Reversible second-order conditional sequences in incidental sequence learning tasks

Abstract

In sequence learning tasks, participants’ sensitivity to the sequential structure of a series of events often overshoots their ability to express relevant knowledge intentionally, as in generation tasks that require participants to produce either the next element of a sequence (inclusion) or a different element (exclusion). Comparing generation performance under inclusion and exclusion conditions makes it possible to assess the respective influences of conscious and unconscious learning. Recently, two main concerns have been expressed concerning such tasks. First, it is often difficult to design control sequences in such a way that they enable clear comparisons with the training material. Second, it is challenging to ask participants to perform appropriately under exclusion instructions, for the requirement to exclude familiar responses often leads them to adopt degenerate strategies (e.g., pushing on the same key all the time), which then need to be specifically singled out as invalid. To overcome both concerns, we introduce reversible second-order conditional (RSOC) sequences and show (a) that they elicit particularly strong transfer effects, (b) that dissociation of implicit and explicit influences becomes possible thanks to the removal of salient transitions in RSOCs, and (c) that exclusion instructions can be greatly simplified without losing sensitivity.

Keywords

Implicit learning transfer chunking generation RSOC sequence

Introduction

Incidental sequence learning has been used extensively as one of the major ways to study implicit learning (Cleeremans, 1993; Cleeremans, Destrebecqz, & Boyer, 1998; Cohen, Ivry, & Keele, 1990; Nissen & Bullemer, 1987; Reed & Johnson, 1994; Shanks & Johnstone, 1999; Stadler, 1993). Implicit learning refers to situations where one acquires information without being aware of what is learnt, that is, in an incidental manner. Experiments in this domain typically involve, first, incidentally exposing participants to experimental material that contains structure, and subsequently, testing them on the knowledge they have acquired about this material (Saffran, Newport, Aslin, Tunick, & Barrueco, 1997). Specifically, in incidental sequence learning, the material consists in repeating series of events or sequences. By varying their complexity, it becomes possible to manipulate the degree of awareness of the participants (Vandenberghe, Schmidt, Fery, & Cleeremans, 2006). However, despite substantial progress concerning our understanding of the mechanisms of sequence learning, many issues remain open. To allow researchers to address them, we introduce a novel methodology based on a new type of sequential material. We first overview the current state of the literature in the introduction and explain why finding new experimental methods would be timely. Then, after introducing our new sequences, we describe how they contribute to the new methodology. We finally provide a working example in the form of a typical incidental sequence learning experiment and demonstrate through our results how this methodological improvement may open new paths of investigation with this paradigm.

Incidental sequence learning

The incidental sequence learning paradigm dates back to the seminal work of Nissen and Bullemer (1987), who asked participants to track a moving visual target, by pressing on the key corresponding to the specific location at which the target appears on each trial. The specific feature of this serial reaction time (SRT) task is that the target moves from location to location according to a predefined sequence. Unbeknownst to participants, the sequence follows a repeating series, typically either a first-order conditional (FOC) or a second-order conditional (SOC) sequence (Cohen et al., 1990; Curran, 1997; Reed & Johnson, 1994), in which it is necessary to consider the previous location (FOC) or the previous two locations (SOC) to predict the next one. Although participants are neither told about the existence of such a structure nor instructed to look for it, results in SRT tasks show that sequence learning still occurs under such incidental exposure, as the reaction times (RTs) of participants decrease significantly with training (Reed & Johnson, 1994). Moreover, the unexpected insertion of a new sequence or of a random sequence after a period of training causes an abrupt increase in RTs (i.e., a transfer effect), indicating that participants have become sensitive to the specific regularities of the training sequence. After this indirect learning phase, direct tests of sequence knowledge are typically administered. In such tests, participants are explicitly instructed to use their knowledge of the sequential regularities. One of those direct tests is the free generation task (Perruchet & Amorim, 1992; Shanks & Johnstone, 1999), in which participants are asked to reproduce the practice sequence repeatedly to the best of their ability. Another test is the fragment-completion task (Norman, Price, Duff, & Mentzoni, 2007; Stefaniak, Willems, Adam, & Meulemans, 2008; Wilkinson & Shanks, 2004), where participants are presented with small chunks of the trained sequence and asked to predict the subsequent location.

In 2001, Destrebecqz and Cleeremans, taking inspiration from Jacoby’s process dissociation procedure (PDP) in the implicit memory literature (Jacoby, 1991), suggested a new type of free generation task in which participants are asked to perform two free generation tasks under very different instructions. In the inclusion task, participants are told to reproduce the practice sequence as best as they can. However, in the exclusion task, they have to produce another—yet similar—sequence that avoids the transitions of the practice sequence. In most cases, subjects perform above chance in the inclusion task, showing that they effectively acquired knowledge about the sensorimotor sequence. Nevertheless, they also often fail to avoid producing the practice sequence when responding to the exclusion instruction, suggesting that the expression of this knowledge falls outside of their intentional control. Hence, the new generation task clearly confirms the implicit nature of the acquired knowledge in SRT tasks, which is why this procedure has been widely used since the publication of Destrebecqz and Cleeremans’s article (Abrahamse, Lubbe, Verwey, Szumska, & Jaskowski, 2012; Dennis, Howard, & Howard, 2006; Deroost & Coomans, 2018; Ferdinand, Mecklinger, & Kray, 2008; Fu, Bin, Dienes, Fu, & Gao, 2013; Fu, Fu, & Dienes, 2008; Gheysen, Gevers, De Schutter, Van Waelvelde, & Fias, 2009; Gheysen, Van Opstal, Roggeman, Van Waelvelde, & Fias, 2010; Goschke & Bolte, 2012; Guzmán Muñoz, 2018; Haider, Eichler, & Lange, 2011; Mong, McCabe, & Clegg, 2012; Sævland & Norman, 2016; Shanks, Rowland, & Ranger, 2005; Viczko, Sergeeva, Ray, Owen, & Fogel, 2018; Wilkinson & Shanks, 2004).

Even so, the new generation task is not as solid as it appears, as many concerns have been expressed regarding the challenges its application involves. Under inclusion, instructions (i.e., “Generate the same sequence that you have practised in the previous phase; simply follow your intuition if you cannot remember”) are straightforward, as much as the task allows clear criteria on which to assess the accuracy of the participants’ responses (by counting the number of generated triplets—succession of three elements—that belong to the practice sequence, for instance). However, this is not the case for the exclusion task. First, participants are often unable to immediately grasp the meaning of instructions such as “Refrain from generating the sequence, even though you may not remember it.” Second, participants may opt to adopt strategies that comply with the instructions yet fail to produce useable data. For instance, merely repeating the same response all the time clearly produces a sequence that is different from the trained material, yet it leaves no room to find out whether people indeed tend to produce familiar responses despite being instructed to avoid doing so. This has prompted many researchers to augment the exclusion instructions with further qualifications such as “do not continuously produce the same response; avoid reversals, repetitions, ascending and descending runs, etc.,” which often leaves participants embrangled in a lengthy list of cautionary dos and don’ts. As it happens, authors frequently have to exclude parts of the results from their analyses for such instructions, additional recommendations included, failed to be properly followed by some participants (Dennis et al., 2006; Fu et al., 2008; Shanks et al., 2005). Finally, the procedure, because of the complexity of its instructions, appears unsuitable for use with special populations, such as amnesic patients, children, and elderly people (Graf & Komatsu, 1994; but see Gaillard, Destrebecqz, Michiels, & Cleeremans, 2009, for a successful application to older adults).

The new methodology that we will introduce later in this article was designed to remedy this first problem. Furthermore, it also addresses a second issue, which concerns the experimenter’s inability to properly control and measure the influences of implicit and explicit processes in incidental sequence learning tasks, as described in more details below.

Implicit and explicit processes

Implicit learning has also been demonstrated in many other domains, such as language acquisition (see Cleeremans et al., 1998, for a review). In a seminal article, Miller (1958) showed that subjects memorise grammatical strings—that is, strings complying with an artificial grammar (AG) defined in terms of a set of transition probabilities between letters—better than random strings of letters. According to Miller, encoding of information is carried out explicitly in short-term memory by information recoding and regrouping in chunks such that, once fully stored, such chunks may serve as a set of probability rules to assess the grammaticality of newly presented test instances—but see also Servan-Schreiber and Anderson (1990) and Perruchet and Pacton (2006), who suggested that chunking might be an implicit process instead. Reber (1967) favoured another view. He argued that if transition probabilities were indeed encoded as a set of explicit rules in memory, then subjects should be able to report these rules verbally. But his results instead showed that subjects could discriminate test instances above chance while they remained unable to verbalise the transition rules underlying the grammar. Thus, these and other findings (Cleeremans, 1993; Hasher & Zacks, 1984; Jiménez, Méndez, Pasquali, Abrahamse, & Verwey, 2011; Saffran et al., 1997) rather suggest that the learning process supporting a chunking of the AG rules, which these authors deem explicit indeed, might not be the only mechanism in play during such a task, as the statistical structure of the AG also appears to be encoded in memory incidentally. Such an implicit process, in contrast, would operate on an instance-per-instance basis and make it possible to convey a sense of familiarity with the structure, yet no proper recollection. Other authors (Dominey, Lelekov, Ventre-Dominey, & Jeannerod, 1998; Marcus, Vijayan, Bandi Rao, & Vishton, 1999) described a similar dichotomy in our learning abilities in such an experimental context, considering a first, implicit and automatic mechanism, responsible for learning the specific transitions embedded within each instance of the grammar—as a process of statistical learning would account for—and a second, explicit and effortful mechanism, which extracts the probability rules (also mentioned as abstract relationships) that commonly exist between grammar elements in all instances—hence relatable to a process of chunk learning.

Such a debate naturally extended to implicit learning in sensorimotor situations and thus elicited further investigations (Stadler, 1993; Verwey & Eikelboom, 2003). Over the years, authors have used many types of sequences to explore such questions. In a nutshell, material falls into different categories defined by the complexity of the sequence—that is, the various probabilities by which each element predicts the subsequent occurrence of another element. For instance, with the FOC sequence DBCACBDCBA, D appears twice per sequence loop and predicts B or C with 50% probability, whereas B appears three times and predicts A, C, or D with 33% probability. These discrepancies in element frequencies and transition probabilities make it easier for the participant to become aware of a sequence in the material. In comparison, SOC sequences are harder to notice for they require that the series could not be completely learned on the basis of first-order associations—notably, by having each element appearing equally often—or on the basis of the relative frequency of the sequence elements—notably, by having all first-order transitions equally likely (Curran, 1997). According to this definition, SOC sequences are said to be statistically homogeneous. This explains why SOC sequences, rather than FOC sequences, have become a standard to investigate implicit processes in incidental sequence learning tasks (Jiménez, 2008).

Interestingly though, it was proposed that chunks could also be induced implicitly, either by varying the response-stimulus interval (RSI) to form specific rhythms in the RTs (Stadler, 1995) or by breaking the homogeneity of the sequence material, that is, by introducing particularly salient runs—such as ascending and descending series, or mirrored groups, like in ACBBCA (Koch & Hoffmann, 2000). Such methods appeared controversial though (Jiménez, 2008; Vaquero, Jiménez, & Lupiáñez, 2006), for the patterns they give rise to cannot be exclusively attributed to a single process—implicit and explicit processes congruently contributing to the same effects in such tasks. Because no experimental task can be taken as being process pure (Destrebecqz & Cleeremans, 2003; Jacoby, 1991; Reingold & Merikle, 1988), one important condition to distinguish between learning produced by a chunking process and that derived from learning statistical correlations would therefore be to select structures in which the contributions of both mechanisms could be clearly differentiated (Jiménez, Méndez, & Cleeremans, 1996; Jiménez et al., 2011; Perruchet & Amorim, 1992; Perruchet & Gallego, 1993; Sakai, Hikosaka, & Nakamura, 2004). Notably, the 4-item SOC sequences of 12 elements—although commonly used in this paradigm—is not fit for this purpose. Taking 121342314324 (S1) and 123413214243 (S2) (Reed & Johnson, 1994) as examples, both SOC sequences present many saliencies among the transitions of their elements. In particular, participants may be quicker for transitions belonging to reversals (121 in S1; 424 in S2) and to ascending/descending runs (432 in S1; 1234 and 321 in S2), not only because they are attention-grabbing (such saliencies could serve as explicit, abstract cues for the participants, who may thus tend to segment the sequence and form chunks in accordance with them) but also because their configurations elicit keystroke facilitations naturally (Jiménez, 2008). As a matter of fact, every 4-item SOC sequence of 12 elements that can be built contains such salient features (reversals and ascending/descending triplets simply cannot be avoided), and we therefore wondered what better material could be used in a context of sequence learning.

Precisely, Jiménez et al. (2011), in search of a homogeneous material for their experiment, investigated the use of yet another type of SOC sequences. One of the characteristics of these new sequences is that as they do not contain any transition in common with their reverse, authors could train participants on one instance of such sequences in SRT task and thereafter ask them to generate the reverse sequence in exclusion of the direct test (see how this solves our first problem in section “Reversible second-order conditional sequences”). They notably give the example of the sequence 254613524163 (S3), for which, as is already the case with any SOC sequence, all possible elements and transitions are equally likely—here, each element appears only twice and predicts only two possible elements with the same probability (1 predicts 3 and 6 with 50% probability, 2 predicts 4 and 5, etc.). However, in addition to ascribed SOC features, S3 does not contain any reversal or ascending/descending run. Its elements and transitions are therefore neither attention-grabbing nor facilitating, which constitutes a significant reduction in their relative saliency. For this reason, authors assume that on this material, (a) a continuous process of statistical learning would tend to produce a rather homogeneous improvement of performance over the whole sequence, while (b) segments created across chunk learners would remain arbitrary, depending only on random or idiosyncratic factors. Interestingly, Jiménez and colleagues acknowledge that, under such circumstances, such segmentation effects would cancel out among a group of learners—because they would all learn different chunks—thus rendering their effects undetectable on the average measures of RTs. Hence, they instead suggest observing this discontinuous performance through the measures of variance of the individual RTs, which would tend to increase significantly when a participant starts learning selectively about certain segments of the sequence (see section “Method” for further details). Such variance measures could, according to their theory, reflect the amount of chunk learning taking place at any given block, a baseline being established when statistical learning occurs alone (i.e., residual variance caused by any other factor, such as practice or fatigue).

Therefore, the second goal of this study is to design and test a new type of homogeneous material, that is, a structure in which all the legal elements and first-order transitions are equally likely—such as in SOC sequences—and in addition for our purposes, neither are attention-grabbing nor elicit keystroke facilitation—hence presenting no or small difference in their relative saliency (Jiménez et al., 2011). With such an experimental material, any observed discontinuity in the effects of sequence learning could be safely attributed to chunk learning rather than to any other factor, such as an incidental extraction of statistical irregularities in the material.

Reversible second-order conditional sequences

All in all, SOC sequences appear to have stronger homogeneity than FOC sequences and that is why they have been more widely used since 1994. Nevertheless, newly raised methodological concerns led us to believe that statistical homogeneity remains insufficient and that abstract homogeneity—a new term we will properly define further below—is warranted as well when investigating the difference of influences between implicit and explicit processes. We therefore looked for a better structure and hereby formally introduce the reversible second-order conditional (RSOC) sequence, defined such that (a) an RSOC is a SOC, and (b) an RSOC contains no common transition with its own reverse—that is, the mirrored RSOC created by reversing the reading direction. For example, both 124132546356 and its reverse 653645231421 are SOC sequences: each element appears with the same frequency—two times per sequence loop—and there are no transitions more likely than others—they all have a 50% probability. In addition, we can verify that no transitions are shared by these two sequences, and therefore, 124132546356 and its reverse are also RSOC sequences. When using RSOCs though, reversals and ascending and descending runs must still be eliminated. The reason behind such a procedure is to get rid of potential abstract cues and keystroke facilitations in the material, for they may favour specific chunking strategies (i.e., some chunks being more likely to be acquired by the participants than others). We believe that only such refined sets of RSOC sequences should be considered as being abstractly homogeneous, that is, here in the context of behavioural experiments (such a refinement might be irrelevant in simulation, for instance), sequences for which it can be argued that “no transition would be responded faster than others in the absence of learning” (cf., Jiménez, 2008, for a method to test this property). After applying such a refinement, we are left with a total of 60,480 six-item RSOCs of 12 elements to choose from (such a structure was explored in Jiménez et al., 2011; see Supplementary Material 1 for an exhaustive list). When the sequence learning task only involves one hand, the same process of refined selection generates a total of 840 five-item RSOCs of 10 elements (such a structure was explored in Schmitz, Pasquali, Cleeremans, & Peigneux, 2013; see Supplementary Material 2 for an exhaustive list). Being simpler, these sequences could also better fit experiments involving special populations.

Properly used, RSOCs provide a straightforward way to assess learning, as the reverse sequence could be used as a control sequence that does not share any single transition with its counterpart but is otherwise fully analogous to it (i.e., training and testing materials share the same non-salient abstract structure). In the SRT task, for instance, they may constitute a better material for testing the transfer effect than random sequences or even standard SOC sequences (see Schmitz et al., 2013, for the first practical use in this context). Indeed, with such sequences (cf., S1 and S2 mentioned in Introduction), participants may still benefit from having practised the FOC transitions (12, 13, 14, 21 . . . 42, 43) during a transfer block, and extensive practice over the material—ensuring that SOC transitions were also learned—is required to elicit any significant effect. Conversely, the presentation of the reverse RSOC in transfer ensures that every bit of information acquired during the training phase becomes useless in transfer, the participant being exposed to an entirely new sequence (surely one could argue that the absence of saliency in both sequences constitutes information per se, but the whole point of any experiment is precisely to cancel out any factor that is not under study). In that case, transfer effects could be tested earlier and would be stronger in general.

Crucially, the use of RSOC sequences also substantially simplifies the generation task. Indeed, instead of asking participants in exclusion to “Refrain from generating the practice sequence,” the experimenter may now ask to “Generate the practice sequence, but in reverse” (see Jiménez et al., 2011; Schmitz et al., 2013, for the first practical uses of this new instruction). Importantly, this new exclusion task carries the same assumption as the previous one, that is, implicit and explicit influences diverge, but the task has become more comparable to the inclusion task, for there is—now as well—only one (obvious) solution to the task. It is henceforth unnecessary to ask participants to avoid reversal transitions, repetitions, or ascending/descending runs because these can be attributed neither to explicit nor to implicit influences. Therefore, whereas it keeps its initial purpose, the new instruction in exclusion is not only easier for participants to understand and follow, it also offers a clearer interpretation of the results to the experimenter—please refer to section “Results” to see how the PDP (Destrebecqz & Cleeremans, 2001) can be extended to this new situation.

Experiment

Method

We carried out our experiment following the procedure of a typical SRT task, that is, a task that includes a transfer block (indirect test) and is followed by a generation task (direct test). The goal of this experiment is twofold. First, we aim to validate the new methodology for both tests, which will rely on the reverse RSOC sequence in transfer and on the new instruction in exclusion. Second, we want to show how statistically and abstractly homogeneous material—specifically, RSOCs—may allow identification of the chunking process during learning. According to the theory by Jiménez et al. (2011), when all legal transitions of a material are equally predictable and their relative saliency reduced, it becomes possible to dissociate the effects of chunk learning from that of statistical learning by comparing the variance of the individual RTs for each transition of the material on each block. In a hypothetical case of pure statistical learning, all transitions—due to their homogeneity—should be learned uniformly, at the same speed. Hence, RTs would be about the same for all transitions at any given block and would evenly decrease from block to block. Accordingly, the variance of the RTs would remain low for all blocks. Contrastively, under a chunking process, segmenting the material should induce differences in the RTs associated with each transition. Indeed, as the material gets progressively memorised in chunks, participants would respond faster to transitions within chunks (transitions being memorised) than between chunks (transitions not yet memorised). Notably, encoding in chunks may involve hierarchical representations in memory, such that further retrieval of the successive elements within those chunks be facilitated by such an organisation of knowledge. Thus, within a same block, RTs should vary depending on whether they relate to a transition within or between chunks. Interestingly, participants are likely to start by memorising small, first-order chunks—eliciting a bit more variance than a statistical process—then concatenate them into larger, second-order chunks throughout exposure—eliciting even more variance—to finally remember the whole sequence as just one higher order chunk repeating itself when knowledge will get fully automatised—with a variance back to that of a statistical process. In practice, the variance of the RTs can be understood as an index of the ratio between the number of transitions between chunks and that of transitions within chunks. The closer this ratio is to 50%, the higher the variance. This is because in such a case there are as many slow RTs as there are fast RTs. To summarise, the theory predicts that the variance measure (referred to as Variance_total in Jiménez et al., 2011) should remain roughly flat across blocks for pure statistical learners (at a level of residual variance), but should depict an inverted U-shape for pure chunk learners (rising above the level of residual variance).

However, no task being ever purely implicit or explicit (Destrebecqz & Cleeremans, 2003; Jacoby, 1991; Reingold & Merikle, 1988), chunk and statistical learning would always occur concomitantly. Therefore, we may only approximate these learning conditions experimentally by modulating, for instance, the instructions given to the participants. Resulting from an implicit process of statistical learning, all participants should automatically learn about statistical information to a similar extent during the task. Proper instructions, however, should influence the extent of chunk learning that, contrastively, depends on explicit and effortful mechanisms. We will therefore differentiate two groups of learners in our experiment, hereby defined in terms of either incidental or intentional learning conditions. Specifically, task instructions will ensure that the former group gets much fewer opportunities to learn chunks than the latter group. Within such a framework, although a certain amount of variance is indeed expected for both groups (let us call them V₁ and V₂, respectively)—a same baseline due to residual factors in statistical learning (V_S, of same order for both groups), plus varying amounts of variance due to the difference in chunk learning (V_C1 and V_C2)—,any discrepancy found between the variance measures of the two groups can safely be attributed to an excess of chunk learning (considering V₁ = V_S + V_C1 and V₂ = V_S + V_C2, we have V₂ – V₁ = V_C2 – V_C1).

Design

In total 62 adults (18-30 years old; age and gender balanced among experimental groups) participated in the experiment, in exchange for either €6 or course credits. The SRT task involved six possible locations for the stimulus to appear on the screen, and thus six corresponding keys on the keyboard, on which the participant had to press to trigger a change in stimulus location. An RSOC sequence (463124165325) was chosen as the practice sequence for a half of the participants, whereas the reverse RSOC (523561421364) was assigned to the other half. The whole group was divided again so that one experimental group practised the SRT task under intentional learning condition and the other under incidental learning condition. Instructions for the incidental learning condition only described the task as a speed contest (Incidental Condition). Participants in the intentional learning condition followed the same instructions but were also asked to memorise the sequence that they were about to learn before the start of the session (Intentional Condition). Successful memorization was checked by the experimenter before the beginning of the training session. Participants in the intentional learning condition were further instructed to make use of their knowledge of the sequence to optimise response speed during the task. All participants then practised on the SRT task over 30 blocks. All blocks, except the 29th, consisted of 100 trials featuring eight repetitions of the practice sequence, which started on each block with the second transition of the preceding block so as to avoid a common starting point on each block that would provide participants in the incidental condition with an extra cue about the structured nature of the material. Block 29 was arranged as the transfer block and consisted of eight repetitions of the reverse of the practice sequence. A constant RSI of 250 ms was maintained over the whole task. The first four trials of each block were removed from the results, thus leaving eight full sequences per block for further analysis. We also replaced the RT of each trial where the participant either made an error or was correct but took more than 1 s to respond (as it was often the case after an error, for instance), by the mean of the RTs for the block.

After training on the SRT task, awareness was assessed through an unexpected generation task. In inclusion, participants were asked to generate the practice sequence over 100 trials, from which we only analysed the last 94 triplets. Then, the old and new instructions in exclusion were tested such that only half of the participants were told to produce the practice sequence in reverse (i.e., New Generation), whereas the other half had to refrain from producing the practice sequence—avoiding reversals, repetitions, and ascending/descending runs as well (i.e., Old Generation). In all generation tasks, participants were advised to take some time to visualise and practise the sequence that they were about to generate before starting the subsequent block.

Results

We compared RTs (Figure 1) during training within a 2 (Incidental and Intentional Conditions) × 30 (Blocks) analysis of variance (ANOVA) and measured the transfer effect with a 2 (Incidental and Intentional Conditions) × 2 (average of the 28th and 30th blocks, and the 29th block) ANOVA. As expected, measures of RTs revealed strong effects of Blocks during both training, F(29, 1740) = 125.24, p < .001, $η_{p}^{2} = . 67$ , and transfer, F(1, 60) = 317.18, p < .001, $η_{p}^{2} = . 84$ (M_29-28&30 = 221.29 ms, SD = 100.87 ms). Although Conditions alone appeared significant neither during training, F(1, 60) = 2.19, p = .144, nor transfer, F < 1, we nonetheless observed interactions between Conditions and Blocks in training, F(29, 1740) = 1.90, p < .005, $η_{p}^{2} = . 031$ , and in transfer, F(1, 60) = 4.85, p < .05, $η_{p}^{2} = . 075$ (Incidental: M_29-28&30 = 193.93 ms, SD = 78.72 ms; Intentional: M_29-28&30 = 248.64 ms, SD = 113.78 ms). We also tested the effectiveness of having the reverse RSOC in place of transfer sequence by submitting the RTs of the first block, M = 499.33 ms, SD = 11.21 ms, and the 29th block, M = 532.23 ms, SD = 9.40 ms, to a similar analysis. Although neither an effect of Conditions nor an interaction with Blocks, F(1, 60) < 1 for both, was observed, a strong effect of Blocks, F(1, 60) = 14.574, p < .001, $η_{p}^{2} = . 195$ (M_29-1 = 32.90 ms, SD = 67.63 ms), indicated that the insertion of the reverse RSOC caused performance to drop even deeper than when participants had started the task without any a priori knowledge—suggesting that at least this portion of the transfer effect is in fact due to the element of surprise.

Figure 1.

Reaction times per block during the SRT task (error bars represent the standard error of the mean).

In general, error rates from the fourth to the 28th block remained low (Incidental: M = 3.51%, SD = 1.81%; Intentional: M = 4.35%, SD = 2.05%) and constant across Blocks and Conditions, Blocks: F(24, 1440) = 0.652, p = .899; Conditions: F(1, 60) = 2.908, p = .093; Blocks ×Conditions: F(24, 1440) = 1.016, p = .441 —slightly lower errors were observed during the first three blocks and almost twice as many errors during the transfer block for both conditions, such blocks and the last being removed from the above-mentioned analysis to demonstrate the absence of speed–accuracy trade-off or of influence of the instructions on the error rates during the main part of the task. During transfer, a 2 (Incidental and Intentional Conditions) × 2 (average of the 28th and 30th blocks, and the 29th block) ANOVA revealed the strong effect of Blocks mentioned above, F(1, 60) = 59.86, p < .001, $η_{p}^{2} = . 499$ , but only a marginal effect of Conditions, F(1, 60) = 3.64, p = .061, $η_{p}^{2} = . 057$ , and no interaction, F(1, 60) = 0.109, p = .742. Error rates indeed increased significantly but similarly for both Conditions during transfer (Incidental: M_28&30 = 2.52%, SD_28&30 = 0.30%, M₂₉ = 6.42 %, SD₂₉ = 0.66%; Intentional: M = M_28&30 = 4.05%, SD_28&30 = 0.39%, M₂₉ = 7.63%, SD₂₉ = 0.91%), suggesting that the insertion of the reverse RSOC in transfer induced participants in both Conditions to make more errors to a similar extent.

One might conclude from these different results that our instructions did not have a strong impact on learning, eliciting stronger effects of Blocks during training and in transfer for participants in the Intentional Condition but no effect of the Conditions per se. However, an inspection of the RT variance suggests a different conclusion (Figure 2). Analogous ANOVAs conducted on the variance measures during training revealed a strong effect of Blocks, F(29, 1740) = 7.59, p < .001, $η_{p}^{2} = . 112$ , and this time, a significant effect of Conditions as well, F(1, 60) = 6.60, p < .05, $η_{p}^{2} = . 099$ , despite no interaction, F(29, 1740) = 1.13, p = .288. The effect of Blocks in transfer was also significant, F(1, 60) = 17.59, p < .001, $η_{p}^{2} = . 227$ (M_29-28&30 = −5.19 × 10³ ms², SD = 9.76 × 10³ ms²), but neither an effect of the Conditions, F(1, 60) = 1.94, p = .169, nor an interaction, F(1, 60) = 1.19, p = .280, could be captured.

Figure 2.

Total variance per block during the SRT task (error bars represent the standard error of the mean).

Interestingly, variance measures showed an initial increase over the first blocks of training: an ANOVA conducted over the RT variance of the first six blocks captured an effect of Blocks, F(5, 300) = 15.445, p < .001, $η_{p}^{2} = . 205$ , of Conditions, F(1, 60) = 4.519, p < .05, $η_{p}^{2} = . 070$ , and an interaction between Blocks and Conditions, F(1, 60) = 4.268, p < .05, $η_{p}^{2} = . 066$ . The fact that this increase appears to be much stronger in the Intentional Condition (M_6-1 = 8.59 × 10³ ms², SD = 10.72 × 10³ ms²) than in the Incidental one (M_6-1 = 4.80 × 10³ ms², SD = 3.28 × 10³ ms²) is in line with our hypothesis that participants in the former group would acquire knowledge in chunks more than those in the latter. Notably, their variance measures seem to hit a ceiling at Block 14 and, after that, to cap and even start a slow decrease. As Jiménez et al. (2011) have argued, the hit should correspond to the reach of a bimodal distribution in which there would roughly be, in average of all participants in the Intentional Condition, as many elements of the sequence that correspond to transitions between chunks as to transitions within chunks. We may then assume that as the chunks slowly grow bigger and bigger, a decrease in variance measures may later on occur, which the figure only suggests. Nonetheless, for we did not attempt to induce any particular segmentation of the material, each chunk learner could segment the practice sequence into a different set of chunks to commit it to memory. Hence, we have no mean to push the group analyses further here and rather leave the door open for future investigations with the assurance that refined RSOCs are sufficiently homogeneous to allow dissociation of influences from the chunking process during learning.

Let us now focus on the results of the generation task (Figure 3). First, to confirm that our new instruction in exclusion of the generation did not affect the sensitivity of the PDP, we carried out a 2 (Old and New Generations) ×2 (Incidental and Intentional Conditions) × 2 (Inclusion and Exclusion Tasks) ANOVA restricted to the number of generated triplets from the training sequence—participants exposed to the Old Generation were not susceptible to produce triplets from the reverse sequence in the first place. Analyses show that, as expected in a PDP, an effect of Tasks, F(1, 58) = 89.65, p < .001, $η_{p}^{2} = . 607$ , an effect of Conditions, F(1, 58) = 7.455, p < .01, $η_{p}^{2} = . 114$ , and an interaction between these factors is observed, F(1, 58) = 11.505, p = .001, $η_{p}^{2} = . 166$ . However, neither the effect of Generations, F(1, 58) = 1.993, p = .163, nor any interaction with other factors appears to be significant, all F(1, 58) < 3, p > .1. This confirms that our New Generation did not impede the PDP in any way.

Figure 3.

Generated triplets per sequence type during the Generation task (error bars represent the standard error of the mean).

We may hence proceed with a more detailed analysis. In inclusion, both the incidental, M = 42.04%, SD = 32.12%, and the intentional, M = 68.08%, SD = 34.38%, groups were able to generate more triplets of the practice sequence than those expected by chance, respectively, t(30) = 5.55, p < .001, and t(30) = 9.41, p < .001. The chance level was computed as the number of triplets in the sequence divided by the total number of possible triplets, that is, 12/(6 ×5 × 4) = 10%. Nonetheless, participants in the Intentional Condition produced more triplets from the practice sequence than participants in the Incidental Condition, as revealed by a between-subjects ANOVA, F(1, 60) = 9.50, p < .005, $η_{p}^{2} = . 137$ , thus reflecting their stronger knowledge of the sequence. In exclusion, only the incidental group, M = 14.52%, SD = 8.17%, continues to produce triplets of the practice sequence above chance, t(30) = 3.08, p < .005, whereas the intentional group, M = 10.54%, SD = 6.22%, is able to refrain this behaviour, t(30) = 0.48, p = .635. This group difference is further confirmed in a between-subjects ANOVA, F(1, 60) = 4.66, p < .05, $η_{p}^{2} = . 072$ .

Given that our results are perfectly in line with our expectations, we can now suggest a methodological extension of the PDP in this context. This time, we will restrict analyses to the results of participants exposed to the New Generation. Thus, we conducted a 2 (Training and Reverse Sequences) × 2 (Incidental and Intentional Conditions) × 2 (Inclusion and Exclusion Tasks) ANOVA. Our motivation here is to apply the PDP to both the Training and the Reverse Sequences at the same time and thus contrast their analyses. A successful result is obtained if a three-way interaction is revealed, for it confirms that the Training and Reverse Sequences elicit reverse interactions on the PDP. We effectively obtained such a three-way interaction, F(1, 58) = 8.69, p < .01, $η_{p}^{2} = . 130$ . The ANOVA also confirmed effects of Sequences, F(1, 58) = 27.43, p < .001, $η_{p}^{2} = . 321$ , of Conditions, F(1, 58) = 5.16, p < .05, $η_{p}^{2} = . 082$ , and of Tasks, F(1, 58) = 5.00, p < .05, $η_{p}^{2} = . 079$ , as well as an interaction between Sequences and Tasks, F(1, 58) = 42.45, p < .001, $η_{p}^{2} = . 423$ . Other interactions, however, were not significant, F(1, 58) < 1 for both Sequences ×Conditions and Conditions × Tasks. Faithful to our approach, we should then apply the PDP to the results on the reverse sequence—still focusing only on participants who performed the New Generation—to confirm an interaction in a 2 (Incidental and Intentional Conditions) × 2 (Inclusion and Exclusion Tasks) ANOVA. The analysis revealed an effect of Tasks, F(1, 29) = 16.08, p < .001, $η_{p}^{2} = . 357$ , and the expected interaction with Conditions, F(1, 29) = 5.17, p < .05, $η_{p}^{2} = . 151$ , but this time no effect of Conditions taken alone, F(1, 29) = 3.36, p = .077 (a result that does not affect our interpretations).

We may then deepen the analyses again, now keeping in mind that recursively focusing on each subgroup will limit statistical power even further. Specifically in exclusion, participants in the Intentional Condition, M = 29.45%, SD = 27.53%, were able to generate more triplets from the reverse sequence than expected by chance, t(15) = 2.83, p < .05, but participants in the Incidental Condition, M = 13.26%, SD = 11.63%, were not able to mentally rehearse and reconstruct the sequence so as to generate the reverse transitions above chance, t(14) = 1.09, p = .296. This difference in the scores of the two groups in reverse generation is also revealed by a between-subjects ANOVA, F(1, 29) = 4.44, p < .05, $η_{p}^{2} = . 133$ . Interestingly, the size of this effect is higher than that obtained when we were comparing the two groups on their generation of triplets from the training sequence in exclusion (effect size of $η_{p}^{2} = . 072$ ). Thus, these last results make even clearer, despite the loss of statistical power, that the knowledge acquired by the participants in the Intentional Condition was more explicit than that of the participants in the Incidental Condition. This strongly supports our claim that the new instruction in exclusion of the generation constitutes a significant methodological improvement.

Discussion

Through this study, we attempted to overcome two inherent limitations of current methodology in incidental sequence learning. The first relates to the free generation task introduced in 2001 by Destrebecqz and Cleeremans. Although widely employed, this technique relies on complicated instructions in exclusion and therefore presents a great deal of difficulty for many participants and experimenters (Dennis et al., 2006; Fu et al., 2008; Shanks et al., 2005; see also Graf & Komatsu, 1994). Here, we aimed to introduce a methodological improvement that substantially simplifies the exclusion task. Second, we observed that the material used in this paradigm—whether it consists of pseudo-random, FOC, SOC, or ad hoc sequences—does not seem to allow experimenters to unequivocally distinguish between implicit and explicit influences during training. A solution to this issue would be timely, for the debate about a possible dissociation between processes operating statistical and chunk learning still remains after more than five decades of research in this field (to name just a few protagonists: Cleeremans, 1993; Jiménez, 2008; Jiménez et al., 1996; Miller, 1958; Nissen & Bullemer, 1987; Perruchet & Amorim, 1992; Perruchet & Pacton, 2006; Reber, 1967; Saffran et al., 1997; Stadler, 1993; Vaquero et al., 2006; Verwey & Eikelboom, 2003).

Thus, we designed a new type of sequences, labelled RSOC sequences and characterised by the fact that they are SOC sequences that do not share any transitions with their reverse. This new type of material allowed us to overcome the two limitations mentioned above. Indeed, solving the first limitation, the experimenter may now provide participants with simpler instruction in the exclusion task while keeping intact the original principle and sensitivity of the PDP (Destrebecqz & Cleeremans, 2001; Jacoby, 1991). We believe that the new instruction, “Generate the practice sequence, but in reverse,” (a) can now be understood by a much larger population; (b) is less subject to the interpretation of the participant; (c) renders the exclusion task more similar to the inclusion task, given that there is only one obvious solution to either; and (d) can lead to stronger effects and thus clearer evidence for the experimenter.

Regarding the second limitation, we first note that the reverse sequence may be used as a control sequence, for instance in a block of transfer, so that the transfer effect would represent the full amount of acquired knowledge and not only, as is the case with other types of sequences, a portion of it. Second, because we have defined a refinement policy (which, in addition to the already established statistical homogeneity of SOC sequences, aims at further reducing the presence of abstract cues in the sequential material, hence allowing a kind of abstract homogeneity as well), RSOCs may be used to create a material that is suitable for investigating dissociations between statistical and chunk learning processes. Notably, as a statistical process would, in theory, acquire knowledge about each transition of such a material at the same speed, any difference in learning speed across transitions can be accounted to the exclusive influences of a chunking process.

Experimentally, we applied the new methodology in the context of an SRT task (indirect test) followed by an unexpected generation task (direct test) presenting the new instruction in exclusion. For the purpose of contrasting implicit and explicit influences, we differentiated two conditions, with participants being either (a) incidentally exposed to the task or (b) knowledgeable about the existence of the sequence and therefore intentional. Considering the indirect test, on one hand, we effectively found a very strong drop in performance during transfer for both groups, here stronger in range than when having started the task per se—a difference accountable to the violation of expectations, may they be implicit, explicit, or both here, and thus conceivably slightly stronger for participants in the intentional condition. In addition, we tested Jiménez et al.’s (2011) hypothesis that variance measures during training would be larger for chunk learners. In the intentional condition, participants effectively produced RTs characterised by far greater variance than in the incidental condition, with such a difference reflecting a higher disparity in the learning speeds of the transitions of the sequence and, therefore, an exclusive influence of the chunking process (although some part of this influence is already captured in the measures of the participants in the incidental condition—who may have relied on chunk learning to a lesser extent—along with other influential factors constituting a baseline of residual variance). Then, on the other hand, participants’ score on the direct test was consistent with our expectations (i.e., better performance in both inclusion and exclusion for the intentional group) and confirmed that the PDP (Destrebecqz & Cleeremans, 2001) works better in this new version. Notably, participants in the intentional condition could mentally reverse the sequence and generate the new transitions above chance in exclusion, whereas the incidental group was not able to do so and instead continued to generate the practice sequence.

We believe that these three results offer additional evidence in support of Jiménez et al.’s (2011) hypothesis, notably on the idea that chunking and statistical learning are dissociable learning processes—respectively explicit and implicit—running concomitantly in the brain. We would extend it as such: On the one hand, the chunking process would rely on a rule-based encoding of the material’s regularities—involving the working memory for building the chunks and the episodic memory (hippocampus) as a first storage space before further distillation back to the cortex could take place. Chunking would allow fast encoding of knowledge—though at the cost of a high cognitive demand—in hierarchical representations that would further facilitate the retrieval (i.e., recollection) of each component of the chunks. Statistical learning, on the other hand, would automatically encode each instance of the material in a much slower fashion. However, it would leave attentional capacities intact, being supported by neural plasticity alone in sensorimotor pathways (cortex, thalamus, striatum, and cerebellum), and would convey a sense of familiarity to any similar instance encountered thereafter. Notably, this type of learning would bear long-lasting effects and, thus, would form the basis of long-term habits.

To conclude, we trust that this experiment effectively validates the RSOC sequences as a methodological improvement in the context of incidental sequence learning tasks. We then strongly suggest balancing different RSOCs to obtain a diversified set of sequences to be presented across subjects, as it would reduce any effect of keystroke facilitation even further and thus provide an experimental material with additional abstract homogeneity. In Supplementary Materials, we provide two exhaustive lists of refined RSOCs, with structures we thought would be the most useful for the field, and a sample code (C++/MFC) for generating all refined RSOCs given a number of elements and a sequence length, thus hopefully prompting researchers in the field to use them.

Supplemental Material

Supplementary Material 1

List of Reversible Socs with 6 elements and of length 12 with no reversal or a/de-scending runs

Supplemental Material

Supplementary Material 2

List of Reversible Socs with 5 elements and of length 10 with no reversal or a/de-scending runs

Supplemental Material

Supplementary Material 3

Sample code (C++/MFC) for generating refined RSOCs given a number of elements and a sequence length

Footnotes

Acknowledgements

We thank Luis Jiménez for his help and guidance on the use of variance measures, as well as Iring Koch and two other anonymous reviewers for their useful comments on the previous versions of the article.

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

This work was supported by the National Fund for Scientific Research (F.R.S.-FNRS, Belgium) under Research Fellowship and by the Japan Society for the Promotion of Science (JSPS, Japan) under Postdoctoral Fellowship to A.P.; A.C. is a Research Director with the F.R.S.-FNRS (Belgium); V.G. was supported by Brussels Institute for Research and Innovation—Innoviris grant BB2B 2012-1-107. This work was also partly supported by BELSPO IAP grant P7/33 and by ERC Advanced Grant RADICAL.

Supplementary materials

Supplementary Material 1. List of reversible socs with 6 elements and of length 12 with no reversal or a/de-scending runs

Supplementary Material 2. List of reversible socs with 5 elements and of length 10 with no reversal or a/de-scending runs

Supplementary Material 3. Sample code (C++/MFC) for generating refined RSOCs given a number of elements and a sequence length

References

Abrahamse

E. L.

Lubbe

R. H. J.

Verwey

W. B.

Szumska

Jaskowski

(2012). Redundant sensory information does not enhance sequence learning in the serial task. Advances in Cognitive Psychology, 8, 109–120. doi:10.5709/acp-0108-y

Cleeremans

(1993). Mechanisms of implicit learning: Connectionist models of sequence processing. Cambridge, MA: The MIT Press.

Cleeremans

Destrebecqz

Boyer

(1998). Implicit learning: News from the front. Trends in Cognitive Science, 2, 406–416. doi: 10.1016/S1364-6613(98)01232-7

Cohen

Ivry

R. I.

Keele

S. W.

(1990). Attention and structure in sequence learning. Journal of Experimental Psychology: Learning, Memory, and Cognition, 16, 17–30.

Curran

(1997). Higher-order associative learning in amnesia: Evidence from the serial reaction time task. Journal of Cognitive Neuroscience, 9, 522–533.

Dennis

N. A.

Howard

J. H.

Howard

D. V.

(2006). Implicit sequence learning without motor sequencing in young and old adults. Experimental Brain Research, 175, 153–164. doi:10.1007/s00221-006-0534-3

Deroost

Coomans

(2018). Is sequence awareness mandatory for perceptual sequence learning: An assessment using a pure perceptual sequence learning design. Acta Psychologica, 183, 58–65. doi:10.1016/j.actpsy.2018.01.002

Destrebecqz

Cleeremans

(2001). Can sequence learning be implicit? New evidence with the process dissociation procedure. Psychonomic Bulletin & Review, 8, 343–350. doi:10.3758/bf03196171

Destrebecqz

Cleeremans

(2003). Temporal effects in sequence learning. In Jiménez

(Ed.), Attention and implicit learning (pp. 181–213). Amsterdam: John Benjamins.

10.

Dominey

P. F.

Lelekov

Ventre-Dominey

Jeannerod

(1998). Dissociable Processes for Learning the surface and abstract structure sensorimotor sequences. Journal of Cognitive Neuroscience, 10, 734–751. doi:10.1162/089892998563130

11.

Ferdinand

N. K.

Mecklinger

Kray

(2008). Error and deviance processing in implicit and explicit sequence learning. Journal of Cognitive Neuroscience, 20, 629–642. doi:10.1162/jocn.2008.20046

12.

Bin

Dienes

Gao

(2013). Learning without consciously knowing: Evidence from event-related potentials in sequence learning. Consciousness and Cognition, 22(1), 22–34. doi:10.1016/j.concog.2012.10.008

13.

Dienes

(2008). Implicit sequence learning and conscious awareness. Consciousness and Cognition, 17, 185–202. doi:10.1016/j.concog.2007.01.007

14.

Gaillard

Destrebecqz

Michiels

Cleeremans

(2009). Effects of age and practice in sequence learning: A graded account of ageing, learning, and control. European Journal of Cognitive Psychology, 21, 255–282. doi:10.1080/09541440802257423

15.

Gheysen

Gevers

De Schutter

Van Waelvelde

Fias

(2009). Disentangling perceptual from motor implicit sequence learning with a serial color-matching task. Experimental Brain Research, 197, 163–174. doi:10.1007/s00221-009-1902-6

16.

Gheysen

Van Opstal

Roggeman

Van Waelvelde

Fias

(2010). Hippocampal contribution to early and later stages of implicit motor sequence learning. Experimental Brain Research, 202, 795–807. doi:10.1007/s00221-010-2186-6

17.

Goschke

Bolte

(2012). On the modularity of implicit sequence learning: Independent acquisition of spatial, symbolic, and manual sequences. Cognitive Psychology, 65, 284–320. doi:10.1016/j.cogpsych.2012.04.002

18.

Graf

Komatsu

(1994). Process dissociation procedure: Handle with caution! European Journal of Cognitive Psychology, 6, 113–129. doi:10.1080/09541449408520139

19.

Guzmán Muñoz

F. J.

(2018). The influence of personality and working memory capacity on implicit learning. Quarterly Journal of Experimental Psychology. Advance online publication. doi:10.1177/1747021817749582

20.

Haider

Eichler

Lange

(2011). An old problem: How can we distinguish between conscious and unconscious knowledge acquired in an implicit learning task. Consciousness and Cognition, 20, 658–672. doi:10.1016/j.concog.2010.10.021

21.

Hasher

Zacks

R. T.

(1984). Automatic processing of fundamental information: The case of frequency of occurrence. American Psychologist, 39, 1372–1388. doi:10.1037/0003-066x.39.12.1372

22.

Jacoby

L. L.

(1991). A process dissociation framework: Separating automatic from intentional uses of memory. Journal of Memory and Language, 30, 513–541. doi:10.1016/0749-596x(91)90025-f

23.

Jiménez

(2008). Taking patterns for chunks: Is there any evidence of chunk learning in continuous serial reaction-time tasks? Psychological Research, 72, 387–396. doi:10.1007/s00426-007-0121-7

24.

Jiménez

Méndez

Pasquali

Abrahamse

Verwey

(2011). Chunking by colors: Assessing discrete learning in a continuous serial reaction-time task. Acta Psychologica, 137, 318–329. doi:10.1016/j.actpsy.2011.03.013

25.

Jiménez

Méndez

Cleeremans

(1996). Comparing direct and indirect measure of sequence learning. Journal of Experimental Psychology: Learning, Memory, and Cognition, 22, 948–969. doi:10.1037/0278-7393.22.4.948

26.

Koch

Hoffmann

(2000). Patterns, chunks, and hierarchies in serial reaction time tasks. Psychological Research, 63, 22–35. doi:10.1007/pl00008165

27.

Marcus

G. F.

Vijayan

Bandi Rao

Vishton

P. M.

(1999). Rule learning by seven-month-old infants. Science, 283, 77–80. doi:10.1126/science.283.5398.77

28.

Miller

G. A.

(1958). Free recall of redundant strings of letters. Journal of Experimental Psychology, 56, 485–491. doi:10.1037/h0044933

29.

Mong

H. M.

McCabe

D. P.

Clegg

B. A.

(2012). Evidence of automatic processing in sequence learning using process-dissociation. Advances in Cognitive Psychology, 8, 98–108.

30.

Nissen

M. J.

Bullemer

(1987). Attentional requirements of learning: Evidence from performance measures. Cognitive Psychology, 19, 1–32. doi:10.1016/0010-0285(87)90002-8

31.

Norman

Price

M. C.

Duff

S. C.

Mentzoni

R. A.

(2007). Gradations of awareness in a modified sequence learning task. Consciousness and Cognition, 16, 809–837. doi:10.1016/j.concog.2007.02.004

32.

Perruchet

Amorim

P. A.

(1992). Conscious knowledge and changes in performance in sequence learning: Evidence against dissociation. Journal of Experimental Psychology: Learning, Memory, and Cognition, 18, 785–800. doi:10.1037/0278-7393.18.4.785

33.

Perruchet

Gallego

(1993). Association between conscious knowledge and performance in normal subjects: Reply to Cohen and Curran (1993) and Willinghan, Greeley and Bardone (1993). Journal of Experimental Psychology: Learning, Memory, and Cognition, 19, 1438–1444. doi:10.1037/0278-7393.19.6.1438

34.

Perruchet

Pacton

(2006). Implicit learning and statistical learning: One phenomenon, two approaches. Trends in Cognitive Sciences, 10, 223–238. doi:10.1016/j.tics.2006.03.006

35.

Reber

A. S.

(1967). Implicit learning of artificial grammars. Journal of Verbal Learning and Verbal Behavior, 6, 855–863. doi:10.1016/s0022-5371(67)80149-x

36.

Reed

Johnson

(1994). Assessing implicit learning with indirect tests: Determining what is learned about sequence structure. Journal of Experimental Psychology: Learning, Memory, and Cognition, 20, 585–594. doi:10.1037/0278-7393.20.3.585

37.

Reingold

E. M.

Merikle

P. M.

(1988). Using direct and indirect measures to study perception without awareness. Perception & Psychophysics, 44, 563–575. doi:10.3758/bf03207490

38.

Sævland

Norman

(2016). Studying different tasks of implicit learning across multiple test sessions conducted on the web. Frontiers in Psychology, 7, 808. doi:10.3389/fpsyg.2016.00808

39.

Saffran

J. R.

Newport

E. L.

Aslin

R. N.

Tunick

R. A.

Barrueco

(1997). Incidental language learning: Listening (and learning) out of the corner of your ear. Psychological Science, 8, 101–105. doi:10.1111/j.1467-9280.1997.tb00690.x

40.

Sakai

Hikosaka

Nakamura

(2004). Emergence of rhythm during motor learning. Trends in Cognitive Sciences, 8, 547–553. doi:10.1016/j.tics.2004.10.005

41.

Schmitz

Pasquali

Cleeremans

Peigneux

(2013). Lateralized implicit sequence learning in uni- and bi-manual conditions. Brain and Cognition, 81, 1–9. doi:10.1016/j.bandc.2012.09.002

42.

Servan-Schreiber

Anderson

J. R.

(1990). Learning artificial grammars with competitive chunking. Journal of Experimental Psychology: Learning, Memory, and Cognition, 16, 592–608. doi:10.1037/0278-7393.16.4.592

43.

Shanks

D. R.

Johnstone

(1999). Evaluating the relationship between explicit and implicit knowledge in a sequential reaction time task. Journal of Experimental Psychology: Learning, Memory, and Cognition, 25, 1435–14551. doi: 10.1037/0278-7393.25.6.1435

44.

Shanks

D. R.

Rowland

L. A.

Ranger

M. S.

(2005). Attentional load and implicit sequence learning. Psychological Research, 69, 369–382. doi:10.1007/s00426-004-0211-8

45.

Stadler

M. A.

(1993). Implicit serial learning: Questions inspired by Hebb (1961). Memory & Cognition, 21, 819–827. doi:10.3758/bf03202749

46.

Stadler

M. A.

(1995). Role of attention in implicit learning. Journal of Experimental Psychology: Learning, Memory, and Cognition, 21, 674–685. doi:10.1037/0278-7393.21.3.674

47.

Stefaniak

Willems

Adam

Meulemans

(2008). What is the impact of the explicit knowledge of sequence regularities on both deterministic and probabilistic serial reaction time task performance? Memory & Cognition, 36, 1283–1298. doi:10.3758/mc.36.7.1283

48.

Vandenberghe

Schmidt

Fery

Cleeremans

(2006). Can amnesic patients learn without awareness?: New evidence comparing deterministic and probabilistic sequence learning. Neuropsychologia, 44, 1629–1641. doi:10.1016/j.neuropsychologia.2006.03.022

49.

Vaquero

J. M. M.

Jiménez

Lupiáñez

(2006). The problem of reversals in assessing sequence learning with serial reaction time tasks. Experimental Brain Research, 175, 97–109. doi:10.1007/s00221-006-0523-6

50.

Verwey

W. B.

Eikelboom

(2003). Evidence for lasting sequence segmentation in the discrete sequence-production task. Journal of Motor Behaviour, 35, 171–181. doi:10.1080/00222890309602131

51.

Viczko

Sergeeva

Ray

L. B.

Owen

A. M.

Fogel

(2018). Does sleep facilitate the consolidation of allocentric or egocentric representations of implicitly learned visual-motor sequence learning? Learning & Memory, 25(2), 67–77. doi:10.1101/lm.044719.116

52.

Wilkinson

Shanks

D. R.

(2004). Intentional control and implicit sequence learning. Journal of Experimental Psychology: Learning, Memory, and Cognition, 30, 354–369. doi:10.1037/0278-7393.30.2.354

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.81 MB

0.01 MB

0.00 MB