Stimulus dependence and cross-modal interference in sequence learning

Transfer in artificial grammar learning

Most of previous transfer studies use the above-described artificial grammar learning paradigm (Reber, 1967). Altmann, Dienes, and Goode (1995) conducted four experiments to show the presence of positive transfer. In their study, participants learned an artificial grammar with stimuli in one modality and were tested in another. Results showed an above-chance performance on the transfer task. While these results seem convincing, Tunney and Altmann (1999) argue that these results are due to a specific cue—that is, the illegal starting element of ungrammatical sentences—and not to transfer. Removal of this specific cue results in the absence of positive transfer. The importance of sequence-initial clusters in transfer was also highlighted by an earlier study by Gomez (1997). While there are some further studies reporting transfer in the AGL task (Hendricks, Conway, & Kellogg, 2013; Tunney & Altmann, 2001; Vokey & Higham, 2005), it is still not completely clear what the requirements are for abstraction and transfer in the AGL task.

The notion of abstraction as well as the transfer phenomenon has been discussed by Redington and Chater (1996). One of their important, but less highlighted, arguments is that all studies consider some trivial abstraction. That is, even the studies arguing for stimulus dependence focus on learning sequences of letters/symbols, and so on, disregard the question of whether it is the letter sequences or patterns of light intensity changes that are learned.

Another crucial transfer-related issue is the question of control conditions. Redington and Chater (1996) argue that above-chance performance does not necessarily require training: Participants are able to identify and utilize cues of grammaticality within the test session. Hence different patterns of results might be interpreted as transfer due to improper use of control measures. In accordance, a number of later studies used two separate analyses to test learning: one to test the deviation from chance, and one to test the deviation from an untrained control group (e.g., Conway & Christiansen, 2005). Similarly, according to Redington and Chater (1996) it is also possible that participants do not learn abstract rules, but map the post-transfer sequences onto pre-transfer sequences during test phase processing. That is, transfer paradigms might not necessarily test abstract learning.

In a different design, Conway and Christiansen (2006) trained participants with a dual AGL task. There were two separate grammars with two distinct sets of stimuli. The two sets of stimuli were from (a) two different modalities, (b) two different dimensions within a modality, or (c) from the same dimension. In the test phase, participants faced grammatical strings from the two grammars, but using only one of the stimulus sets. Results showed that participants were able to learn both grammars if the two vocabularies were from different dimensions or modalities. On the other hand, if the two vocabularies were from the same dimension, participants performed above chance on only one of the grammars. That is, intra-dimensional stimulus sets caused a negative transfer in learning.

Transfer in the serial reaction-time task

A different stream of research uses the serial reaction-time (SRT) task (Nissen & Bullemer, 1987). In this task, participants see a target stimulus appearing in one of four different locations, and their aim is to press the button that corresponds to the given location. Unknown to the participants, the target stimulus follows a repeated order. Response latencies decrease as long as the sequence is present, but as soon as the sequence is removed, RTs immediately increase (Destrebecqz & Cleeremans, 2001). The random block, and hence the test of sequence specific learning, is usually referred to as transfer block (Schwarb & Schumacher, 2012).

The notion of transfer is also used for an effect similar to the above outlined: A structure that has been learned in one setting is carried over to another. The current study focuses on the latter sense. Studies on this latter type of transfer are related to the hypotheses of what exactly is learned throughout the task: whether participants learn to predict the next stimulus (Remillard, 2003), the next response location (Willingham, Wells, Farrell, & Stemwedel, 2000), or the next movement (Deroost, Zeeuws, & Soetens, 2006) that is required to carry out the task (for a review, see Kemény & Lukács, 2011). There are studies showing positive transfer from one hand to the other (Japikse, Negash, Howard, & Howard, 2003), as well as from finger responses to arm responses, or between manual and verbal responses (Keele, Jennings, Jones, Caulton, & Cohen, 1995). In sum, results mostly reported positive transfer.

Previous studies have also addressed interference in the SRT task. There have been quite a few experiments in which a secondary task was introduced to decrease conscious access to the sequential stimuli. Apart from the effect on consciousness, secondary tasks usually impair learning performance. Results differ in whether they propose that the performance decrease is due to interference with learning (Nissen & Bullemer, 1987; Shanks & Channon, 2002) or with the expression of the acquired information (Frensch, Lin, & Buchner, 1998). These studies, however, mainly focused on interference between sequence learning and other domains.

Goedert and Willingham (2002) tested whether learning two sequences interferes over consolidation. Participants were trained on a sequence, A. After an interval of 5 minutes, 1 hour, 5 hours, or 24 hours (only once, participants were randomly assigned to conditions) participants were trained again, but on another sequence, B (sharing no triplets with Sequence A). The consolidation effect was tested 48 hours after the first training with a three-block session with Sequence A again. Results showed both proactive (Sequence A decreases learning of Sequence B) and retroactive (Sequence B negatively affecting knowledge of Sequence A) interference, and neither proactive nor retroactive interference depended on the time elapsed between the first and the second training blocks. This study provides evidence for negative transfer between non-overlapping sequences.

Different methods for assessing transfer in the AGL and SRT tasks

As seen above, positive and negative transfer studies of artificial grammar learning and those of the SRT task are not parallel. In artificial grammar learning, transfer is the abstraction of rules from stimulus sequences and applying those to new stimuli. There is no mapping between stimuli of the pre- and post-transfer phases. Participants receive a training phase with a set of symbols, and are tested on CVC (consonant–vowel–consonant) syllables. To differentiate between “hes, kav, jix, pel” “hes sog pel jix” (taken from Experiment 4, Appendix C in Altmann et al., 1995, p. 912, the former being ungrammatical, the latter being grammatical) they need to correctly map “kav” onto Symbol 7, “jix” onto Symbol 2, and “pel” onto Symbol 3, and not in any other combination (Figure 3, Altmann et al, 1995). As there is no mapping in the task, we cannot expect participants to be able to differentiate between these sequences. This may explain the lack of proper transfer in such a setting, as shown by Tunney and Altmann (1999). In later studies, Gomez, Gerken, and Schvaneveldt (2000) and Tunney and Altmann (2001) also showed that mapping is required for transfer (at least at the level of repetition structure).

Transfer in the SRT task requires preserving at least one domain. The perceptual or conceptual features of the task are preserved after an inter-manual transfer, while the required movements change. Hence pre-transfer and post-transfer stimuli are linked to each other. To Stimulus A participants give Response A (key press with right index finger) before the transfer, and Response B (key press with left little finger) after the transfer. Responses A and B are both linked to the same stimulus; hence there is a transparent mapping between them.

In summary, in previous studies that investigated the nature of the transfer using different kinds of methodology, the mapping was either undetectable (in AGL studies) or manifest (in SRT experiments). Covert and overt mapping between pre- and post-transfer stimuli elicits different questions; hence it is not easy to compare the results of AGL and SRT transfer studies. To this end we aimed to test whether transfer can be observed between learning and test phase if the mapping between the learning and the test phase is consciously unavailable, but encoded in the physical features of the stimuli/responses. Therefore, we designed an experimental protocol, in which we introduced a new, motor-free implicit training procedure (observational perceptual learning), and transfer to an audiomotor sequence learning task in the test phase. We expect that our design gives an answer on whether the representation developed throughout the task is abstract and can be transferred to another modality, or bound to the input stimuli and hence cannot be transferred.

The test phase is identical in all conditions: Participants hear one of four possible CV syllables and have to press the response key that corresponds to the syllable. The response locations (as well as the stimulus identities) follow a 12-element deterministic second order conditional sequence. There are three groups differing in the preceding training phase. In all three groups, participants are exposed to four horizontally aligned circles. One of the circles is coloured, while the rest are empty. The colour can be red or grey, varying randomly, but with only below 15% being grey. The aim of the participants is to count the grey dots. In two conditions, the appearance of the target stimulus (either grey or red) follows a fixed sequence. The sequence is either identical to the one used in the test phase or a different one (the reverse sequence). Note that in the identical condition the spatial organization of the stimuli of the training phase is identical to the spatial organization of the responses of the test phase. The third condition is a control condition in which the appearance of the dots is random. The experiment uses a second control condition where no training phase took place. While conditions differ in the training phase, we examine learning on the test phase. The central questions are (a) whether learning on the test phase is affected by the use of the same or different sequences in the pre- and post-transfer phases; (b) whether the use of different sequences have different effects on learning compared to a baseline; and (c) whether general fatigue plays a role in sequence learning.

Previous studies have shown that visual location sequence learning may take place in the absence of motor responses (Howard, Mutter, & Howard, 1992). Results showed that participants having to respond to all stimuli and participants having to respond to just a few stimuli in each block show a comparable sequence learning performance. Even response latencies in general were very similar. Our question is whether a location sequence learned through observation can be transferred from the visual perceptual domain (that is, from the screen) to the motor domain (to the response buttons). Our hypothesis is that if the representation developed during the observational task is stimulus independent, then sequence knowledge will be better in the test phase for sequences that are identical to the training sequence, than if the training sequence is the reverse of the test sequence.

As reviewed above, previous studies of the artificial grammar learning were based on the assumption that mapping is not required for sequential transfer. If mapping is not required, transfer should take place in both cases, regardless whether the pre- and post-transfer sequences are identical. To test this hypothesis we included a control group without a training phase. Comparing the no-training phase with the two sequential training phases reveals the effect of the sequential spatial pre-training on later auditory sequence learning.

Another possibility is that sequences are bound to the input stimulus. This would suggest that learning operates on superficial, uninterpreted stimuli. In this case one would expect that even sequences with identical structure but different stimuli behave differently. As a result one can expect a negative transfer between sequences, regardless of whether the structure is identical or different. Hence, one can expect decreased sequence learning compared to the control group.

On the other hand, a prediction of a negative effect in longer tasks might always have a simple interpretation of general fatigue. To control for this interpretation we included a second control condition in which participants observed a spatial training phase with no sequential structure. If performance in the latter control condition is identical to that in the experimental groups, we can assume that the decrease in performance is only due to the dot counting task, and is not a sequential effect.

Method

Design and stimuli

The experiment consisted of two main phases: the training and the test phase. Completing the experiment took approximately 40 minutes. Stimuli were displayed, and answers were collected using E-Prime 1.2 (Psychology Software Tools, Inc., Pittsburgh, PA).

Training phase

Participants were randomly divided into three different groups in order to test the cross-modal learning effect—two experimental and the control group—while data from a second control group, the random-dot group, were collected later. Participants in the same, mirror, and random-dot groups (see details below) were trained with a six-block visual sequence observation and dot counting task. The control group had no training phase.

In the training phase, participants saw four circles (diameter: 33 pixels) in the horizontal midline of the screen with 50-pixel gaps between the circles (640 × 480 pixels resolution was used). Three of the circles were empty, while the fourth one was red or grey (Figure 1a). Each stimulus was on screen for 750 ms. Inter-stimulus interval was set to 250 ms. In order to maintain attention, the task of the participants was to count the grey dots. Different colours appeared randomly; the likelihood of the appearance of the grey circle was 15%. The result of counting had to be reported at the end of each block.

OPEN IN VIEWER

Figure 1.

The design of the experiment. Stimuli in (a) the visual training phase, and responses in (b) the auditory test phase have spatial structures. These spatial structures may be mapped onto each other. The figure illustrates the same condition, where the spatial structures of the two tasks are identical.

In the training phase, participants faced six blocks of 120 stimuli with a short self-paced break between blocks. The location of the stimuli was random in the random-dot group, but followed a 12-element sequence in the two experimental groups, the same and the mirror group. In the case of the same group the sequence of the target location—irrespective of their colour—was 121423413243 (where numbers represent locations from left to right—the same sequence was used for all participants in the test phase. The sequence is a second-order conditional deterministic sequence, which is identical to the sequence used in previous studies, like Kemény & Lukács, 2011; or Vaquero, Jiménez, & Lupiáñez, 2006). The mirror group faced the mirror of this sequence (342314324121).

Test phase

Participants of all four groups faced the same test phase, which was an auditory version of the SRT task (Nissen & Bullemer, 1987). They heard one of four different CV syllables: /fi/, /ga/, /lu/, /pe/. CV syllables were 600 ms long (44100 Hz, 16 bit, stereo). The task of the participants was to press the key that corresponds to the stimulus: to press “y” for /fi/, “c” for /ga/, “b” for /lu/, and “m” for /pe/. Participants were explicitly asked to respond as fast and as accurately as possible. Note that we used Hungarian QWERTZ keyboards, where these buttons are all located in the lowest row with one key between each adjacent response button (Figure 1b). Each item was heard only once, and required a response. Upon incorrect response, a warning message (“Wrong answer”) appeared for 500 ms in the middle of the screen. Response-to-stimulus interval was set to 250 ms.

The task consisted of six blocks with 120 button-presses in each block. Blocks 1–4 and Block 6 were sequence blocks, while the order of auditory stimuli was pseudorandom in Block 5. The sequence in the sequential blocks was 12143413243, where 1 is /fi/, 2 is /ga/, 3 is /lu/, and 4 is /pe/. That is, the sequence was identical to the observed sequence in the same condition, and the reverse of the observed sequence in the mirror condition. The important design-element is that the spatial structure of the response sequence (which button was to be pressed) was either identical to or the reverse of the spatial structure of the stimulus sequence in the training phase. Figure 1 illustrates the design. In Block 5 the appearance of the elements was pseudorandom: No consecutive trials were identical. The blocks were separated by short self-paced breaks.

Results

Similar to previous studies (Kemény & Lukács, 2011), practice and sequence-learning data were analysed separately. First we analysed how reaction times decrease due to practice during Blocks 1 to 4 in the auditory test phase. Then we computed a disruption score by extracting the mean of Block 4 and Block 6 medians from Block 5 medians—that is, the median RTs for the random block minus the mean of median RTs of the surrounding sequence blocks. In both analyses, condition was included as a between-subjects variable. As the mean accuracy of participants was above 95% in all conditions, we do not report accuracy measures. However, to avoid any speed–accuracy trade-offs, we excluded all participants with accuracy below 90%. Three participants were excluded from the same, one from the mirror, and three from the control condition. No participants performed below 90% in the random-dot condition.

Practice effect

A 4 × 4 repeated measures analysis of variance (ANOVA) was conducted with Block (1 through 4) as within-subject variable and condition (same vs. mirror vs. random-dot vs. no-training) as between-subjects variable. Huynh–Feldt correction was used due to sphericity violation indicated by a significant Mauchly's test of sphericity. Results revealed a significant main effect of block, F(2.000, 333.925) = 92.659, p < .001, η p 2 = . 357 . Polynomial contrast revealed a monotonous decrease in reaction times, as indicated by a significant linear contrast for block, F(1, 167) = 132.730, p < .001, η p 2 = . 443 . Results also revealed an overall difference between the conditions, F(3, 167) = 5.599, p < .01, η p 2 = . 091 . Post hoc pairwise comparisons showed that the RTs of the random-dot condition were significantly lower than RTs in all other conditions (comparison with the control condition: p < .05; other two conditions: p < .01). There was also a significant Block × Condition interaction, F(5.999, 333.925) = 4.724, p < .001, η p 2 = 0.078 . For mean reaction times, see Figure 2.

OPEN IN VIEWER

Figure 2.

Mean reaction times (RTs) by block and by condition. Error bars indicate standard error.

OPEN IN VIEWER

Figure 3.

Mean reaction times (RTs) by block and by category (sequence present or absent in the training phase). Error bars indicate standard error.

To further evaluate the effect of sequence-based pre-training, we grouped conditions with and without sequential training (sequential-training and no-sequence-training/no-training, respectively). The first category included the same and mirror conditions, while the second was composed of the random-dot and no-training groups. A 4 × 2 ANOVA with block (1–4) as within-subject variable and category (sequential-training vs. no-sequence-training/no-training) as between-subjects variable showed that RTs were linearly decreasing, as revealed by a significant linear polynomial contrast, F(1, 169) = 132.888, p < .001, η p 2 = . 440 . RTs of the no-sequence-training category were significantly lower, as revealed by a significant category main effect, F(1, 169) = 11.977, p < .001, η p 2 = . 066 . There was also a significant Block × Category interaction, F(1.977, 334.041) = 10.028, p < .001, η p 2 = . 056 .

To further investigate the Block × Category interaction, we conducted a univariate ANOVA for each block with category as between-subjects variable. The ANOVAs revealed that there was no difference between the categories in Block 1, F(1, 169) = 1.674, p = .023, η p 2 = . 010 , but there was a significant difference for all other blocks [Block 2: F(1, 169) = 5.293, p < .05, η p 2 = . 030 ; Block 3: F(1, 169) = 13.476, p < .001, η p 2 = . 074 ; Block 4: F(1, 169) = 21.335, p < .001, η p 2 = . 112 ]. For RTs in each category throughout the blocks, see Figure 3.

Interference between pre- and post-transfer sequences

As explained above, the observation of a visual sequence interferes with the later learning of an audiomotor sequence. A possible interpretation was general fatigue—that is, participants became tired after a long training session and performed worse in the later training. This possibility, however, was ruled out by employing a control condition with a non-sequential training phase. That is, participants of the random-dot condition were exposed to the same amount of stimuli as experimental participants, but the training phase had no structure. Still their test phase performance did not decrease. Hence, general fatigue is not a plausible explanation.

It is still possible, however, that differences in learning are linked to the use of resources. That is, processing of different sequences load to the same system, and that system has a limited capacity. Numerous studies have discussed the possibility of a modality- and domain-independent sequence learning mechanism. In a number of papers, Aslin, Saffran and colleagues showed that a similar mechanism helps infants in the extraction of statistical dependencies in auditory verbal (Saffran et al., 1996), auditory non-verbal (Saffran et al., 1999), and spatial–visual stimuli (Fiser & Aslin, 2002). There is also evidence that rule-based learning takes place similarly across modalities (Saffran, Pollak, Seibel, & Shkolnik, 2007). On the other hand, there are also studies showing that infants only learn rules if speech is involved (Marcus, Fernandes, & Johnson, 2007; but see also Saffran et al., 2007), and that adults are only able to learn two simultaneously presented grammars if their stimuli differ at least in dimension (Conway & Christiansen, 2006). The latter set of results led to a conclusion by Frost and colleagues (Frost, Armstrong, Siegelman, & Christiansen, 2015) that separate parallel statistical learning mechanisms exist. According to Frost et al. (2015), these mechanisms are overlapping in the neurological foundations, but operate separately for the different modalities.

The current experiment provides evidence for the overlap, but not for the separate mechanisms. A possible interpretation of the current results is that the development of different core sequences uses the same underlying learning mechanism. The extensive use of the mechanism decreases later learning, possibly due to neural fatigue. This decrease is affected by the sequential nature of learning, which explains why the random-dot conditions did not result in the same decreased sequence learning as the experimental conditions. If such a source-independent sequence learning mechanism truly exists, the current design would be a good candidate for further neuroscience research.

Neural fatigue of the mechanism responsible for sequence learning is not the only interpretation of the results. As explained above, previous studies have shown that a later learned sequence can retroactively inhibit a previously learned sequence, and the previously learned sequence can proactively inhibit the later learned sequence (Goedert & Willingham, 2002). It is plausible to assume that results are not due to the fatigue of the source-independent sequence learning mechanism, but an already acquired sequence proactively inhibits learning of a second sequence. This would also account for the results showing no group-based difference in the early blocks of the test phase, only emerging later. However, we argue that comparable reaction times in Block 1 reflect similar baseline RTs. It is also important to note that RTs of Block 1 are also numerically similar to RTs of the random block (Block 5). Studies using the SRT task generally focus on at least two parallel effects: sequence-independent and sequence-dependent effects (for a summary, see Robertson, 2007, p. 10073). The decrease of reaction times during the sequenced blocks is considered as a sum of both sequence-independent and sequence-dependent effects. That is, participants become faster as they get to know the sequence, but they also learn general skills, like visuomotor associations (audiomotor in the current experiment) between stimuli and responses, which can also decrease response latencies. This effect can be observed, for example, at the alternating serial reaction-time task, a task in which sequenced and random elements alternate. Results generally show that reaction times for random elements show a strong decrease, which can be interpreted as general skill acquisition (Howard & Howard, 1997).

It is plausible to assume that while the training sequence is proactively inhibiting the acquisition of the test sequence, there is still a decrease in reaction times. This decrease can be mainly driven by general skill acquisition. In the end, however, the test sequence was also learned, shown by a significant sequence learning effect. This latter explanation of proactive inhibition, however, points to the same direction as the previously discussed neural fatigue scheme, suggesting a general sequence learning mechanism, which processes sequences from multiple sources, but has a limited capacity. The aim of further studies should be to contrast the predictions of neural fatigue and proactive interference and to develop the time-course for the clash between consecutive sequences. That is, how long does being exposed to a sequence interfere with acquisition of later sequences.

In the previous section we discussed why being exposed to a sequence decreases learning another sequence. However, by itself neither neural fatigue nor proactive interference explains the observed negative transfer if the training phase and the test phase use the same sequence with different stimuli. In the next section we discuss characteristics of sequence representations in order to give an account for the negative transfer.

Structure-independent sequence transfer

A plausible explanation for the observed negative transfer between training and test phase even in the case of identical sequential structures can be based on the nature of representation developed during sequence learning. Our interpretation is built around stimulus dependence of the sequential representation. Stimulus dependence is used in a dual sense. On the one hand, it may refer to a common category for modality and domain dependence (see Kemény & Lukács, 2013)—an interpretation we cannot use now. On the other hand, stimulus dependence may reflect the phenomenon that a learning mechanism is constrained to its input, hence there is no transfer to other stimulus sets (Bregman, Patel, & Gentner, 2012). This is a central issue in sequence learning. Sequences are generally interpreted as the fixed order of single items, which leads to a dual nature of sequences: a kind of algorithm on the one hand, and unrelated values on the other (or syntax and vocabulary, as paralleled with the artificial grammar learning, see Gomez & Gerken, 2000; or Saffran, 2002). The question of stimulus dependence is whether sequential structure and elements being ordered can truly be dissociated, or a specific sequence is formed by the integration of the structure and the specific stimuli. The dissociation between structure and elements has been observed in human neuropsychological studies too. Sequence learning is mostly discussed in terms of the procedural–declarative memory distinction (Skosnik et al., 2002), which suggests a difference between well-defined, factual information and less defined, process-like knowledge (Squire, Knowlton, & Musen, 1993). As a consequence, procedural representations are expected to be less flexible.

This has already been observed in a procedural but non-sequential task, the Weather Prediction task (Reber, Knowlton, & Squire, 1996). In this task, participants are expected to abstract and accumulate statistical information from stimulus–stimulus contingencies. Reber et al. (1996) examined patients with amnesia, characterized by a general deficit in declarative memory, but intact procedural functioning. Their probabilistic classification performance was comparable to the performance of a healthy control group. That is, the clinical group was able to extract statistical information; the use was inflexible, however. Clinical participants could identify overall contingencies, but could not break them down to separate elements. Healthy participants, on the other hand, were able to trace back statistical information. This result shows that stimulus dependence is a characteristic of the procedural system, while declarative representations are not bound to stimuli.

A vast number of comparative psychological studies also focus on stimulus dependence. Previous studies found that non-human animals have a great difficulty to generalize and transfer long sequences. While Shackleton, Ratcliffe, and Weary (1992) showed that chickadees are able to transfer relative pitch to a different absolute pitch, this effect was only present in the case of two elements. However, in the case of longer sequences, relative pitch was not utilized by either European starlings (Page, Hulse, & Cynx, 1989) or Rhesus monkeys (Wright, Rivera, Hulse, Shyan, & Neiworth, 2000). These results show that sequential representations should be interpreted as inflexible and stimulus dependent, at least in non-human animals. ¹

The above studies provide evidence that structure and structured elements may not necessarily be dissociated from each other. That is, we can consider a sequence identical to another if both the sequential information and the elements themselves are identical. If two sequences share the same structure but differ on stimuli, the two sequences are just as different as two sequences with identical stimulus sets, but different ordering of the stimuli. This predicts a similar transfer between sequences of identical structure with different stimuli and between sequences with different structure. The current results provide evidence in favour of this hypothesis as the same and mirror sequences showed the same negative transfer effect.

This interpretation can be integrated with a two-system model of sequence learning (Daltrozzo & Conway, 2014; following Pierrehumbert, 2003). This model suggests that initial sequence binding takes place as a stimulus driven, bottom-up process. In the next step, higher order processes trigger top-down feedback mechanisms. It is plausible to assume that pure observation in the pre-training phase resulted in a bottom-up binding of sequential elements. This could explain the stimulus-dependent nature of the representation, as a higher order sequence has not had the opportunity to develop. However, the current study was not designed to test the predictions of the two-systems model, and hence this is only a possible scenario that requires further testing.

An alternative interpretation of the results is rooted in previous studies on statistical learning. A number of previous studies showed that triplet learning is expressed regardless of the order of the triplets. That is, after a short training session, familiarity or grammaticality judgements of backward triplets are higher than those of triplet fragments (Jones & Pashler, 2007; Tanaka & Watanabe, 2014; Turk-Browne & Scholl, 2009). These results argue that the representations of these triplets are flexible enough to recognize the reversed order. A major difference, however, is rooted in the length of the sequence: While statistical learning studies used triplets, we use a 12-element sequence. Although we have no evidence that the length of the sequence decreases flexibility, it would be surprising if participants were able to reverse an implicitly learned 12-element-long spatial sequence. Still, further studies are needed to address this question.

Future research, however, should be aimed to identify the minimal sequence difference that leads to negative transfer. It seems obvious that when participants are exposed to the same sequence with the same stimuli, we should observe no interference. Such a case would be two consecutive blocks of the same SRT task. Strictly speaking, however, no two sequences are identical. There is a very high number of task-independent environmental changes between two consecutive blocks of the same experiment. We consider these as random external noise. This noise is excluded though, and the sequence is abstracted from the random elements. We could expect the same effect in the current experiment. That is, participants exclude stimulus-based information and only focus on the spatial sequence. But results show the opposite. Why is it that some information can be excluded, while others cannot? Is it in connection with the structured nature of the information? That is, random elements are truly random, but modality-specific information was sequenced in this task. Or is it related to the cross-modal nature of the task, and no interference would be observed with intra-modal transfer?

The current study showed that a single mechanism with limited capacity is responsible for processing completely different sequences: in the current case, a visuospatial and an audiomotor sequence. While the design argues for a quite general sequence processing mechanism, its scope is yet unknown. The current experiment was designed to test transfer between visual and auditory spatial sequences, and hence it is possible that the scope is limited to sequences with a spatial component, despite being modality independent. These questions should be addressed by future research.