RUNNING HEAD: CATEGORY KNOWLEDGE INTERACTION 1

PROCEDURAL AND DECLARATIVE KNOWLEDGE SIMULTANEOUSLY CONTRIBUTE

TO CATEGORY RESPONSE SELECTION

Authors [redacted for blind review]

INTERACTION IN LEARNING 2

Abstract

Skilled behaviour in real-world contexts often relies on a combination of both declarative

and procedural learning. However, precisely how declarative and procedural knowledge interact

is not yet fully understood. Previous findings have shown that procedural and declarative

learning may interact or compete at the systems level during encoding, consolidation, and

retrieval, but beyond this, it is not known whether declarative and procedural representations

themselves interact. The goal of the current study is to investigate whether procedural and

declarative knowledge can contribute simultaneously to categorization response selection

behavior. We designed a stimulus set in which information learned by each system sometimes

supports different responses, and created trials in the test phase that are designed to maximize

such divergence. Participants were instructed to use a completely diagnostic, verbalizable, shape-

based rule to categorize exemplars, receiving feedback after each trial. However, unbeknownst to

participants, the categories also differed probabilistically in their color distributions. Participants

used both color (learned procedurally) and shape (learned declaratively) to categorize exemplars,

responding more quickly when both sources indicated the same category judgement, and more

slowly when they conflicted. Debriefing confirmed participants were unaware of the color

distributions. These results showing simultaneous trial-level contributions from both declarative

and procedural memory systems. Our findings represent a novel form of interaction between the

two systems and have implications for domains beyond the laboratory, such as decision-making

and classroom instruction.

.
INTERACTION IN LEARNING 3

The last century has seen great strides in our understanding of human learning and

memory. Compelling evidence indicates that there are multiple dissociable memory systems with

different characteristics and instantiated in different neural substrates (Gabrieli, 1998). One of

the best-characterized systems is referred to as the declarative memory system, which requires

intact medial temporal lobe structures and can yield verbalizable knowledge, which is often

acquired within a single trial (Eichenbaum & Cohen, 2001; Graf & Schacter 1985; Cohen &

Squire, 1980). Within the multiple forms of memory that are not available to awareness and do

not require the medial temporal lobe (Reber, 2013), another well-characterized system is the

procedural memory system, which relies on a fronto-striatal network and is characterized by

gradual learning across multiple learning episodes, yielding non-verbalizable knowledge that is

more easily expressed through performance (Squire & Zola-Morgan 1988; Squire 1992, 2004).

The last few decades have moved beyond establishing the dissociable systems to

examining ways that they may interact, since both are available in healthy adults. Many of these

investigations have focused on competition (Poldrack & Packard, 2003), or possible

collaboration (Freedberg et al., 2020) between systems during the learning process (encoding).

What has not been examined as closely is whether and how information from both systems may

contribute to downstream processes such as decision making and response selection.

As we review below, much existing evidence suggests that information from only a

single system at a time contributes to these downstream processes. However, outside of the

laboratory, we observe scenarios in which information from both systems appears to be used

simultaneously. For example, when making diagnoses, medical experts seem to use a

combination of both conditional reasoning based on declarative knowledge as well as

probabilistic reasoning based on experience (Norman & Brooks, 1997). Similarly, professional
INTERACTION IN LEARNING 4

musicians performing from memory seem to simultaneously draw on both a non-verbalizable

representation of motor sequences as well as a verbalizable understanding of the structure, form,

and meaning of a musical piece (Chaffin, Logan, & Begosh, 2009 cited in Reber, 2013)

Procedural and Declarative Category Learning

Procedural learning is noted to be “inflexible, slow or incremental, unconscious,

automatic, and insensitive to reinforcer devaluation” (Seger & Spiering, 2011). Acquisition of

this type of memory has been referred to as procedural learning, habit learning, or skill learning

(Foerde, 2018; Foerde & Poldrack, 2008; Knowlton, Mangels & Squire, 1996; Salmon &

Butters, 1995; Seger, 2006). Importantly, procedural learning is not limited to motor skill

acquisition, but also includes cognitive skill acquisition (Poldrack et al., 1999; VanLehn, 1996;

Wan et al., 2012). Moreover, procedural learning is characterized by the acquisition of some

form of abstract information, not only simple associative information (Reber, 1989; Seger,

1994).

Category learning—learning that stimuli (such as objects) can be assigned to categories,

and that members of a category can often be treated in the same way —can be accomplished

through either declarative or procedural learning, depending on task demands and category

structure. In a typical category-learning experiment, a participant is shown an example stimulus

(e.g., a face, butterfly, or abstract shape) and asked to classify it into one of two categories, with

feedback given after each trial. Evidence of learning in these tasks is seen in increasingly

accurate responses as training progresses, and can be measured against chance responding, which

is usually 50% accuracy given the two-alternative forced-choice format.

Category learning paradigms may be deterministic or probabilistic. In deterministic

category learning, each stimulus is associated with only one category, and the given feedback is
INTERACTION IN LEARNING 5

consistent regarding category membership. In contrast, in probabilistic category learning tasks

(also known as probabilistic classification tasks, such as the “Weather Prediction Task”;

Knowlton, Squire, & Gluck, 1994), the association between each cue and the outcome (category)

is probabilistic. The consistency of the cues may vary (e.g., a cue may predict outcome A 90% of

the time, and B 10%, or the split may be 55%/45%). Thus, the same combination of cues may

sometimes lead to one outcome (e.g., Category A) and sometimes the other (e.g., Category B;

Knowlton, Mangels, & Squire, 1994). The probabilistic structure deters explicit hypothesis

testing and produces lower rates of explicit knowledge inference by participants. Despite this

difference in feedback consistency, probabilistic and deterministic category learning have similar

task demands, and individual performance differences in these two tasks are correlated,

suggesting shared processes or mechanisms (Kalra et al, 2019).

Within deterministic category learning, optimal performance could depend on either

declarative or procedural learning mechanisms, depending on the configuration of stimulus space

and placement of the category boundary. The stimulus space is often two-dimensional, such as

sine wave gratings that vary on the dimensions of frequency and orientation. For simple,

verbalizable category structures, such as those that require comparison along only one dimension

(often referred to as “Rule-Based” categories or paradigms), declarative learning is efficient for

reaching high levels of accuracy. In contrast, complex deterministic category structures that

cannot be easily verbalized, such as those requiring integrating information across two

dimensions simultaneously (“information-integration” category structures) are more effectively

learned by an implicit, feedback-driven procedural approach. Behavioral, neuropsychological,

and neuroimaging evidence suggest that rule-based category learning depends on declarative

processes, many mediated by medial temporal lobe structures (Ashby et al., 1998; Maddox,
INTERACTION IN LEARNING 6

Ashby, Ing, & Pickering, 2004; Waldron & Ashby, 2001; Zeithamova & Maddox, 2006), while

information-integration category learning depends on procedural learning mechanisms that are

mediated primarily by the striatum (Ashby & Ennis, 2006; Filoteo, Maddox, Salmon, & Song,

2005; Nomura et al., 2007).

In probabilistic category learning paradigms, cue combinations and cue-outcome

association probabilities are generally too complex to be effectively learned through declarative

methods (such as memorizing particular cue-outcome pairs). Although participants do attempt to

learn declaratively at first, and by the end of training may have gained some declarative

knowledge of cue-outcome associations (Gluck et al., 2002; Meeter et al., 2008; Shohamy et al.,

2004), optimal performance nevertheless usually depends on gradual, non-conscious, and

feedback-driven procedural learning. Like information-integration deterministic category

learning, probabilistic category learning depends on procedural learning mechanisms that are

mediated primarily by the striatum (Foerde & Shohamy, 2011b; Knowlton et al., 1994;

Knowlton, Squire, et al., 1996; Shohamy et al., 2004; Squire et al., 1994; Knowlton, Mangels, &

Squire, 1996).

Existing research on interactions between memory systems in category learning

Inspired by findings in rodent models that lesioning one system seemed to improve

function of the other system (McDonald et al., 2004; McDonald & Hong, 2013; McDonald &

White, 1995), early models of interaction between declarative and procedural memory in humans

emphasized competition for resources during encoding (learning) 1. Supporting this view,

Poldrack and colleagues (2001) examined fMRI data from participants performing either a

probabilistic classification task or a paired associate (declarative memory) task with the same

1
Competition during consolidation has been observed in sequence learning, but to our knowledge not in category
learning. (Brown et al., 2009; Brown & Robertson, 2007; Galea et al., 2010)
INTERACTION IN LEARNING 7

stimuli. They found an inverse relationship such that activity in caudate was greater in the

probabilistic classification task than the paired associate task, with medial temporal lobe activity

following the opposite pattern. Across participants, they also found an inverse relationship such

that participants who showed greater medial temporal lobe activity showed less caudate activity

(and vice versa). They interpreted this pattern of results as evidence for competition between the

systems during learning.

However, other aspects of the rodent work suggested another mode of interaction

between procedural and declarative knowledge. Rather than competing for resources during

encoding, these studies suggested that both procedural and declarative encoding could proceed

simultaneously without interfering with each other, creating sometimes redundant

representations. For example, rodents who had shifted from medial-temporal based

representations to striatal based representations to navigate a maze reinstated the previously-

learned medial temporal-based representations when their striata were inactivated with lidocaine

(Packard & McGaugh, 1996). Models of human behavior based on this work suggested that

competition between the systems might take place later, when knowledge is retrieved and

applied (Poldrack, & Rodriguez, 2004), rather than at initial encoding.

In a human behavioral study, Crossley & Ashby (2015) used a parallel approach. Instead

of lesioning one system, they selectively impaired procedural learning for some participants

using delayed feedback. Delayed feedback is known to impair procedural but not declarative

encoding (Foerde & Shohamy, 2011a; Maddox et al., 2003; Maddox & Ing, 2005; Smith et al.,

2014). All participants were trained on a complex stimulus space, but shown only stimuli for

which rule-based learning was more effective (although an information-integration encoding

strategy could also be used). After training, participants were tested on another section of the
INTERACTION IN LEARNING 8

stimulus space, for which information-integration learning was optimal. During this transfer

phase, participants who received immediate feedback during training were able to successfully

switch to using an information-integration representation, presumably acquired tacitly alongside

rule-based learning. However, participants who received delayed feedback performed poorly on

the transfer phase, indicating a failure to switch to an information-integration strategy. The

authors interpret these findings as evidence that the declarative system initially dominated

response selection in the first phase of the experiment, but that procedural learning took place

simultaneously in the participants with immediate feedback, and that the procedural system was

therefore able to “take over” response selection in those participants when it provided optimal

responses and the declarative system did not.

Human neuroimaging studies also support the possibility that procedural and declarative

encoding can take place simultaneously, without interference. Foerde et al. (2007) compared

probabilistic classification performance under both single-task (normal) and dual-task conditions.

Performing a secondary task during probabilistic classification reduced the declarative

knowledge acquired by participants (as reflected in an explicit test of cue-outcome associations),

but did not interfere with procedural learning (as reflected in classification performance).

Similarly, Miles & Minda (2011), using deterministic category structures, found that a secondary

verbal task disrupted rule-based but not information-integration category learning. Furthermore,

Foerde and colleagues (2006) found that activity in striatal regions was similar in both single

task and dual-task conditions, suggesting that procedural learning occurred regardless of

declarative learning. Classification accuracy correlated with medial temporal lobe activity in the

single-task condition and with striatal activity in the dual-task condition, further supporting the
INTERACTION IN LEARNING 9

interpretation that the medial temporal lobe and striatum support declarative and procedural

learning, respectively.

Another imaging study provided evidence that declarative and procedural learning,

beyond merely working in parallel independently of each other, could possibly work together

(Dickerson et al., 2011). In an event-related fMRI study of probabilistic classification, Dickerson

and colleagues (2011) found that both the medial temporal lobe and striatum showed a greater

BOLD response to difficult cues than easy cues. Furthermore, they observed significant

correlation in the time course of BOLD responses between the two areas, demonstrating

functional connectivity. The authors interpret these results to mean that not only could both

systems work in parallel without interfering with each other, but speculate that they may even

collaborate to enhance overall learning2.

Finally, some evidence suggests that if provided with an explicit cue, participants can

flexibly switch between rule-based and information-integration categorization strategies on

different trials within the same session (Turner et al., 2017). This finding demonstrates that

coordination between the two systems is possible and may be managed by a third, coordinating

system or mechanism3. However, trial-by-trial switching appears to be rare or difficult in the

absence of explicit cueing (Erickson, 2008; Turner et al., 2017).

To summarize, in addition to early suggestions of competition between procedural and

declarative systems for category learning during encoding, converging evidence now suggests

that (at least sometimes) both systems may simultaneously encode information in parallel during

training, without necessarily competing with one another. Additional findings suggest that there

2
Cooperation between systems in the form of increased functional connectivity has also been observed for non-
categorization tasks, including sequence learning (Albouy et al., 2013; Freedberg et al., 2020)
3
The idea of a “third party” mediating the balance of function between procedural and declarative memory has been
introduced in other discussions of interactions between procedural and declarative memory, but not in the context of
category learning specifically. (Cabeza & Moscovitch, 2013; McDonald et al., 2004; McDonald & Hong, 2013)
INTERACTION IN LEARNING 10

may be competition at a later downstream point, such as response selection, which is resolved in

a “winner-take-all” fashion, so that performance reflects only the contributions of a single

system.

This “winner-take-all” competition idea is included in the dominant computational model

of procedural and declarative category learning, COVIS (COmpetition between Verbal and

Implicit Systems). Based on neuropsychological, behavioral, and imaging findings, COVIS

proposes a 3-part apparatus: a procedural learning module, a declarative learning module, and a

gate-keeping mechanism (Ashby & Ell, 2001; Ashby & O’Brien, 2005). In COVIS, on each trial,

the procedural and declarative modules each propose a classification response (Category A or

Category B) which is fed forward to the gate-keeping mechanism along with a confidence rating.

The gate-keeping mechanism chooses the proposed response with the greater confidence rating

and feeds it forward for execution. COVIS is thus consistent with findings of simultaneous

encoding followed by competition at response-selection (Crossley & Ashby, 2015) as well as the

possibility of trial-by-trial switching (Turner et al., 2017). However, COVIS (as it is currently

specified) does not allow for the possibility that both procedurally and declaratively learned

representations could contribute to response selection simultaneously within a single trial. Yet, as

we review briefly below, there is some limited evidence to suggest that information from both

systems could contribute to downstream processes simultaneously, within a single trial.

Evidence for possible interaction/mixing

To our knowledge, no study has been specifically designed to test whether information

from both systems simultaneously contributes to a response within a single trial. However, some

studies have hinted at this possibility. For example, Brooks and Allen (1991) observed that even

though participants were given a perfectly predictive classification rule, their responses on a
INTERACTION IN LEARNING 11

transfer task were nevertheless affected by similarity to previously seen exemplars along

irrelevant dimensions—almost as if the similarity information were “contaminating” the rule-

based classification. However, this study was originally designed to contrast rule-based and

exemplar-based learning, not declarative and procedural learning, so it is possible that that the

generalization based on exemplars could have been mediated by either procedural or declarative

mechanisms. Similarly, Schoenlein and Schloss (2022) trained participants to classify stimuli

based on a completely diagnostic shape difference, but using a cool-biased color distribution for

one category and a warm-biased color difference for the other category. Participants were able to

use the shape rule effectively, and they did not demonstrate explicit knowledge of color

differences between categories. However, after training, participants were asked to rate how

associated different colors were with each category (using the category names) on a continuous

scale labeled from “not at all” to “very much.” The participants rated cool colors as more

associated with the cool-biased category and warm colors as more associated with the warm-

biased category, even generalizing to warm and cool colors that had not been included in the

training set. Again, these findings demonstrate that unconscious learning of probabilistic

information can occur simultaneously with explicit rule use. However, since the color-category

associations were assessed outside of the categorization task, this study could not test whether

this probabilistic color information directly influenced category decisions during online

classification performance.

More closely, Batterink and colleagues (Berger & Batterink, in press; Batterink et al.,

2014) found that participants who were taught an explicit rule (near/far) governing the use of

novel articles showed reaction-time and accuracy effects of the training distribution along a

covert, second dimension (animate/inanimate objects), even when they were unaware of the
INTERACTION IN LEARNING 12

article-animacy contingency. This finding shows that information acquired without awareness

can affect intentional response decisions that are made on the basis of a deterministic rule,

delaying response times when the two forms of information conflict. These results suggest that

both information that the participant is aware of, as well as information that the participant is not

aware of, are interacting at the point of response selection. In the current study, we test whether

this finding may be observed in the context of visual category learning.

The current study

The goal of the current study is to observe whether procedural and declarative knowledge

contribute simultaneously to categorization response selection behavior, after initial encoding

has occurred. It can be difficult to disentangle the contributions of multiple learning systems in a

single task because the systems typically converge on common responses. Here, we have

designed a stimulus set in which information learned by each system sometimes supports

different responses, and we have created trials in the test phase that are designed to maximize

such divergence.

For the declarative system, we provide—explicitly, verbally, and with examples—a

deterministic, verbalizable category rule based on feature combinations. For the procedural

system, our training items follow a distribution of colors that differs probabilistically between the

two categories. If an association between color distribution and category membership is learned,

it must be learned gradually, with feedback. We also include measures of whether the color-

category knowledge is available to awareness or not. If it is learned gradually, based on

feedback, and without awareness, we may infer that this learning is mediated by the striatal

procedural system. The structure and task demands of the probabilistic color-category

association are very much like the probabilistic classification task, which has been reliably
INTERACTION IN LEARNING 13

demonstrated to use the striatal procedural system (Foerde & Shohamy, 2011b; Knowlton et al.,

1994; Knowlton, Squire, et al., 1996; Shohamy et al., 2004; Squire et al., 1994; Knowlton,

Mangels, & Squire, 1996.).

During the test phase, if we see a difference between trials in which procedural and

declarative learning point to the same response (congruent) compared to those trials in which

they point to different responses (incongruent), we will conclude that both sources of information

contribute to response selection within a given trial. If we do not observe such a difference, we

will conclude that information from both systems is not used simultaneously in response

selection, at least in the current paradigm.

Experiment 1

Methods
Participants
A total of 243 participants were recruited using the Connect Cloud Research platform

(Hartman et al., 2023) following an approved ethics protocol (ages 18-68; mean age 38.31 years;

51% Male, 49% Female by self-report). Participants were compensated US$4.50 for completing

the task. To encourage engagement with the task, participants were told that bonuses were

possible for performance over 90% (the bonus given was $0.50). After exclusions, a total of 202

participants were included for analysis (see Analysis section for detail on exclusion criteria).

Materials

Alien Stimuli

“Alien” images were created using custom code and Python’s PIL package to combine

geometric figures (ovals, rectangles, etc.). Each alien consisted of a large oval with some
INTERACTION IN LEARNING 14

configuration of the following features: number of eyes (1-4), mouth type (round or angular),

nose type (round or angular), ears (present or absent). All thirty-two possible combinations of

these features were generated and used in the experiment (see Figure 1A for example images).

Categorization Features and Explicit Rule. Stimuli were divided into two categories

based on an exclusive or (XOR) rule over the eye and mouth dimensions (Figure 1). Stimuli in

Category A had either a round mouth and odd number of eyes OR a square mouth and even

number of eyes. Conversely, stimuli in Category B had either a square mouth and odd number of

eyes OR a round mouth and even number of eyes. Thus, each category included 16 possible

configurations of features (including both the diagnostic and nondiagnostic feature dimensions

[nose and ears]). Because of the disjunctive nature of the rule, each category could be further

divided into two subcategories based on which part of the XOR rule applied (e.g., for each

category, odd number of eyes and even number of eyes belonged to different subcategories).

Color selection and distribution. Each configuration of features (category token) was

then generated in a variety of colors (Figure 1). Colors were divided into warm and cool colors

based on their hue angle. Even hue-angle spacing was used to choose hue angles of 0, 30, 60, 90,

120, 150, 180, 210, 240, 270, 300, 330 (where 0 corresponds to red and 180 to cyan). Warm

colors were defined as those within 90 degrees (+/-) of 0; cool colors were defined as those

greater than 90 degrees (+/-) from 0. For each hue angle position, saturation (chroma) was

adjusted to minimize differences in saturation across hue angles and between warm and cool

color groups. Given the inherent asymmetry of the visual color space, it was not possible to

equalize luminance (brightness) between warm and cool colors; cool colors were systematically

lower in luminance than warm colors. The COLORMATH package for Python (Taylor, 2018) was

used to select colors in CIELChuv color space (CIE, 1986) and to calculate the distances between
INTERACTION IN LEARNING 15

colors in that color space (delta e). Degree of warmness and differences in warmness between

colors were calculated based on hue angle and distance from 0. LHC and RGB values can be

found in Table 1 of the Supplemental Materials.

Training phase color distributions. Unbeknownst to participants, stimuli in the training

condition followed biased color distributions. Two color distributions were created: one warm-

biased and one cool-biased, and each was assigned to a category in counterbalanced fashion (i.e.,

for half of participants, Category A followed the warm-biased distribution and Category B the

cool-biased distribution, and vice versa for the other half; see Figure 1). Each color distribution

included a total of 88 tokens distributed across all 12 hue values. Each distribution contained 64

congruent (e.g., warm colors in the warm distribution) and 24 incongruent (e.g., cool colors in

the warm distribution) items. Within each category, each color was also distributed evenly across

subcategory and number of eyes. The distribution of non-diagnostic features (nose type and ears/

no ears) was also matched between the two categories.

Procedure
Overview
Participants completed the task on their personal computers (option to complete the task

on tablet or phone was disabled). After giving informed consent, participants viewed an

explanation of the explicit rule on a screen with visual examples. Participants then progressed

through the training phase, followed by the test phase. Finally, participants answered

demographic and debriefing questions and completed a brief test of color-blindness using

Ishihara plates. All instructions were provided in text form on the computer screen. Consent,

demographic, debriefing, and colorblindness items were administered through Qualtrics.

Training and test phase categorization tasks were programmed in PsychoPy (Pierce et al., 2019),
INTERACTION IN LEARNING 16

uploaded to the Pavlovia online experiment platform ([Link]), and embedded in

Qualtrics (Qualtrics, Provo UT).

Training phase
Participants were given a cover story in which they were asked to match aliens to their

appropriate vehicles—aliens of one category used rockets, while the other category used saucers

(no verbal labels were given for vehicle type). Participants were instructed on the explicit rule

through slides that explained the rule (e.g. “Group 1 aliens have square mouths and an odd

number of eyes”) and provided examples. The training task then began. A short practice block of

12 trials preceded the four training blocks. The stimuli in the practice block were sequenced to

highlight the diagnostic features differentiating the categories (e.g., Category A stimulus with

one eye followed by Category B stimulus with one eye). In all other ways practice trials were

identical to training trials.

On each trial, participants were shown an alien and two images of spaceships—one

rocket, one saucer-shaped. The rocket was always on the left side of the screen and the saucer on

the right. Participants categorized each alien by choosing rocket or saucer (left/right) with a key

press (f/j). Participants received feedback in the form of the words “correct” or “incorrect”

displayed on the screen. Feedback was based solely on the explicitly instructed XOR mouth/eyes

rule. Category-ship combination was counterbalanced across participants, so for half the

participants Category A aliens used the rocket and Category B aliens used the saucer, and vice

versa for the other half. To encourage accurate performance, incorrect responses were followed

by a 3 second delay with countdown, serving as a penalty for errors. The task screen also

included a “power bar” showing cumulative accuracy; in the pre-task instructions, participants

were told that cumulative accuracy above 70% was necessary “to win the game.”
INTERACTION IN LEARNING 17

Stimuli were presented in 4 blocks of 44 trials; each block was roughly even in terms of

category, color, subcategory, and non-diagnostic features. Stimulus order was pseudorandom

such that no more than 3 trials from the same category appeared consecutively and consecutive

same-color or same-eye-number trials were similarly limited. Trial order within a block was

fixed, but block order was random across participants.

Test phase
On each trial, participants again saw an alien and were asked to match the alien with its

ship, using f/j keys as before. Unlike the prior Training phase, no feedback was provided, with

only a fixation cross presented between trials. Trials were either congruent, such that the

category membership based on shape-based rule and color probability were the same, or

incongruent such that the shape-based rule and color probability would predict different category

membership (e.g., a warm-colored example that followed the shape rules for the cool-trained

category). Unbeknownst to participants, the test phase stimuli consisted of an even number of

congruent and incongruent items from each category. Stimuli in each category were matched for

the frequency of diagnostic and non-diagnostic features. All test phase stimuli were novel (i.e.,

not presented in the training phase).

Post-tests and survey questions

Colorblindness items. Participants were asked to type the numbers visible to them in a set of

5 Ishihara plates selected to probe for deficiencies in color vision. Before any exclusions, about

1% of participants responded to the Ishihara plates in ways consistent with some form of

colorblindness. However, performance for these participants was comparable to that of other

participants and the main RT difference between congruent and incongruent trials at test was

similar in colorblind and non-colorblind participants. It is likely the case that they were able to

perceive the differences in color distributions based on brightness differences even though they
INTERACTION IN LEARNING 18

may not have been able to perceive all hue differences. For this reason, we did not exclude

participants on the basis of their Ishihara plate responses or self-reports of colorblindness.

Strategy use and color awareness questions. Participants were also asked to complete the

following open-ended questions:

1. Describe the rule you used to classify the aliens (in your own words, to the best of your

ability).

2. Other than the rule you were instructed to use, did you use any strategy or rule of thumb

to decide which aliens went with which ships? (if yes, please describe briefly if you can)

3. Did you notice anything about the colors of the aliens? If yes, please describe below

4. Describe what (if anything) you noticed about the colors of the aliens.

5. Did you use the colors to help the aliens find their ships? (yes/no)

As we were primarily interested in effects of implicit sensitivity to the color distribution on

categorization performance, responses to any of these questions that suggested a use of color to

classify stimuli, or awareness of the biased color distributions resulted in the participant’s data

being excluded from further analysis (roughly 10% of total participants were excluded on this

basis).

Explicit color knowledge task. A final test for explicit knowledge of the color-category

associations was given in the form of a classification task like the training task, but with the

cover story that now the aliens were viewed from behind, i.e. their “facial features” were not

visible, only colors. Participants completed 16 trials (one trial for each color plus 4 additional

trials using peak colors), and then after each classification trial, to rate their confidence in their

decision on a scale from 1 to 4 (1=guessing, 4=completely confident).

Analysis
INTERACTION IN LEARNING 19

Participant exclusions

Low performance. Fifty (30) participants performed below 70% accuracy on the training

phase; most of these participants scored near chance (M = .513, SD = .079), and many had

suspiciously low reaction times (M = 2.395 seconds, SD = 1.200 seconds), so they were excluded

from further analysis on the assumption that they had not engaged with the task in earnest.

In addition, 11 participants were excluded from test phase analysis for test phase scores

lower than 70% accuracy.

Survey responses. In addition, participants were excluded for the following reasons

based on their responses to open-ended questions: Answered ‘yes’ to “did you use colors to

classify aliens” (N=10); response demonstrated awareness of color distribution (N=8); response

demonstrated lack of understanding of instructions (N=3). These categories are not mutually

exclusive, so they do not sum to the total number of excluded participants. After all exclusions,

the final sample size for analysis was 202.

Peak colors

Within each color distribution, four colors appeared most frequently (warm: yellow,

goldenrod, orange red; cool: blue, blue-green, green-blue, cyan) and these colors also had the

strongest association to one category over another (appearing with the highest frequency in one

category and lowest frequency in the other; see Figure 1). These we refer to as “peak colors.”

The other, intermediate, colors (magenta, red-purple, blue-purple, green) appeared with lower

total frequency and appeared with similar (though not identical) frequencies in each category.

Because of these frequency characteristics, the non-peak colors were not expected to show strong

color-category associations. In order to examine color-category associations, we used only trials

INTERACTION IN LEARNING 20

with peak colors in the test phase analysis. Because the colors were evenly distributed in the test

phase, this meant that we used only the 2/3 of trials with peak colors in our analysis.

Training phase
Accuracy analysis. Mean accuracy was computed for each participant and each block.

Mean accuracy was analyzed using a one-way ANOVA with blocks (1-4) as a within-subjects

factor.

Reaction time analysis. Mean reaction time was computed for each participant and each

block. Mean reaction time was analyzed using a one-way ANOVA with blocks (1-4) as a within-

subjects factor.

Test phase
Accuracy analysis. Each participant’s mean accuracy by condition

(congruent/incongruent) was calculated separately. A paired t-test was performed between

conditions.

Reaction time analysis. Before reaction time by participant and condition was analyzed,

reaction time data were cleaned as follows: 1) only correct trials were included 2) each

participant’s mean and SD for RT were calculated; trials that were above or below 3 SDs of that

participant’s mean were excluded. In addition, examining the histograms for reaction time means

and standard deviations, 8 participants were excluded based on exceptionally long (mean >4.5

seconds) or exceptionally variable (SD>5 seconds) reaction times were excluded.

Finally, each participant’s mean reaction time by condition (congruent/incongruent) was

calculated, and a paired t-test was performed between conditions.

Stimuli, code for creating stimuli, analysis scripts, and de-identified data for both experiments

(including full survey responses) can be found at:

INTERACTION IN LEARNING 21

[Link]

[anonymized link for review]. This study was not preregistered.

Results
Task performance
Training phase
Accuracy. Mean accuracy across all participants and blocks was high (M = .912, SD

= .143) and increased across blocks, reflecting improved classification performance over the

course of the task (F (4,880) = 70.67, p<.001; see Figure 2).

Reaction time. Participants varied in mean reaction time, but the overall mean was (M =

2.395 seconds, SD = 1.200 seconds). On average, reaction times decreased across blocks,

providing additional evidence that performance improved as the task went on (F (4,880) =

195.44, p<.001).

Test Phase
Overall performance. Participants performed well overall on the test, with an average

accuracy of .979 (SD =.14).

Accuracy analysis. Accuracy was slightly lower for incongruent trials than congruent

trials, but the difference was not significant (Mcongruent = .980, Mincongruent = .977; t (208) = 1.47, p

= .143).

Reaction time analysis. Critically, reaction times were significantly longer for

incongruent trials than congruent trials (Mcongruent = 1.74 seconds, Mincongruent = 1.81 seconds; t

(201) = 2.28, p = .023; see Figure 2).

Explicit Color Knowledge Post-Test

On average, participants scored at chance on the explicit knowledge post-test (where

categorization accuracy was based solely on alien color, in the absence of facial feature
INTERACTION IN LEARNING 22

information; M = .488, SD = .123, t (141) = 1.08, p = .281)4. Overall, participants reported low

confidence in their categorization decisions on the task: the mean slider response was 1.7 (on a

range where 1=guessing to 4=completely confident). Furthermore, there was no significant

difference in confidence ratings for correct versus incorrect trials (Mcorrect = 1.70; Mincorrect = 1.69; t

(141) = .423, p=.673).

Strategy use (self-report)

89% of participants gave some response to the question “Describe the rule you used to

classify the aliens (in your own words, to the best of your ability).” Of these, 78% gave a

response that referred generally to the eyes and mouth or to the parity of the eyes and the shape

of the mouth, or referred to using a rule but did not describe the rule. We interpret these

responses as evidence that shape (XOR rule)-based classification in these participants was

supported by declarative knowledge.

Interim Discussion (Experiment 1)

Participants were explicitly taught a difficult disjunctive rule for categorization of stimuli

(“aliens”) based on a combination of eye and mouth characteristics, and successfully applied this

rule during a training phase with feedback. During training, stimuli in each category were

presented according to a biased color distribution, in effect creating a probabilistic color-

category association (for a precedent, see Schoenlein & Schloss, 2022). Critically, in a

subsequent test phase without feedback, participants showed faster categorization performance

for trials in which the color-category associations were preserved (“congruent” trials) than on

trials in which the association was violated (“incongruent” trials). This reaction time effect
4
This was true for the sample mean regardless of whether “aware” participants were included or excluded, although
those individuals did show higher-than-chance mean accuracy. The mean accuracy for “aware” participants was
above chance M=.578, SD = .225; and the mean confidence rating for correct trials by aware participants was higher
than their mean confidence rating for incorrect trials (Mcorrect = 1.76; Mincorrect = 1.56). However, due to the small
number of aware participants who completed all tasks (n=4), these differences were not significant.
INTERACTION IN LEARNING 23

occurred in the absence of participants’ conscious awareness of color differences between the

categories, as assessed through questionnaire responses. Furthermore, participants’ performance

on an explicit color-based categorization task was at chance, their confidence in making these

explicit color-based category decisions was low. Therefore, participants were not consciously

aware of the color-category association, but nonetheless over time acquired implicit sensitivity to

this association, as demonstrated by facilitated responses to congruent trials. Importantly, their

confidence was unrelated to their categorization performance, meeting the “zero-correlation

criterion” (Dienes & Berry, 1997). According to this criterion, the lack of correlation between

performance and confidence can be interpreted as evidence that participants lack awareness that

they have learned, and therefore that the learning is indeed implicit (“below the subjective

threshold”); the zero-correlation criterion is considered stronger evidence of implicitness than

free self-report.

These results provide initial evidence that declarative and procedural learning systems

contribute to category selection within a single trial. On correct incongruent trials, participants

correctly applied the explicit categorization rule, while their performance was simultaneously

slowed by the atypical colour associated with the item. Participants not only reported no

knowledge of the color information, but were at chance in an explicit color-category association

test, demonstrating that the color information was learned unconsciously and use automatically,

meeting part of the definition for procedural learning. However, it is unclear whether these

effects are due to the acquisition of abstract knowledge (another feature of procedural learning:

Seger 1994; Reber 1989, 1991) or if they can be explained by a simpler mechanism, such as

stimulus-response association or conditioning.

INTERACTION IN LEARNING 24

In order to perform an internal replication of the effects and to test whether the effects

observed in Experiment 1 were due to changes beyond simple stimulus-response association

learning, we performed a second experiment to examine transfer of acquired knowledge for the

color-category association. Unlike Experiment 1, in Experiment 2 the task format for the test

phase differed from that of the training phase; in all other respects, the two experiments were

identical.

Experiment 2

Methods
Participants
A total of 249 undergraduates (ages 17 to 25, mean age 18.4 years; 44% male, 54%

female, 2% other/decline to state/non-binary/genderqueer) participated via an online platform for

course credit. After all exclusions (see below), the final sample size for analysis was n=190.

Procedure
Overview
As in Experiment 1, participants completed a training phase and test phase, as well as

post-task survey and demographic questions.

Training phase
The training phase for Experiment 2 used the same stimuli and paradigm used for training

in Experiment 1.

Test phase
Recall that in Experiment 1, each test trial involved presenting participants with a single

alien and two ships. In contrast, in Experiment 2, participants were shown two aliens, one from

Category A and one from Category B, and one ship, and were instructed to choose which alien

corresponded to the presented ship. We manipulated the color congruency of both aliens as a

pair. Specifically, in half the trials, the colors of both aliens were congruent with the training
INTERACTION IN LEARNING 25

color distribution (e.g., A-warm/B-cool); in the other half of trials, both aliens were presented in

incongruent colors. All stimuli presented were novel (i.e., not previously seen in the training

task). To avoid confusion with the training left/right configuration of the ships, the aliens were

stacked vertically and response keys were u/n (upper or lower). Participants were presented with

a total of two blocks of 38 trials each (one saucer block, one rocket block).

Pairs in the congruent and incongruent conditions were balanced for total shared features,

shared diagnostic features, shared non-diagnostic features, binned warmness difference, and

mean distance between colors in color space (delta e).

Analysis

Participant exclusions

Low performance. Thirteen (13) participants performed below 70% accuracy on the

training phase and were excluded from further analysis on the assumption that they had not

engaged with the task in earnest. In addition, 17 participants were excluded from test phase

analysis for test phase scores lower than 70% accuracy.

Survey responses. In addition, participants were excluded for the following reasons

based on their responses to open-ended questions. Participants whose responses to survey

questions indicated knowledge of the color-category association (n=9) and/or who answered

“Yes” to “did you use color to categorize the aliens” (n=11) were excluded from further analysis

(total n=17). After these exclusions, the final sample size for accuracy analysis was 190.

Colors used in analysis

In order to maintain balance across the values of warmness difference and colipase

distance, pairs including all colors, not just peak colors, were used in the analysis.
INTERACTION IN LEARNING 26

Training phase

Accuracy analysis. As in experiment 1, mean accuracy was computed for each

participant and each block. Mean accuracy was analyzed using a one-way ANOVA with blocks

(1-4) as a within-subjects factor.

Reaction time analysis. As in experiment 1, mean reaction time was computed for each

participant and each block. Mean reaction time was analyzed using a one-way ANOVA with

blocks (1-4) as a within-subjects factor.

Test phase

Accuracy analysis. Each participant’s mean accuracy by condition

(congruent/incongruent) was calculated as in experiment 1

Reaction time data cleaning and analysis. Before reaction time by participant and

condition was analyzed, reaction time data were cleaned as follows: 1) only correct trials were

included 2) each participant’s mean and SD for RT were calculated; trials that were above or

below 3SDs of that participant’s mean were excluded 3) participants with exceptionally long

(mean >6 seconds) or exceptionally variable (SD > 6 seconds) were excluded; these cutoffs were

determined by visual inspection of the histograms for participant means and standard deviations

of reaction time, respectively. After these exclusions, n=173 for reaction time analysis5.

After data were cleaned as described above, each participant’s mean reaction time by

condition (congruent/incongruent) was calculated.

5
Without excluding participants with exceptionally long or variable reaction times, the main effected reported for
this experiment is still significant: (MINCONGRUENT =3.28s, MCONGRUENT = 3.07s; MRTDIFF = .143s, t(189)= 4.005,
p<.0001 )
INTERACTION IN LEARNING 27

Results
Task performance
Training phase
Mean accuracy (M=.93 SD= .05) and mean reaction time (M = 2.07s. SD =1.46s) on the training

task were comparable to Experiment 1. As in Experiment 1, performance improved over training

blocks, as seen in significant effects of block on accuracy (F (4, 760) = 47.81, p<.001) and

reaction time (F (4, 756) = 228.43, p<.001).

Test Phase

Overall performance. Mean accuracy was comparable to test phase accuracy in

Experiment 1(M =.964, SD = .187). Mean response times were longer than in Experiment 1 (M =

2.90s SD = 2.54s) after reaction time data cleaning (see above), reflecting the greater difficulty of

the task.

Accuracy analysis. Although accuracy was numerically higher for congruent trials than

incongruent trials, this difference did not reach significance (MCONGRUENT = .966, MINCONGRUENT

= .960; t (189) = 1.75, p = .08).

Reaction time analysis. As in Experiment 1, participants’ RTs were significantly slower

for incongruent than congruent trials: (MINCONGRUENT =2.862s, MCONGRUENT = 2.719s; MRTDIFF

= .143s, t (172) = 4.68, p<.0001). Figure 3 shows the comparison of RTs across condition, within

participant.

Strategy use (self-report)

87% of participants gave some response to the question “Describe the rule you used to

classify the aliens (in your own words, to the best of your ability).” Of these, 85% gave a

response that referred generally to the eyes and mouth or to the parity of the eyes and the shape

of the mouth. An additional 3% referred to using a rule but did not describe the rule. We interpret
INTERACTION IN LEARNING 28

these responses as evidence that shape (XOR rule)-based classification in these participants was

supported by declarative knowledge.

Interim Discussion (Experiment 2)

Participants completed a training phase with a biased color distribution and a test phase

that required transferring the knowledge gained in training to a novel test format. As in

Experiment 1, responses were faster when the colors presented were congruent with the training

phase color distribution, suggesting an influence of the implicit color distribution on

categorization decisions. As in Experiment 1, this effect occurred outside of participants’ explicit

awareness of the color distribution. This finding is noteworthy given that some forms of non-

declarative learning are not seen when response format is altered, and suggests that the observed

effect reflects more than simple stimulus-response association (Vaquero et al., 2019), but rather

reflects some abstract representation of the stimulus space.

General Discussion

In two experiments, we demonstrated that procedural learning can occur simultaneously

with declarative learning (consistent with previous findings), and that both procedurally-learned

information and declaratively-learned information can contribute to category response selection

within a single trial (not previously demonstrated). Participants successfully used a declarative,

verbalizable rule to categorize novel stimuli, while simultaneously demonstrating effects of a

procedurally learned color distribution, as demonstrated by slower response times to trials in

which the rule and color distribution conflicted. These results cannot be explained by current

models of winner-take-all competition between procedural and declarative learning at encoding

(McDonald et al., 2004; McDonald & Hong, 2013), consolidation (Brown & Robertson, 2007;

Galea et al., 2010), or retrieval (Crossley & Ashby, 2015).

INTERACTION IN LEARNING 29

There are at least two forms of evidence that the XOR rule was learned and implemented

by a declarative form of learning. In their short debrief responses, a majority (>85%) of

participants in both experiment 1 and experiment 2 explicitly referred to the eyes and mouth

(diagnostic features), or to “the rule that was given,” and sometimes reported verbalizable

heuristics based on the XOR rule. Furthermore, performance accuracy in the first block of

training trials was near ceiling (Figure 2a and 3a) in both experiment 1 and experiment 2,

suggesting that participants immediately applied the explicitly-instructed rule from the previous

instruction phase, rather than searching for a rule or gradually accumulating information about

the shape-category relation. In addition, as all stimuli shown in the test phase were novel (i.e.

not shown in the training phase), participants could not use memory for specific exemplars to

categorize the test stimuli. Given that participants were explicitly instructed to use the XOR rule

in their decisions, practiced implementing this rule prior to the learning blocks, and performed

with high accuracy (>90%) according to this rule, it is reasonable to assume that declarative

knowledge of the XOR rule guided their categorization decisions.

In addition, there are also multiple forms of evidence that the color distribution

information was learned procedurally (and not declaratively). In their short debrief responses,

very few participants (<10%) reported any explicit knowledge of the biased color distributions;

those who did were excluded from the analysis intended to show interaction between declarative

and procedural information. Additionally, in the post-test after Experiment 1, participants were at

chance when asked to classify stimuli based on color alone, and their confidence ratings did not

significantly differ between correctly and incorrectly classified colors (meeting the zero-

correlation criterion; Dienes & Berry, 1997). Thus, taken together, participants in our final

sample demonstrated no awareness of the hidden color distribution, but nevertheless were slower
INTERACTION IN LEARNING 30

on incongruent trials. For these reasons, we infer that the color distribution was learned

procedurally. Despite the fact that the reversed form of the test phase could potentially

disadvantage a procedural learning system (Anderson & Fincham, 1994; Vaquero et al, 2020),

we interpret the maintenance of the effect across this reversal as evidence of abstract learning

beyond stimulus-response association (Reber,1991; Seger, 1994), such as the formation of a

probabilistic color-category stimulus space.

Our findings our consistent with previous research showing that procedural encoding can

take place simultaneously and “covertly” during declarative encoding (Foerde et al., 2006;

Packard & McGaugh, 1996; Song et al., 2007). However, other previous research has

demonstrated competition at consolidation (Brown & Robertson, 2007; Galea et al., 2010) and/or

retrieval (Crossley & Ashby, 2015; Packard & McGaugh, 1996). The results of the current study

demonstrate that, at least in some situations, the competition at retrieval is not a zero-sum game;

the output from one system is not completely disregarded or discarded in the process of decision-

making. When procedural knowledge and declarative knowledge converge on a categorization

decision, responses are facilitated; when they diverge, responses are hindered.

Many of the previous studies on interaction between procedural and declarative learning

have focused on whether both systems are active during encoding or retrieval, and whether one

system’s activity inhibits or facilitates activity in the other (see e.g., Freedberg, 2020; for

exceptions see Robertson, 2022). However, the current study involves interactions between the

representations or knowledge formed by each system, not just their relative activity. In doing so,

it presents a challenge for computational models such as COVIS that consider only the
INTERACTION IN LEARNING 31

confidence and bias of each module’s decision, but do not consider whether the modules’

responses converge or diverge for downstream response selection6.

One practical implication of the current findings is that in complex, real-world situations,

both declarative and procedural decision-making processes may interact. For example, in formal

instruction, students are often given rules or necessary and sufficient criteria for category

membership (e.g., mammals have hair and produce milk to feed offspring). However, even when

students are given such a rule or criteria, if they experience only a biased selection from the

space of possible category members (e.g. only dogs and cats as mammals), they may have

difficulty transferring their knowledge of mammals in general to unfamiliar exemplars (whales,

armadillos). Our main finding is consistent with an emerging body of evidence that supports the

idea of sampling broadly from the example space during instruction to facilitate later transfer

(Carvalho, Chen, & Yu; Nosofsky et al. 2018). However, other studies have suggested that more

narrow sampling, particularly focused on a category’s central tendency may be beneficial for

learners (Bowman & Zeithamova, 2020, 2023; Homa & Vosburgh, 1976; Homa & Cultice,

1984). Further research will be required to determine when and why narrow versus broad

sampling is beneficial for robust learning and transfer. Considering interactions between

procedural and declarative knowledge may clarify the conditions under which broad versus

narrow sampling is beneficial in instruction.

Our results also provide insights into the formation of stereotypes and potential ways to

counteract learned negative associations. Similar to many studies using the Implicit Association

Test (Greenwald & Banaji, 2017), the current findings provide further experimental evidence

6
A simple modification to the gating mechanism of COVIS could make it compatible with the current results.
Briefly, the gating mechanism would need to take into account the relative direction (sign) of each module’s
decision (e.g. Category A positive, Category B negative) in addition to the existing confidence and bias parameters.
The modified gating mechanism could be likened to the mixture-of-experts seen in the ATRIUM model (Krushke,
1992, 2011).
INTERACTION IN LEARNING 32

that behavior is often influenced by a combination of implicit representations based on

accumulated experiences and explicit understandings or beliefs. Generally, the IAT reveals

reaction time differences whereby pairing a social group with traits that are stereotypically

associated with that group (e.g., male – engineer) results in faster reaction times than pairing

with traits that are stereotypically not associated (e.g., male – nurse; Banaji & Hardin, 1996).

Importantly, these reaction time differences are found even when individuals hold explicit beliefs

that contradict the stereotypes (Greenwald & Banaji, 1995; Greenwald et al., 1998); somehow,

exposure to stereotypes in the environment (for example through media portrayals) creates an

implicit bias that can operate despite an individual’s best intentions to hold unbiased attitudes

towards particular social groups. Our findings are consistent with this general model of

interaction between implicit and explicit representations. While many successful interventions

for counteracting implicit bias recruit explicit reasoning and declarative knowledge (Devine et

al., 2012; Forscher et al., 2016), our findings support suggestions that interventions that recruit

procedural learning mechanisms, such as the presentation of counter-stereotypical exemplars,

may be an important complement (Fitzgerald et al., 2019; Forscher et al., 2016), since both

procedural and declarative knowledge can contribute to real-time decision making.

Another area for consideration is the developmental trajectory of interaction between

procedural and declarative learning. We know that procedural and declarative learning have

different developmental trajectories, with procedural learning maturing sooner (Finn, Kalra, et

al., 2015). For this reason, we might predict that interactions between procedural and declarative

learning in young children might show a different pattern than that observed in healthy, young

adults, with stronger contributions from the procedural system. However, several studies of

category learning with children have suggested that children perseverate in using simple rule-
INTERACTION IN LEARNING 33

based strategies based on declarative learning (Huang-Pollock et al., 2011; Rabi et al., 2015;

Rabi & Minda, 2014). The explanation proposed by these studies is that the children are unable

to inhibit the output of the declarative system, which is consistent with the relative immaturity of

prefrontal areas responsible for inhibition as well as other means of adjudicating between

different response options (Gogtay et al., 2004; Lenroot & Giedd, 2006; Shaw et al., 2006).

However, these studies contrasted category induction for rule-based (declarative) and

information-integration (procedural) category structures; the children had to infer the rule rather

than being told a rule before the task started. Pure induction reflects the way naturalistic

categories are learned, but is not necessarily how categories and concepts are typically learned in

formal instructional contexts, like classrooms. In these contexts, instruction often emphasizes

declarative knowledge without considering the role that covertly learned procedural knowledge

may play. Future studies, such as those that adapt the current paradigm for use with children, will

be needed to fully address the issue of developmental differences in interactions between

declarative and procedural learning, particularly when considering instructional implications.

On that note, one important limitation of the current study is the lack of ecological validity.

Although we have suggested instructional implications, we hasten to point out that the paradigm

used in this study bears only a faint resemblance to learning in a classroom setting. In particular,

fewer exemplars can be presented in a classroom, and they are usually distributed more widely

across time and without immediate feedback. It is possible that the selection of examples from

the space creates an implicit representation that could conflict with a rule-based category

definition in real-world learning settings, just as in our laboratory paradigm, but this must be

confirmed empirically in studies with greater ecological validity.

INTERACTION IN LEARNING 34

The current results also raise an important theoretical question for theories of memory

systems. Specifically, do the procedural and declarative learning systems create independent,

discrete representations of the stimulus space that are integrated into a common response only at

the time of response selection, or do they contribute to a single representation of the stimulus

space that directly guides retrieval? Although the current behavioural results cannot speak to this

issue, we are currently conducting a fMRI-representational similarity analysis (RSA) study to

address this question. These neuroimaging results will also clarify where in the brain and when

in the course of processing (at encoding or later at response selection) the interaction or overlap

takes place; candidate areas include regions that support reasoning and decision-making, such as

prefrontal cortex (Krawczyk, 2002).

Finally, looking backward, rather than forward, it may be possible to reinterpret some

traditionally puzzling findings (such as Brooks & Allen, 1991; Armstrong, Gleitman, &

Gleitman, 1983) in the context of interaction between rule-based and similarity-based

representations of a stimulus space. Brooks and Allen (1991) found that participant

categorization decisions were skewed by similarity along non-diagnostic (but partially

predictive) dimensions; it could be the case that this interaction between rule-based and

similarity-based reasoning was an interaction between declarative and procedural knowledge, as

in our study, but to an extent that influenced accuracy as well as response times. Armstrong et al.

(1983) demonstrated that participants would, when prompted, gave responses that suggested a

graded representation even for categories with strict criteria (such as odd numbers). These

apparently paradoxical findings could potentially be explained in terms of task demands that

preferentially recruit procedural or declarative knowledge for decision making. Future research

could include replications of these classic studies with specific controls and measures in place to
INTERACTION IN LEARNING 35

monitor the relative contributions of procedural and declarative knowledge. Thus, the findings of

the current study open new directions for both basic and applied research, potentially deepening

our fundamental understanding of human cognition as well as addressing important aspects of

human learning and development outside the laboratory.

 INTERACTION IN LEARNING 36

References

Albouy, G., King, B. R., Maquet, P., & Doyon, J. (2013). Hippocampus and Striatum: Dynamics and
Interaction During Acquisition and Sleep-Related Motor Sequence Memory Consolidation.
HIPPOCAMPUS, 23(11), 985–1004. [Link]
Allen, S. W., & Brooks, L. R. (1991). Specializing the operation of an explicit rule. Journal of
experimental psychology: General, 120(1), 3.
Anderson, J. R., & Fincham, J. M. (1994). Acquisition of procedural skills from examples. Journal of
experimental psychology: learning, memory, and cognition, 20(6), 1322.
Ashby, F. G., Alfonso-Reese, L. A., Turken, A. U., & Waldron, E. M. (1998). A neuropsychological
theory of multiple systems in category learning. Psychological Review, 105(3), 442–481. https://
[Link]/10.1037/0033-295X.105.3.442
Ashby, F. G., & Crossley, M. J. (2010). Interactions between declarative and procedural-learning
categorization systems. Neurobiology of Learning and Memory, 94(1), 1–12.
[Link]
Ashby, F. G., & Ell, S. W. (2001). The neurobiology of human category learning. Trends in Cognitive
Sciences, 5(5), 204–210. [Link]
Ashby, F. G., & Ennis, J. M. (2006). The role of the basal ganglia in category learning. Psychology of
Learning and Motivation, 46, 1-36.
Ashby, F. G., & O’Brien, J. B. (2005). Category learning and multiple memory systems. TRENDS IN
COGNITIVE SCIENCES, 9(2), 83–89. [Link]
Banaji, M. R., & Hardin, C. D. (1996). Automatic Stereotyping. Psychological Science, 7(3), 136-
141. [Link]
Batterink, L. J., Oudiette, D., Reber, P. J., & Paller, K. A. (2014). Sleep facilitates learning a new
linguistic rule. Neuropsychologia, 65, 169-179.
Berger, S. & Batterink, L.J. (in press) Children extract a new linguistic rule more quickly than adults.
Developmental Science
Bowman, C. R., & Zeithamova, D. (2023). Coherent category training enhances generalization in
prototype-based categories. Journal of Experimental Psychology: Learning, Memory, and
Cognition.
Bowman, C. R., & Zeithamova, D. (2020). Training set coherence and set size effects on concept
generalization and recognition. Journal of Experimental Psychology: Learning, Memory, and
Cognition, 46(8), 1442.
Brown, R. M., & Robertson, E. M. (2007). Inducing motor skill improvements with a declarative task.
Nature Neuroscience, 10(2), 148–149. [Link]
 INTERACTION IN LEARNING 37

Brown, R. M., Robertson, E. M., & Press, D. Z. (2009). Sequence Skill Acquisition and Off-Line
Learning in Normal Aging. PLOS ONE, 4(8). [Link]
Cabeza, R., & Moscovitch, M. (2013). Memory Systems, Processing Modes, and Components:
Functional Neuroimaging Evidence. Perspectives on Psychological Science, 8(1), 49–55. https://
[Link]/10.1177/1745691612469033
Carvalho, P. F., Chen, C. H., & Yu, C. (2021). The distributional properties of exemplars affect
category learning and generalization. Scientific reports, 11(1), 11263.
Chaffin, R., Logan, T. R., & Begosh, K. T. (2009). Performing from Memory. Oxford Handbook of
Music Psychology, 352.
CIE (Commission Internationale de l’Eclairage). (1986) Colorimetry. 2nd Edition, Publication CIE
No. 15.2. Commission Internationale de l’Eclairage, Vienna.
Cohen, N. J., & Squire, L. R. (1980). Preserved learning and retention of pattern-analyzing skill in
amnesia: Dissociation of knowing how and knowing that. Science, 210(4466), 207-210.
Crossley, M. J., & Ashby, G.F. (2015). Procedural learning during declarative control. Journal of
Experimental Psychology: Learning Memory and Cognition, 41(5), 1388–1403.
[Link]
Devine P.G., Forscher P.S., Austin, A.J., and Cox, W.T.L. (2012) Long-term reduction in implicit
race bias: A prejudice habit-breaking intervention. Journal of Experimental Social Psychology
48(6): 1267-1278
Dickerson, K. C., Li, J., & Delgado, M. R. (2011). Parallel contributions of distinct human memory
systems during probabilistic learning. Neuroimage, 55(1), 266–276.
[Link]
Dienes, Z., & Berry, D. (1997). Implicit learning: Below the subjective threshold. Psychonomic
Bulletin & Review, 4(1), 3–23. [Link]
Eichenbaum,H.,& Cohen,N. J. (2001).From conditioning to conscious recollection: Memory systems
of the brain. New York: Oxford University Press.
Filoteo, J. V., & Maddox, W. T. (2014). Procedural-based category learning in patients with
Parkinson’s disease: impact of category number and category continuity. Frontiers in Systems
Neuroscience, 8, 14. [Link]
Filoteo, J. V., Maddox, W. T., Ing, A. D., Zizak, V., & Song, D. D. (2005). The impact of irrelevant
dimensional variation on rule-based category learning in patients with Parkinson’s disease.
Journal of the International Neuropsychological Society : JINS, 11(5), 503–513.
[Link]
Filoteo, J. V., Maddox, W. T., Salmon, D. P., & Song, D. D. (2007). Implicit category learning
performance predicts rate of cognitive decline in nondemented patients with Parkinson’s disease.
Neuropsychology, 21(2), 183–192. [Link]
 INTERACTION IN LEARNING 38

Filoteo, J. V, Maddox, W. T., Salmon, D. P., & Song, D. D. (2005). Information-integration category
learning in patients with striatal dysfunction. Neuropsychology, 19(2), 212–222.
[Link]
FitzGerald, C., Martin, A., Berner, D. et al. Interventions designed to reduce implicit prejudices and
implicit stereotypes in real world contexts: a systematic review. BMC Psychol 7, 29 (2019).
[Link]
Foerde, K., Knowlton, B. J., & Poldrack, R. A. (2006). Modulation of competing memory systems by
distraction. Proceedings of the National Academy of Sciences of the United States, 103(31),
11778–11783. [Link]
Foerde, K., Poldrack, R. A., & Knowlton, B. J. (2007). Secondary-task effects on classification
learning. Memory & Cognition, 35(5), 864–874.
Foerde, K. (2018). What are habits and do they depend on the striatum. Current Opinion In
Behavioral Sciences, 20, 17–24. [Link]
Foerde, K., & Poldrack, R. A. (2008). Procedural Learning in Humans. Psychological Review,
December, 1083–1091. [Link]
Foerde, K., & Shohamy, D. (2011a). Feedback Timing Modulates Brain Systems for Learning in
Humans. Journal of Neuroscience, 31(37), 13157–13167.
[Link]
Foerde, K., & Shohamy, D. (2011b). The role of the basal ganglia in learning and memory: Insight
from Parkinson’s disease. Neurobiology of Learning and Memory, 96(4, SI), 624–636.
[Link]
Forscher, P. S., Lai, C. K., Axt, J. R., Ebersole, C. R., Herman, M., Devine, P. G., & Nosek, B. A.
(2019). A meta-analysis of procedures to change implicit measures. Journal of Personality and
Social Psychology, 117(3), 522–559.
Freedberg, M., Toader, A. C., Wassermann, E. M., & Voss, J. L. (2020). Competitive and cooperative
interactions between medial temporal and striatal learning systems. In Neuropsychologia (Vol.
136). Elsevier Ltd. [Link]
Gabrieli, J. (1998). Cognitive neuroscience of human memory. Annual Review of Psychology, 49, 87–
115. [Link]
Galea, J. M., Albert, N. B., Ditye, T., & Miall, R. C. (2010). Disruption of the Dorsolateral Prefrontal
Cortex Facilitates the Consolidation of Procedural Skills. Journal of Cognitive Neuroscience,
22(6), 1158–1164. [Link]
Gluck, M. A, Shohamy, D., & Myers, C. (2002). How do people solve the “weather prediction” task?:
individual variability in strategies for probabilistic category learning. Learning & Memory (Cold
Spring Harbor, N.Y.), 9(6), 408–418. [Link]
 INTERACTION IN LEARNING 39

Gogtay, N., Giedd, J. N., Lusk, L., Hayashi, K. M., Greenstein, D., Vaituzis, A. C., Nugent, T. F.,
Herman, D. H., Clasen, L. S., Toga, A. W., Rapoport, J. L., & Thompson, P. M. (2004).
Dynamic mapping of human cortical development during childhood through early adulthood.
Proceedings of the National Academy Of Sciences of the United States of America, 101(21),
8174–8179. [Link]
Graf, P., & Schacter, D. L. (1985). Implicit and explicit memory for new associations in normal and
amnesic subjects. Journal of Experimental Psychology: Learning, memory, and cognition, 11(3),
501.
Greenwald, A. G., & Banaji, M. R. (1995). Implicit social cognition: Attitudes, self-esteem, and
stereotypes. Psychological Review, 102(1), 4–27. [Link]
Greenwald, AG, DE McGhee, JLK Schwartz (1998) Measuring individual differences in implicit
cognition: the implicit association test. Journal of personality and social psychology 74(6): 1464
Hartman, R., Moss, A. J., Jaffe, S. N., Rosenzweig, C., Litman, L., & Robinson, J. (2023).
Introducing Connect by CloudResearch: Advancing Online Participant Recruitment in the
Digital Age. [Link] Retrieved [Date].
Homa, D., & Vosburgh, R. (1976). Category breadth and the abstraction of prototypical
information. Journal of Experimental Psychology: Human Learning and Memory, 2(3), 322.
Homa, D., & Cultice, J. C. (1984). Role of feedback, category size, and stimulus distortion on the
acquisition and utilization of ill-defined categories. Journal of Experimental Psychology:
Learning, Memory, and Cognition, 10(1), 83.
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.
[Link]
Knowlton, B. J., Mangels, J. A, & Squire, L. R. (1996). A neostriatal habit learning system in
humans. Science (New York, N.Y.), 273(5280), 1399–1402.
[Link]
Knowlton, B. J., Squire, L. R., & Gluck, M. A. (1994). Probabilistic Classification Learning in
Amnesia. Learning & Memory (Cold Spring Harbor, N.Y.), 1, 106–120.
[Link]
Knowlton, B. J., Squire, L. R., Paulsen, J. S., Swerdlow, N. R., Swenson, M., & Butters, N. (1996).
Dissociations within nondeclarative memory in Huntington’s disease. Neuropsychology, 10(4),
538–548.
Krawczyk, D. C. Contributions of the prefrontal cortex to the neural basis of human decision
making. Neuroscience & Biobehavioral Reviews 26.6 (2002): 631-664.
Lenroot, R. K., & Giedd, J. N. (2006). Brain development in children and adolescents: insights from
anatomical magnetic resonance imaging. Neuroscience and Biobehavioral Reviews, 30(6), 718–
729. [Link]
 INTERACTION IN LEARNING 40

Maddox, W. T., Ashby, F. G., & Bohil, C. J. (2003). Delayed feedback effects on rule-based and
information-integration category learning. JOURNAL OF EXPERIMENTAL PSYCHOLOGY-
LEARNING MEMORY AND COGNITION, 29(4), 650–662. [Link]
7393.29.4.650
Maddox, W. T., Ashby, F. G., Ing, A. D., & Pickering, A. D. (2004). Disrupting feedback processing
interferes with rule-based but not information-integration category learning. Memory &
Cognition, 32(4), 582-591.
Maddox, W. T., & Ing, A. D. (2005). Delayed feedback disrupts the procedural-learning system but
not the hypothesis-testing system in perceptual category learning. JOURNAL OF
EXPERIMENTAL PSYCHOLOGY-LEARNING MEMORY AND COGNITION, 31(1), 100–107.
[Link]
McDonald, R. J., Devan, B. D., & Hong, N. S. (2004). Multiple memory systems: The power of
interactions. Neurobiology of Learning and Memory, 82(3), 333–346.
[Link]
Mcdonald, R. J., & Hong, N. S. (2013). How does a specific learning and memory system in the
mammalian brain gain control of behavior? Hippocampus, 23(11), 1084–1102.
[Link]
McDonald, R. J., & White, N. M. (1995). Hippocampal and Nonhippocampal Contributions to Place
Learning In Rats. Behavioral Neuroscience, 109(4), 579–593. [Link]
7044.109.4.579
Meeter, M., Radicsa, G., Myers, C. E., Gluck, M. A., & Hopkins, R. O. (2008). Probabilistic
categorization: How do normal participants and amnesic patients do it? NEUROSCIENCE AND
BIOBEHAVIORAL REVIEWS, 32(2), 237–248. [Link]
Miles, S. J., & Minda, J. P. (2011). The effects of concurrent verbal and visual tasks on category
learning. Journal of Experimental Psychology. Learning, Memory, and Cognition, 37(3), 588–
607.
Norman, G. R., & Brooks, L. R. (1997). The Non-Analytical Basis of Clinical Reasoning. In
Advances in Health Sciences Education (Vol. 2). Kluwer Academic Publishers.
Nomura, E. M., Maddox, W. T., Filoteo, J. V, Ing, A. D., Gitelman, D. R., Parrish, T. B., Mesulam,
M.-M., & Reber, P. J. (2007). Neural correlates of rule-based and information-integration visual
category learning. CEREBRAL CORTEX, 17(1), 37–43. [Link]
Nosofsky, R. M., Sanders, C. A., & McDaniel, M. A. (2018). A Formal Psychological Model of
Classification Applied to Natural-Science Category Learning. Current Directions in
Psychological Science, 27(2), 129–135. [Link]
Packard, M. G., & McGaugh, J. L. (1996). Inactivation of hippocampus or caudate nucleus with
lidocaine differentially affects expression of place and response learning. Neurobiology of
Learning and Memory, 65(1), 65–72. [Link]
 INTERACTION IN LEARNING 41

Peirce, J. W., Gray, J. R., Simpson, S., MacAskill, M. R., Höchenberger, R., Sogo, H., Kastman, E.,
Lindeløv, J. (2019). PsychoPy2: experiments in behavior made easy. Behavior Research
Methods. 10.3758/s13428-018-01193-y
Poldrack, R. A., & Packard, M. G. (2003). Competition among multiple memory systems: converging
evidence from animal and human brain studies. Neuropsychologia, 41(3), 245–251.
[Link]
Poldrack, R. A., Prabhakaran, V., Seger, C. A., & Gabrieli, J. D. E. (1999). Striatal activation during
acquisition of a cognitive skill. Neuropsychology, 13(4), 564–574. [Link]
4105.13.4.564
Poldrack, R. A., & Rodriguez, P. (2004). How do memory systems interact? Evidence from human
classification learning. Neurobiology of learning and memory, 82(3), 324-332.
Qualtrics Software Company (2024). Provo, UT, USA available at [Link]
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.
[Link]
Rabi, R., & Minda, J. P. (2014). Rule-Based Category Learning in Children: The Role of Age and
Executive Functioning. PLoS ONE, 9(1), e85316. [Link]
Reber, A. S. (1989). Implicit Learning and Tacit Knowledge. Journal Of Experimental Psychology-
General, 118(3), 219–235. [Link]
Reber, P. J. (2013). The neural basis of implicit learning and memory: A review of
neuropsychological and neuroimaging research. Neuropsychologia, 51(10), 2026–2042.
[Link]
Reber, P. J., & Squire, L. R. (1994). Parallel Brain Systems for Learning with and without Awareness.
Learning & Memory, 1(4), 217–229.
Robertson, E. M. (2022). Memory leaks: information shared across memory systems. Trends in
cognitive sciences, 26(7), 544-554.
Salmon, D. P., & Butters, N. (1995). Neurobiology of skill and habit learning. Current Opinion in
Neurobiology, 5(2), 184–190. [Link]
Schoenlein, M. A., & Schloss, K. B. (2022). Colour-concept association formation for novel concepts.
Visual Cognition, 30(7), 457–479.
[Link]
F
Seger, C. A. (1994). Implicit learning. Psychological Bulletin, 115(2), 163–196.
[Link]
 INTERACTION IN LEARNING 42

Seger, C. A. (2006). The basal ganglia in human learning. The Neuroscientist : A Review Journal
Bringing Neurobiology, Neurology and Psychiatry, 12(4), 285–290.
[Link]
Seger, C. A., & Spiering, B. J. (2011). A critical review of habit learning and the Basal Ganglia.
Frontiers in Systems Neuroscience, 5(August), 66. [Link]
Shaw, P., Greenstein, D., Lerch, J., Clasen, L., Lenroot, R., Gogtay, N., Evans, A., Rapoport, J., &
Giedd, J. (2006). Intellectual ability and cortical development in children and adolescents.
Nature, 440(7084), 676–679. [Link]
Shohamy, D., Myers, C. E., Onlaor, S., & Gluck, M. A. (2004). Role of the basal ganglia in category
learning: How do patients with Parkinson’s disease learn? Behavioral Neuroscience, 118(4),
676–686. [Link]
Smith, J. D., Boomer, J., Zakrzewski, A. C., Roeder, J. L., Church, B. A, & Ashby, F. G. (2014).
Deferred feedback sharply dissociates implicit and explicit category learning. Psychological
Science, 25(2), 447–457. [Link]
Squire, L. R. (1992). Memory and the Hippocampus - A Synthesis From Findings with Rats,
Monkeys, and Humans. Psychological Review, 99(2), 195–231. [Link]
295x.99.2.195
Squire, L. R. (2004). Memory systems of the brain: a brief history and current
perspective. Neurobiology of learning and memory, 82(3), 171-177.
Squire, L. R., Hamann, S., & Knowlton, B. (1994). Dissociable Learning and Memory-Systems of the
Brain. Behavioral And Brain Sciences, 17(3), 422–423.
Squire, L. R., & Zola-Morgan, S. (1988). Memory: brain systems and behavior. Trends in
neurosciences, 11(4), 170-175.
Taylor, G. (2018) Colormath package for Python. Available at: [Link]
colormath. See also: [Link]
Turner, B. O., Crossley, M. J., & Ashby, F. G. (2017). Hierarchical control of procedural and
declarative category-learning systems. NeuroImage, 150, 150–161.
[Link]
VanLehn, K. (1996). Cognitive Skill Acquisition. Annual Review of Psychology, 47(1), 513–539.
[Link]
Vaquero, J. M., Lupiáñez, J., & Jiménez, L. (2020). Asymmetrical effects of control on the expression
of implicit sequence learning. Psychological Research, 84, 2157-2171.
Waldron, E. M., & Ashby, F. G. (2001). The effects of concurrent task interference on category
learning: Evidence for multiple category learning systems. Psychonomic Bulletin & Review, 8(1),
168–176.
 INTERACTION IN LEARNING 43

Wan, X., Takano, D., Asamizuya, T., Suzuki, C., Ueno, K., Cheng, K., Ito, T., & Tanaka, K. (2012).
Developing Intuition: Neural Correlates of Cognitive-Skill Learning in Caudate Nucleus.
Journal Of Neuroscience, 32(48), 17492–17501. [Link]
12.2012
Zeithamova, D., & Maddox, W. T. (2006). Dual-task interference in perceptual category
learning. Memory & cognition, 34(2), 387-398.
INTERACTION IN LEARNING 44

Figures

Figure 1 Experiment Design A.) Participants were told the XOR eye/mouth rule for classifying aliens. In
the training phase, items for each category were presented across a biased color distribution, either cool-
biased (Category A) or warm-biased (Category B), creating a probabilistic color-category association. In
the test phase, an even distribution of congruent and incongruent items from each category were
presented.

B.) Schematic illustrations of Experiment 1 and Experiment 2. In both experiments, feedback was given
during the training phase but not the test phase; all test phase stimuli were novel (not seen in training).
INTERACTION IN LEARNING 45

Figure 2. Experiment 1 results. A) Training accuracy by block; high initial accuracy demonstrates use of
the shape rule. B) Training reaction time decreased over successive blocks. C) Test accuracy by
condition: responses to color congruent trials were more accurate than incongruent trials, but the
difference was not significant. D) Test reaction time by condition: Trials in which color and shape
information conflicted (incongruent) were significantly slower than trials in which they converged
(congruent).

Figure 3. Experiment 2 results. A) Training accuracy by block; high initial accuracy demonstrates use of
the shape rule. B) Training reaction time decreased over successive blocks. C) Test accuracy by
condition: responses to color congruent trials were more accurate than incongruent trials, but the
difference was not significant. D) Test reaction time by condition: Trials in which color and shape
INTERACTION IN LEARNING 46

information conflicted (incongruent) were significantly slower than trials in which they converged
(congruent)