80–95.
Sommerville, J.A., Woodward, A.L., & Needham, A. (2005).
Action experience alters 3-month-old infants’perception of
others’actions. Cognition,96 (1), B1–B11.
Stanislaw, H., & Todorov, N. (1999). Calculation of signal
detection theory measures. Behavior Research Methods,
Instruments, & Computers,31 (1), 137–149.
Tummeltshammer, K.S., Wu, R., Sobel, D.M., & Kirkham,
N.Z. (2014). Infants track the reliability of potential infor-
mants. Psychological Science,25 (9), 1730–1738.
Walker, C.M., & Gopnik, A. (2014). Toddlers infer higher-
order relational principles in causal learning. Psychological
Science,25 (1), 161–169.
Warmington, M., & Hitch, G.J. (2014). Enhancing the learning
of new words using an errorless learning procedure: evidence
from typical adults. Memory,22 (5), 582–594.
Warmington, M., Hitch, G.J., & Gathercole, S.E. (2013).
Improving word learning in children using an errorless
technique. Journal of Experimental Child Psychology,114 (3),
456–465.
Woodcock, R.W., McGrew, K.S., & Mather, N. (2001).
Woodcock-Johnson III. Itasca, IL: Riverside Publishing.
Yamauchi, T., & Markman, A.B. (2000). Inference using
categories. Journal of Experimental Psychology: Learning,
Memory, and Cognition,26 (3), 776–795.
Received: 29 December 2014
Accepted: 27 May 2015
Supporting Information
Additional Supporting Information may be found in the online
version of this article:
Appendix S1. Description of stimuli composition in a single
task.
Appendix S2. Standardized tests.
Appendix S3A. Panel-A (adults) and Panel-B (children) show
the performances in both the learning and test blocks.
Appendix S3B. Panel-A (adults) and Panel-B (children) show
the mean number of missed responses (timeouts) in each
learning block.
Appendix S4. Simple bootstrap for the post-hoc t-tests
(complements the analysis referred by Figure 3.A, 3.B and
3.C in the manuscript).
Appendix S5A. Non-parametric analysis based on perfor-
mance frequencies (as the empirical cumulative distributions) in
adults (left panels) versus children (right panels), in the early-
phase CL (upper panels) versus late-phase CL (bottom panels).
Appendix S5B. Non-parametric analysis (Chi Square and
Kolmogorov–Smirnov tests). 2
©2015 John Wiley & Sons Ltd
Development of autonomous category learning skills 15
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
Development*of*autonomous*category*learning*skills*
1*
Appendix 1
Description of stimuli composition in a single task. In each task, stimuli were split into four
categories (subspecies of creatures) based on similarities and differences in the two task-relevant
binary feature-dimensions. Within each category, stimuli differed in the two irrelevant binary
feature-dimensions, and were rendered in two distinct points of view. Category labels were
composed of two syllables, where each syllable corresponded with one of the two relevant
feature-dimensions.
Category label
Relevant F1
Relevant F2
Irrelevant F3
Irrelevant F4
Point of view
A-A
0
0
0
0
0/1
A-A
0
0
0
1
0/1
A-A
0
0
1
0
0/1
A-A
0
0
1
1
0/1
A-B
0
1
0
0
0/1
A-B
0
1
0
1
0/1
A-B
0
1
1
0
0/1
A-B
0
1
1
1
0/1
B-A
1
0
0
0
0/1
B-A
1
0
0
1
0/1
B-A
1
0
1
0
0/1
B-A
1
0
1
1
0/1
B-B
1
1
0
0
0/1
B-B
1
1
0
1
0/1
B-B
1
1
1
0
0/1
B-B
1
1
1
1
0/1
Development*of*autonomous*category*learning*skills*
2*
Appendix 2
Standardized tests. The tests were administered in the following order: Letter-Word
Identification, which tested real-word reading; Word Attack, which tested non-word reading;
Picture Vocabulary, which tested picture naming; Picture Recognition, which tested visual
working memory; Rapid Picture Naming, which tested picture naming under time constraints;
Decision Speed, which tested matching pictures for similarities under time constraints; and
Concept Formation, which tested application of rules to visual puzzles. Adults’ and children’s
performance in the standardized tests. Standardized scores are reported, such that the mean score
for the age group is set to 100 with standard deviation set to 15. As expected, adults’ mean raw
scores (not reported here) in all tests were higher than those of children.
Adults
Minimum
Maximum
Mean
Std. Deviation
Concept formation (WJC-5)
95
129
112.89
9.80
Picture recognition (WJC-13)
94
130
106.32
7.93
Decision speed (WJC-16)
79
153
117.26
17.22
Rapid picture naming (WJC-18)
84
133
105.71
13.63
Letter word ID (WJA-1)
91
126
110.84
8.03
Word attack (WJA-13)
86
120
104.76
8.90
Picture vocabulary (WJA-14)
80
132
106.91
11.43
Children
Minimum
Maximum
Mean
Std. Deviation
Concept formation (WJC-5)
103
155
121.67
10.30
Picture recognition (WJC-13)
91
153
113.08
12.30
Decision speed (WJC-16)
80
161
109.24
17.12
Rapid picture naming (WJC-18)
70
123
99.81
10.31
Letter word ID (WJA-1)
95
153
116.97
11.22
Word attack (WJA-13)
96
132
111.16
7.78
Picture vocabulary (WJA-14)
91
174
118.00
17.64
Development*of*autonomous*category*learning*skills*
3*
Appendix 3.A
Panel-A (adults) and Panel-B (children) show the performances in both the learning and test
blocks. Note that comparing between the learning trials from the two experimental conditions is
with no theoretical significance, since in the observational learning trials, as by definition,
participants were told what the correct answer is prior to making their responses. We find
significant age-group simple main effect in both experimental conditions (Panel-C). The simple
main effect in the FBCL condition is consistent with the overall pattern evident in the test blocks
(see Results). The simple main effect in the OCL condition was less expected given that here
participants were explicitly told the correct answer in each learning trial. Indeed, children’s
performances in the observational learning trials were very high, but it seems that even when
they were told the right answer they were more likely than adults to make decision errors. This
may reflect more frequent ‘attention lapses’ and/or lower motor control in children. Patterns of
performances in the observational learning block indicate that in this condition learning was
largely ‘errorless’.
An ANOVA testing if participants had a bias in making more False-Alarms or more
Misses (comparing the False-Alarm rate with the Miss rate), in the observational learning trials,
showed no significant error-type by age-group interaction, F(1, 78) = 0.02, and no error-type
main effect, F(1, 78) = 0.06. In the feedback learning trials there was no significant error-type by
age-group interaction, F(1, 78) = 1.42, p > 0.20, but there was an error-type main effect, F(1, 78)
= 52.74, p < 0.0001, 𝜂!
! = 0.403, with greater False-Alarm rate (mistaking false creature-label
association as true association) than Miss rate (mistaking true creature-label association as false
association), in both age groups.
Development*of*autonomous*category*learning*skills*
4*
Appendix 3.B
Panel-A (adults) and Panel-B (children) show the mean number of missed responses (timeouts)
in each learning block. In this analysis we explored the possibility that in the FBCL condition
learning was compromised due to frequent missed responses in the learning trials, resulting in
participants not extracting sufficient information (see Methods). In the OCL task, on the other
hand, participants were informed about the correctness of the creature-label associations
regardless of their response timing. Here, missed responses may, or may not, indicate ‘attention
lapses’ since the participant could have failed pressing the key for different reasons (to include
lack of motivation to respond only to confirm the cued creature-label association correctness).
We found that missed responses in the FBCL condition were rare, in both age groups.
Nevertheless, children had significantly more missed responses in the FBCL learning trials
(Mean per-block = 0.58 ± SD = 0.59) than adults (0.18 ± 0.16), t(78) = 4.12, p < 0.0002 (equal
variance not assumed), d = 0.93. A similar age group effect was evident in the OCL condition,
t(78) = 4.03, p < 0.0002 (equal variance not assumed), d = 0.91, where in both age groups there
were more missed responses in the OCL condition than in the FBCL condition (both p < 0.005).
In adults, we found only moderate negative correlation between the number of missed
responses (across all learning blocks in the FBCL tasks) and the late-phase CL performances,
r(40) = -0.30, p = 0.063, or with the post CL categorization performances, r(38) = -0.36, p < 0.03
(p-values are not corrected for multiple tests). In children, these two correlations were more
significant, r(40) = -0.32, p < 0.05, and r(38) = -0.50, p < 0.002, respectively. Note that the
smaller effects in adults likely resulted from the close to zero missed responses in most adult
participants. In the OCL condition, in both age groups, the respective correlations were not
statistically significant, all p > 0.15. This confirms that in OCL, missed responses in learning
trials are not necessarily associated with ‘attention lapses’ or misuse of available information.
Development*of*autonomous*category*learning*skills*
5*
Appendix 4
Simple bootstrap for the post-hoc t-tests (complements the analysis referred by Figure 3.A, 3.B
and 3.C in the manuscript). FBCL = Feedback-based learning; OCL = Observational learning;
Ch = Children; Ad = Adults. The two rightmost columns are the lower and upper bounds of the
95% bootstrapped mean difference confidence interval (bootstrap results are based on 10000
bootstrap samples).
T2 to T9
p-value (t-test)
Mean Diff
Lower
Upper
Adults (OCL – FBCL)
0.043
0.054
0.005
0.105
Children (OCL – FBCL)
0.156
-0.045
-0.105
0.016
OCL (Ch – Ad)
0.000
-0.192
-0.258
-0.128
FBCL (Ch – Ad)
0.005
-0.094
-0.157
-0.029
T2 to T4 (early phase)
p-value (t-test)
Mean Diff
Lower
Upper
Adults (OCL – FBCL)
0.086
0.059
-0.006
0.125
Children (OCL – FBCL)
0.036
-0.075
-0.141
-0.006
OCL (Ch – Ad)
0.000
-0.154
-0.226
-0.083
FBCL (Ch – Ad)
0.578
-0.021
-0.093
0.054
Development*of*autonomous*category*learning*skills*
6*
Appendix 5.A
Non-parametric analysis based on performance frequencies (as the empirical cumulative
distributions) in adults (left panels) versus children (right panels), in the early-phase CL (upper
panels) versus late-phase CL (bottom panels). Frequencies/counts are based on the number of
tasks within each observed categorization accuracy range (two tasks in each experimental
condition for each participant; total of 80 tasks). See Appendix 5.B for the complementary
statistical analysis.
Development*of*autonomous*category*learning*skills*
7*
Appendix 5.B
Non-parametric analysis (Chi Square and Kolmogorov–Smirnov tests). FBCL = Feedback-based
learning; OCL = Observational learning; Ch = Children; Ad = Adults. These analyses correspond
with the simple main effects reported in Fig 3.C and Fig 3.D. p-values are not corrected for
multiple tests.
Test blocks performance (total # of tasks)
Χ2-value
p-value (Χ2)
p-value (K-S)
Early [T2 to T4]
Adults (FBCL vs. OCL)
7.61
= 0.055
> 0.40
Children (FBCL vs. OCL)
23.89
< 0.00005
= 0.013
OCL (Ch vs. Ad)
313.63
< 0.00001
< 0.0001
FBCL (Ch vs. Ad)
4.35
= 0.087
> 0.40
Late [T7 to T9]
Adults (FBCL vs. OCL)
5.60
= 0.133
= 0.082
Children (FBCL vs. OCL)
3.78
> 0.20
> 0.20
OCL (Ch vs. Ad)
142.69
< 0.00001
< 0.0001
FBCL (Ch vs. Ad)
42.01
< 0.00001
< 0.0001
*
*
