Phases of procedural learning and memory: characterisation with perceptual-motor sequence tasks

Abstract

Procedural learning and memory has been conceptualised as consisting of cognitive and autonomous phases. Although the Serial Reaction Time Task (SRTT) is a popular task used to study procedural memory (PM), it has not been used to explore the different phases of PM. The present study employed a modified SRTT and investigated whether it can distinguish phases of PM. Our results revealed that performance at the beginning of typing a repeating sequence was marked by a steep learning curve, followed by gradual improvements and ending in high performance levels without further improvement. Steep performance increases characterise the effortful learning of the cognitive phase, gradual increases at higher performances characterise emerging automatisation of the associative phase, and sustained highest performance characterises autonomous procedures when PM has formed. Our study presents an easy-to-use measure, capable of distinguishing phases of PM, and which can be useful to assess PM during brain development.

Introduction

Procedural memories are any acquired skill that once required effort to learn, but which, when formed, needs little conscious exertion to accomplish (Sanchez, Gobel, & Reber, 2010). This type of memory is categorised as implicit (or nondeclarative) memory, and by definition can operate without any participation of conscious thought (Roediger, 1990; Seger, 1994). In order for the procedural memory to be formed, however, the brain must recruit conscious, explicit (or declarative) processes (Beaunieux et al., 2012). The Memory Systems Theory proposes that explicit and implicit memory are functionally independent systems, an idea that is now widely accepted and supported by research (Eichenbaum & Cohen, 2001; Schacter, Wagner, & Buckner, 2000; Squire, Stark, & Clark, 2004). Several studies have sought to understand the involvement of explicit and implicit systems in its formation (Butters, Wolfe, Martone, Granholm, & Cermak, 1985; Schmidtke, Handschu, & Vollmer, 1996; Wilson, Baddeley, Evans, & Shiel, 1994), and aimed to connect specific psychometric measures with stages of procedural learning (Ackerman & Cianciolo, 2000). These studies suggested that procedural learning is not an entirely implicit phenomenon that occurs without conscious awareness and over the course of practice, but instead relies on nonprocedural functions involving various cognitive components such as intellectual ability (Ackerman, 1988), episodic memory (Pitel et al., 2007) and executive functions (Beaunieux et al., 1998, 2006) especially at the beginning of learning. In the earlier works of Fitts and Posner (Fitts, 1964; Fitts & Posner, 1967), the “power law of practice” was proposed to describe the decreasing reaction time nonlinearly with the number of practice trials taken. These ideas were originally theorised in Anderson’s Adaptive Control of Thoughts model (1992), which divided the power function into three distinct phases of procedural memory: cognitive, associative, and autonomous (Anderson, 2000). During the cognitive phase, a skill is acquired via active learning strategies such as working memory functions, resulting in a cognitive model of the task (Anderson, 1982). The cognitive phase is followed by the associative phase, in which the skill improves with repetition and practice; gradually, the cognitive scaffolding for the task is dropped as the skill becomes automated. Finally, in the autonomous phase, the task is performed smoothly with minimal completion time and conscious involvement (Hubert et al., 2007). This last phase is believed to primarily recruit implicit process of perceptual-motor speed (Ackerman, 1988, 1990). This model was empirically validated by Beaunieux and colleagues (2006), showing that the cognitive phase, starting at low performance levels, shows clear improvement in performance, while improvement is markedly reduced in the associative and autonomous phases (see also Ackerman, 1988). Together, these studies provide evidence that the creation of a procedural memory can be conceptualised as a sequence of independent, coordinated explicit and implicit processes.

Several instruments, such as mirror-reading, maze learning, Tower of Toronto, and Rotary Pursuit, have been employed in measuring procedural task ability in studies that revealed dissociations between explicit and implicit memory (Cohen, Eichenbaum, Deacedo, & Corkin, 1985; Poldrack, Desmond, Glover, & Gabrieli, 1998; Ward, Shum, Wallace, & Boon, 2002). However, exploring the sequence of interacting explicit and implicit processes has proven to be similarly complicated across the majority of existing procedural memory tasks (Sanchez & Reber, 2013). In their perceptual-motor sequence learning task study, Sanchez and Reber (2013) concluded that the initial performance was dominated by explicit memory and cognitive skills, as practice continued, coexistence of explicit and implicit learning occurred, but eventually only implicit memory sustained performance (see also Petersen, van Mier, Fiez, & Raichle, 1998). The acquisition of a procedural skill relies on the ability to sequence incoming information and ongoing motor output. The Serial Reaction Time Task (SRTT) is considered to be a robust measure of procedural visuomotor sequence learning, as evidenced by studies using patients with impaired explicit memory (Gobel, Sanchez, & Reber, 2011; Knopman & Nissen, 1987; Muslimovic, Post, Speelman, & Schmand, 2007; Vakil, Bloch, & Cohen, 2017). However, the SRTT has so far not been used to explore the three phases of procedural memory, nor to determine where on the learning curve the autonomous procedural memory is formed. In addition, many existing SRTTs use complicated sequences, which require several hours or even a full day to complete the task. This can lead to fatigue, which affects the individual’s learning. The fatigue factor can further restrict valid SSRT application to certain populations, such as college students, and limit the interpretation of procedural learning in patient populations, who tire more easily. A simplified procedural memory task would be applicable for testing clinical populations and characterising sparing and impairment for the different phases of procedural learning and memory.

In the current study, we accordingly employed a modified SRTT, the Finger-Typing Task (FTT), which aimed to investigate whether it is also a good measurement of procedural learning as the traditional SRTT, and determine if it can be used to distinguish the three phases of procedural learning, and to identify that a procedural motor memory had formed, with simple statistical comparisons. In accordance with Ackerman (1988) and Beaunieux et al. (2006), we expected to see a nonlinear relationship between the performance and the number of trials as described in Fitts and Posner (1967), which will show a considerable motor improvement in performance (steep learning slope) during initial learning through sequence repetition; then steady performance enhancements through rehearsal with a slower (less steep) learning slope, followed by a plateau (no learning slope) of sustained high performance. We further hypothesised that, after a rest period, the sustained high performance would be preserved in the trials following. We also added a control condition, in which we expected no improvement in performance when procedures are not repeated. In addition, subjects underwent a second procedural motor task, Rotary Pursuit (RP), which has been used to study perceptual-motor procedural memory over the course of sixty years (Ammons, 1947; Gabrieli, Stebbins, Singh, Willingham, & Goetz, 1997). Despite the genuine difference between the tasks, our directed hypotheses were that participants with higher performance in the FTT would also perform better in RP, and that participants with a steeper learning curve in the FTT would also produce steeper slopes in RP during the acquisition of procedural memory.

Methods

Participants

22 healthy volunteers participated in the study (15 women; mean age = 36.1 years, range 22–56 years). All participants were right-handed, and reported having no problems moving their fingers, hands, or wrists. Study participants were recruited through word of mouth. Subjects provided written informed consent to participate in this study. The Institutional Review Boards of Stanford University School of Medicine and SRI International approved study procedures.

The Finger-Typing Task (FTT)

The FTT is adapted from the traditional SRTT (Muslimovic et al., 2007; Walker, Brakefield, Morgan, Hobson, & Stickgold, 2002) and is designed to differentiate between the three phases of procedural memory (see also Cohen et al., 1985; Emick & Welsh, 2005). A Dell Latitude E6530 laptop was used to present the task and its instructions using E-Prime 2.0 software (Psychology Software Tools, Pittsburgh, PA) (Schneider, Eschman, & Zuccolotto, 2002). All volunteers performed the task individually in a quiet testing room with an instructor (J.-Y. Hong). Subjects typed letter sequences using the keys of “H”, “U”, “I”, and “L” on an English language keyboard. The four keys were covered by coloured stickers, which were marked with the letters “D”, “C”, “B”, and “A”, respectively. All surrounding keys were removed from the keyboard (the keys “G”, “Y”, “7”, “8”, “9”, “O”, “P”, “;”, “J”, “K”, “B”, “N”, “M”, “<”, and “>”) (Figure 1). To enable a comfortable position for the subject’s hand during typing, padding under the wrist and forearm was provided to support the dominant hand. A removable screen was fitted over the keyboard to prevent participants from seeing their fingers while typing. Typing performance was measured by the quantity of letters typed (speed) for each trial and recorded by E-Prime 2.0 software for later analysis. Each trial lasted 10 seconds.

Figure 1.

The Finger Typing Task (FTT) consisted of 3 sessions: PRACTICE, TESTING, and RETEST sessions. Within each session, the instructions and each trial were shown on the screen following the order of the timeline sequence along the downward arrows. There were 3 trials within the PRACTICE session. 5 Non-Repeating Sequence (NRS) and 30 Repeating Sequence (RS) with a 30-second break in the middle of the RS trials were given in the TESTING session. After minimal one hour, 15 RS trials and 5 NRS trials were given in the RETEST session.

The FTT consists of two test sessions: TESTING and RETEST (Figure 1). Before the test sessions, subjects underwent a brief practice session, during which they were instructed to place their dominant hand on the keyboard with their index finger over the “D” key, their middle finger over the “C” key, their ring finger over the “B” key, and little finger over the “A” key. After confirming that each subject was familiar with the four key positions and the assigned fingers for each key, their dominant hand was covered to prevent them from visually checking their finger positions while typing. Next, subjects were instructed to repeatedly type (from left to right) a sequence of letters shown on the screen during the three practice trials. They were instructed to perform the task silently (and not to read the letters aloud). When these instructions were completed, the program was initiated, and a three-second countdown appeared on the screen. For each of the three practice trials, the word “TYPE” first appeared on the screen for three seconds to prepare the subject for the 16-second practice sequence. Subjects were told that they could stretch and relax their fingers and hands before the letter sequence was presented. Each practice sequence consisted of a sequence of 8 letters. Following the practice (2–3 minute) session, the TESTING session (10–12 minutes) was administered (Figure 1).

TESTING included three conditions: a 5-second TYPE period (similar to the TYPE instruction condition in practice), during which the subjects were permitted to stretch and relax their fingers and hands; the “Non-Repeating Sequence” (NRS), which consisted of a series of non-repeating, unrelated sequences of 6 letters; and the “Repeating Sequence” (RS), a predetermined sequence of 6 letters which appeared repeatedly over several trials. The TESTING session began with five 10-second NRS trials, followed by thirty 10-second RS trials, during which the same sequence of letters was presented repeatedly. A 5-second pause from typing, during which the word “TYPE” was presented on the screen, was given prior to each NRS and RS trial (Figure 1). After 15 RS trials (the halfway mark), there was a 30-second rest period, which allowed subjects to remove their hands from the keypad and stretch to prevent muscle tension or soreness. After the 30-sec rest period another 15 RS trials were shown on the screen. The subjects were asked to repeatedly type each sequence of letters in order (from left to right) as quickly and accurately as possible within the 10-second trial time. No feedback was provided during NRS and RS trials about accuracy or typing speed. The subjects were informed that if they made a mistake, they should not attempt to correct themselves, just to continue with the task.

After a minimal 1-hour break after TESTING, a RETEST session was administered. The RETEST session was similar to TESTING and contained 15 RS trials that were followed by 5 different NRS trials, and there was no 30-second rest break. The instructions were the same as in the TESTING session: subjects were asked to type the letters on the screen from left to right repeatedly, and as quickly and accurately as possible with their hands covered.

Rotary Pursuit (RP) task

This task has been widely used to study hand-eye coordination and motor procedural learning (Corkin, 1968; Harrington, Haaland, Yeo, & Marder, 1990). In the current study, Rotary Pursuit was performed on the Photoelectric Rotary Pursuit Model No. 30014 (Lafayette Instrument Co., Lafayette, Indiana, 47903). During task performance, subjects were standing with their arm outstretched, holding a stylus in their dominant hand. Subjects were instructed to hold the stylus directly above a photoelectric sensor, and to track it while it rotated clockwise beneath a stationary glass plate without touching the stylus to the plate. When the stylus was moved into or out of alignment with the sensor, the subjects received feedback in the form of a clicking sound made by the Photoelectric Rotary Pursuit machine.

The machine was housed on a movable cart, and a standing pad was provided to shorter subjects, in order to ensure an optimal performance height. The initiation and rotation intervals of the rotating sensor was triggered and controlled by the Repeat Cycle Timer Model No. 51013 (Lafayette Instrument Co., Lafayette, Indiana, 47903). The Clock/Counter Model No. 54035 (Lafayette Instrument Co., Lafayette, Indiana, 47903) recorded the length of time that the subject was able to keep the stylus precisely above the photoelectric sensor. The number of times the stylus strayed from its position above the photoelectric sensor was recorded by Impulse Response Counter Model No. 58024C (Lafayette Instrument Co., Lafayette, Indiana, 47903).

In the Rotary Pursuit task, subjects completed one practice and four testing sessions. There were two trials in the practice session: the first trial involved tracking the sensor for 10 seconds at 20 revolutions per minute (rpm), the second trial involved tracking the sensor for 15 seconds at 40 rpm. After completing the practice session, four test sessions were given, with a one hour interim between each test session. Each test session contained eight trials, each with a 20-second tracking time at 40 rpm and a one-minute rest interval after four trials (Figure 2).

Figure 2.

The Rotary Pursuit (RP) task consisted of a PRACTICE and 4 TEST sessions with a one-hour minimum break between each TEST sessions. Within the PRACTICE session, one 10-second sensor tracking trail at 20 rpm was followed by another 15-second tracking trail at 40 rpm. Each TEST session contained 8 trials, each with a 20-second tracking time at 40 rpm and a one-minute rest interval after 4 trials.

The subjects were instructed to try their best to keep the stylus directly over the light as it travelled around the circle. After instructing the subject to begin, the photoelectric sensor was initiated. The length of time the subject was able to keep the stylus directly above the sensor and the number of clicks (indicating when the stylus deviated from its position) were recorded for each trial. These records were not revealed to the subject.

Data analysis

FTT

For each NRS and RS trials, we calculated finger-tapping speed as a performance indicator. Speed was defined by how many letters were typed within 10 seconds. To assess the individual curve of fine motor performance for NRS and RS, each 10-second trial was analysed separately in the order in which they were presented. According to Rickard (2004), Ackerman and Cianciolo (2000) and Ritter, Baxter, Kim, and Srinivasmurthy (2013), there are expected transitions and variances between trials at different time points during practice; therefore, averaging data can generate a gradually improving learning curve, smoothing out the transitions and increasing the reliability of the data. Next, we averaged the speed results of the trials to produce a smoothed (averaged) estimation of the individual learning curve. The first trial of the five NRS trials given during TESTING and RETEST was not included (i.e. NRS1 and NRS6) because subjects tend to miss the onset of the first trial, resulting in inaccurate performance estimates. The second and third trials were averaged, and the fourth and fifth trials were averaged. For simplicity’s sake, we named our averaged data points NRS23 (the average of NRS trial 2 and 3), RS789 (the average of RS trial 7, 8, and 9), and so on. For the 30 RS trials of the TESTING session, and 15 RS trials of the RETEST session, every three RS trials were averaged (TESTING: trials 1–3 (RS123), 4–6 (RS456), 7–9 (RS789), 10–13 (RS101112), 13–15 (RS131415), 16–18 (RS161718), 19–21 (RS192021), 22–24 (RS222324), 25–27 (RS252627), and 28–30 (RS282930); RETEST: trials 31–33 (RS313233), 34–36 (RS343536), 37–39 (RS373839), 40–42 (RS404142), and 43–45 (RS434445)) (Figure 3).

Figure 3.

FTT – Trial Averaging and Phase Identification. Mean typing speed and phases of procedural memory during the TESTING and RETEST sessions. *First significant trial-to-trial performance difference, followed by a series of statistically different performance levels, marking the onset of a new phase.

RP

The time of the “stylus on the target” was recorded for each trial in each session for each individual. We then averaged the individual’s performance (time of the stylus held over the target) between every two trials (Figure 5).

Figure 5.

RP – Trial Averaging and Phase Identification. Mean time-on-target and phases of procedural memory during the four RP sessions. *Identification of the first of a series of statistically different performance levels (time on target) between trials marking the onset of a new phase.

Identification of procedural learning and memory phases

Phase boundaries for the FTT

The motor speed performance for averaged NRS and RS time points for TESTING and RETESTING from each participant were entered for an ANOVA test using SPSS V.23 (Armonk, NY, IBM). In order to determine significant differences between the averaged trial-to-trial time points to identify the onset of a different phase of procedural learning, post-hoc analyses were performed. Significant trial-to-trial differences were tested at a significance threshold of p < .05 using the Bonferroni approach to correct for multiple comparisons. Therefore, phase boundaries were then identified by the first time point with significantly improved performance in relation to the previous points, which were RS456, RS192021 and RS343536 being first statistically different to a set of previous trials indicative of the onset of the next phase.

Phase boundaries for the RP

Same as the FTT, an ANOVA was conducted to determine any differences between averaged trial time points. Post-hoc analyses were performed to identify trial-to-trial differences for the averaged trials using Bonferroni correction for multiple comparisons, with the same thresholds as for FTT. Phase boundaries were also identified through statistical comparison of motor performance level relative to the previous trials.

Slopes and intercepts

For each individual, we calculated the slopes and intercepts of the curves from averaged time points separately for each phase, for both FTT and RP. The slopes provided an estimate of the average change in typing speed during a phase. A positive slope represents an improvement, while a negative slope represents a loss, and a slope of zero indicates that performance continued at the same level. Intercepts provide an estimate of an individual’s performance at the beginning of each phase. To test for significant differences in learning curves and performance between phases, slopes and intercepts from each phase were subjected into repeated measures ANOVA and Bonferroni-corrected post-hoc comparison, for FTT and RP.

FTT and RP comparisons

Spearman’s Rho correlation analyses tested for comparisons in procedural memory acquisition between different motor tasks, the FTT and RP task, for each phase (cognitive, associative, and autonomous).

Results

Procedural learning and memory: the Finger-Typing Task (FTT)

Speed

Supplementary Figure 1 shows the typing speed averaged over 22 subjects for each 10-second trial during TESTING and RETEST sessions. Fine motor speed did not significantly differ between NRS trials. For RS, typing speed steadily improved with repetition of the same sequence (procedural learning).

Phases of procedural learning and memory

To define the three phases (cognitive, associative, and autonomous) of procedural memory acquisition, we used the averaged speed results from each individual, where every two trials of NRS (e.g. NRS23) and every three trials of RS were averaged (e.g. RS123). The first trial of non-repeating sequences (NRS1 and NRS6) was excluded from analyses, i.e. the start of the task (NRS1) and the start of NRS in the RETEST session (NRS6), respectively.

Figure 3 depicts the trial-averaged mean curve of all 22 participants. A repeated measures ANOVA tested for differences in performance between the 19 averaged data points of the curve and revealed point-specific differences in typing speed over all data points (df = 18; F = 15.112; p < .001). The different colours depict the 3 phases of procedural memory acquisition resulting from the repeated measures ANOVA and based on the post-hoc tests after Bonferroni correction (Supplementary Table 1).

TESTING session

NRS: As expected, there was no significant difference in speed between all non-repeating sequences (NRS23, NRS45, NRS78, NRS910) and the first repeating sequence RS123 (all p’s ≥ .999) (Supplementary Table 1). RS: First phase (1-RS-T): Relative to the start of the task (NRS23), the first RS time point showing significantly enhanced typing speed was RS456 (See Supplementary Table 1) depicting a steep speed enhancement with repetition (from RS123 to RS456; see Figure 3), marking the cognitive phase at the beginning of procedural learning. Second phase (2-RS-T): Relative to the non-repeating sequences (NRS23, p = .036), faster typing speed was observed for all the rest of the RS trials starting with RS456 (p ≤ .001, Supplementary Table 1), i.e. greater efficiency in typing the repeated letter sequence pattern. During the second phase, performance speed improved gradually over time, from RS456 to RS192021, marking the next associative phase of procedural learning (Figure 3). Third phase (3-RS-T): Relative to the first repeating sequence RS123, significantly improved performance within the repeating sequences, i.e. faster speed, was not observed until RS192021 (p = .017), and continued for all following RS trials: RS222324 (p = .005), RS252627 (p = .011), RS282930 (p = .005), even including all RS trials of the RETEST session: RS313233 (p = .008), RS343536 (p < .001), RS373839 (p < .001), RS404142 (p < .001), RS434445 (p < .001) (Supplementary Table 1). All RS trials following RS192021 during TESTING did not differ from each other (p’s ≥ .999), i.e. showing a constant of the improved RS typing speed, marking the transition from the associative to the autonomous phase. RS192021 is therefore regarded as the starting point of the autonomous phase (Figure 3).

RETEST session

First phase (1-RS-R): Typing speeds of RS trials in the RETEST session were all significantly faster than that of RS123 trials (p’s < .05, Supplementary Table 1). The high-performance levels acquired through repetition during the testing session were retained during the rest period: The averaged typing speeds for RS313233 at the beginning of the RETEST session were at the same level as the averaged typing speeds for trials at the end of the TESTING session (RS192021, RS222324, RS252627 and RS282930 trials, p’s ≥ .999). Significantly faster typing speed was observed for RS343536 of the RETEST session compared to the RS456 (p = .033), which resembled the steep performance improvement in the cognitive phase in the TESTING session. Second phase (2-RS-R): This second high performance level, starting at RS343536 during RETEST, continued without further improvement in typing speed for all following RS373839, RS404142 and RS434445 trials, thereby marking another autonomous phase. Here, typing speed achieved a new high point, higher than the speed recorded for any of the previous trials, including the TESTING session.

Finally, at the end of the experiment, we confirmed that typing speed slowed considerably for new sequences, i.e. NRS78 and NRS910. Typing speed for NRS at the end of the RETEST session did not differ from that for NRS23 and NRS45 at the beginning of the TESTING session (p’s ≥ .999, Supplementary Table 1). Thus, typing speed for non-repeating sequences remained the same throughout the entire experiment and was significantly slower than for repeating sequences, i.e. all RS except RS123 (p’s < .05, Supplementary Table 1). For validation of the FTT, we acquired another small independent sample (see Supplementary Materials). In addition, we confirmed the phases observed here by using the intercepts and slopes of each phase.

Characterising each phase by intercept and slope

Intercepts.

Table 1 shows the group means for intercepts for each phase, including non-repeating and repeating sequences using repeated-measures ANOVA (F(1,21) = 51.37, p < .0001). Post-hoc pairwise comparisons revealed no significant difference in intercepts between non-repeating sequences at the beginning (NRS-T) and end (NRS-R) of the FTT. Also, NRS-T and the first repeating sequences (1-RS-T) did not differ from each other. With task continuation, typing speed in repeating sequences significantly increased from phase to phase (from 1-RS-T to 2-RS-R, all p’s < .05) with one exception: the performance level dropped slightly but significantly after the break (1-RS-R) compared to the previous phase (3-RS-T). A second significant increase occurred thereafter, during RETEST (from 1-RS-R to 2-RS-R) (Figure 4, black triangles depict intercepts).

Table 1.

FTT Intercepts: Means and standard deviations for typing speed in each phase of the TESTING and RETEST sessions.

Phase	Mean Intercept	Std. Deviation	N	LSD Post-hoc Comparison	Mean Difference	Std. Error	Sig.
NRS-T	1.436364	.6550503	22	NRS-T vs. NRS-R	.161	.111	n.s.
1-RS-T	1.260606	.6113249	22	NRS-T vs. 1-RS-T	.176	.121	n.s.
2-RS-T	2.444848	.9616717	22	1-RS-T vs. 2-RS-T	−1.184	.153	.0001
3-RS-T	2.873106	1.0363306	22	2-RS-T vs. 3-RT-T	−.428	.121	.002
Break
1-RS-R	2.448485	1.2927444	22	3-RT-T vs. 1-RS-R	.425	.172	.022
2-RS-R	3.445455	1.1964916	22	1-RS-R vs. 2-RS-R	−.997	.146	.0001
NRS-R	1.275000	.3340766	22	2-RS-R vs. NRS-R	2.170	.231	.0001

Repeated measures ANOVA for intercept (linear F(1,21) = 51.37, p < .0001). Post-hoc comparison adjusted for multiple comparisons: Least Significant Difference (LSD). Abbreviations: NRS: non-repeat sequences; RS: repeat sequences; T: TESTING session; R: RETEST session; NRS-T: trials NRS23–NRS45; 1-RS-T: trials RS123–RS456; 2-RS-T: trials RS789–RS161718; 3-RS-T: trials RS192021–RS282930; 1-RS-R: trials RS313233–RS343536; 2-RS-R: trials RS373839–RS434445. NRS-R: trials NRS78–NRS910.

Figure 4.

FTT – Phase performance level and learning curves. Mean intercepts and slopes for identified phases. Abbreviations: NRS: non-repeat sequences; RS: repeat sequences; T: TESTING session; R: RETEST session; 1-RS-T: first phase-repeat sequences-testing session; 2-RS-T: second phase-testing; 3-RS-T: third phase-testing; 1-RS-R: first phase-repeat sequences-retest session; 2-RS-R: second phase-retest. *Statistically significant difference.

Slopes.

Table 2 shows the group means of the slopes for each phase and the learning slope comparisons using repeated-measures ANOVA (F(1,21) = 20.17; p < .0001). The learning curve at the start of the repeating sequences during TESTING (1-RS-T slope = .571) was significantly steeper compared to prior non-repeating sequences (NRS-T slope = −.027), as well as for the following repeating sequences (2-RS-T). The second phase showed a moderate learning curve (2-RS-T slope = .074) with significantly faster typing speed at the end of the phase (RS192021) (p = .017). The learning slope of the 2nd phase was less steep than the slope of first phase (p = .0001) and, although descriptively steeper, it was statistically similar to the slope of the 3^rd phase (3-RS-T slope = .019).

Table 2.

FTT Slopes: Means and standard deviations for typing speed slopes of each phase of the TESTING and RETEST sessions.

Phase	Mean Slope	Std. Deviation	N	LSD Post-hoc Comparison	Mean Difference	Std. Error	Sig.
NRS-T	−.027273	.2534182	22	NRS-T vs. NRS-R	−.068	.067	n.s.
1-RS-T	.571212	.4145879	22	NRS-T vs. 1-RS-T	−.598	.115	.0001
2-RS-T	.074416	.0909840	22	1-RS-T vs. 2-RS-T	.497	.095	.0001
3-RS-T	.019470	.1360219	22	2-RS-T vs. 3-RS-T	.055	.037	n.s.
Break
1-RS-R	.477273	.3224224	22	3-RS-T vs. 1-RS-R	−.458.	.069	.0001
2-RS-R	−.034697	.0989076	22	1-RS-R vs. 2-RS-R	512	.069	.0001
NRS-R	.040909	.1444500	22	2-RS-R vs. NRS-R	−.076	.036	.048

Repeated measures ANOVA for slope (quadratic F(1,21) = 20.17, p < .0001). Post-hoc comparison adjusted for multiple comparisons: Least Significant Difference (LSD). Abbreviations: NRS: non-repeat sequences; RS: repeat sequences; T: TESTING session; R: RETEST session; NRS-T: trials NRS23–NRS45; 1-RS-T: trials RS123–RS456; 2-RS-T: trials RS789–RS161718; 3-RS-T: trials RS192021–RS282930; 1-RS-R: trials RS313233–RS343536; 2-RS-R: trials RS373839–RS434445. NRS-R: trials NRS78–NRS910.

After a break, at the beginning of the RETEST session, there was another steep learning curve (1-RS-R slope = .477) that was comparable to the slope for RS trials at the beginning of the TESTING session (1-RS-T slope = .571) (Figure 3). A flat slope (2-RS-R slope = −.034) followed this steep slope, i.e. there was no further improvement in typing speed in the second phase of the RETEST session (Figure 3). The slopes of NRS at the beginning (NRS-T slope = −.027) and end (NRS-R slope = .041) of the FTT were flat and did not differ from each other (p = .32) (Figure 4, grey circles depict the mean slope of each phase).

A steeper learning slope during the first cognitive phase (1-RS-T) correlated with higher levels of performance (intercepts) later on in the task (2-RS-T: Rho = .56, p = .003; 3-RS-T: Rho = .46, p = .015; 1-RS-R: Rho = .56, p = .003; 2-RS-R: Rho = .43, p = .022), but not with performance levels for non-repeating sequences (no procedural memory) (intercept NRSR: Rho = .14, ns).

Procedural learning and memory: the Rotary Pursuit task (RP)

Time on target

In the RP task, performance was defined as the amount of time spent on the target. The average time spent on the target was recorded for a total of 32 trials (8 trials in each of the 4 test sessions) for each subject; group means from all 22 subjects are plotted in Supplementary Figure 2 for each RP trial and session in order of appearance. Over all trials and sessions, RP performance improved steadily.

In the next step, we averaged performance over every two trials (Figure 5) and named the averaged data points T1112 (average of trial 11 and 12), T1314, T1516, etc. An ANOVA tested for differences in performance over all time points and found significant differences in the time spent on the target for the 16 averaged data points (df = 15, F = 4.49; p < .001). The planned paired comparisons using post-hoc tests (after performing the Bonferroni correction) were used to find time points that differed significantly to explore different phases of procedural memory acquisition for the RP task (Supplementary Table 2).

First phase (1-RP): Performance between the first trials of the task (T1112) and all following averaged trials of session 1 and 2 did not significantly differ (p’s > .05), marking them as part of the same phase of procedural memory acquisition. Second phase (2-RP): Significant differences in performance (time spent on the target) was observed between the first trials of session 1 (T1112) and all averaged trials of sessions 3 and 4 (session 3: T3132 p = .016; T3334 p = .006; T3536 p = .004; T3738 p = .006; session 4: T4142 p = .002; T4344 p = .001; T4546 p < .001; T4748 p = .001). Thus, Rotary Pursuit performance revealed two different phases of procedural memory acquisition, the first phase of procedural learning spanning sessions 1 and 2, from T1112 to T3132, and the second phase spanning session 3 and most of session 4, from trials T3132 to T4546 (Figure 5; see Supplementary Table 2).

Third phase (3-RP): Relative to the T1314, the time-spent-on-target differed significantly for the last trials of session 4 (T4546 p = .007, T4748, p = .04), marking the transition to a third phase, starting at T4546. No other significant differences were found between trials (from T1516 to T4748; see Supplementary Figure 2 and Supplementary Table 2). Figure 5 depicts these differences with different colours marking the phases of procedural memory acquisition and the breaks between sessions. Importantly, each of the three aforementioned unique phases can be validated by the intercepts and slopes of each phase.

Characterising each phase by intercept and slope.

Intercepts.

Table 3 shows the group means of the intercepts for each phase and session. A repeated measure ANOVA on the intercepts showed an overall performance increase over all sessions, i.e. longer times the stylus was held over the rotating target (F(1,21) = 94.93, p < .0001). Post-hoc comparisons revealed significant differences in performance level between all three phases with a performance level increase from first to second phases, and further increase from second to third phases.

Table 3.

RP Intercepts: Means and standard deviations for time-on-target for each phase and session.

Phase	Mean Intercept	Std. Deviation	N	LSD Post-hoc Comparison	Mean Difference	Std. Error	Sig.
1-RP	9.547735	3.4069127	22	1-RP vs. 2-RP	3.495*	.43	.0001
2-RP	13.042288	2.7717386	22	2-RP vs. 3-RP	1.427*	.421	.003
3-RP	14.469318	3.0898146	22	1-RP vs. 3-RP	4.922*	.505	.0001

Abbreviations: 1-RP: first phase of Rotary Pursuit (T1112 to T2728), 2-RP: second phase of Rotary Pursuit (T3132 to T4344), 3-RP: third phase of Rotary Pursuit (T4546 to T4748).

Slopes.

Table 4 shows the mean slopes during each phase of the RP task and comparisons between the slopes. A repeated measure ANOVA on the slopes showed a significant decrease in the slopes over the three phases (F(1,21) = 7.93, p = .01). Post-hoc comparisons revealed a significantly steeper slope (learning curve) for the first phase of Rotary Pursuit (T1112 to T2728) than the second (T3132 to T4344) and third phase (T4546 to T4748), but no significant difference in the slopes comparing the second and third phases.

Table 4.

RP Slopes: Means and standard deviations for time-on-target for each phase and session.

Phase-Session	Mean Slope	Std. Deviation	N	LSD Post-hoc Comparison	Mean Difference	Std. Error	Sig.
1-RP	.411479	.2326601	22	1-RP vs. 2-RP	.317*	.086	.004
2-RP	.094679	.2673778	22	2-RP vs. 3-RP	−.468	.249	.222
3-RP	−.373636	1.2343675	22	1-RP vs. 3-RP	−.785*	.279	.031

Abbreviations: 1-RP: first phase of Rotary Pursuit (T1112 to T2728), 2-RP: second phase of Rotary Pursuit (T3132 to T4344), 3-RP: third phase of Rotary Pursuit (T4546 to T4748).

In contrast to our observations for FTT, a steeper learning slope during the cognitive phase in the RP task (1-RP) did not predict higher performance levels (intercepts) later on in the task (intercepts for 2-RP and 3-RP, all p’s > .05). Here, healthy individuals who started out with low RP performance levels, i.e. less time-on-target, actually had steeper learning slopes during the cognitive phase (1-RP slope/intercept: Rho = −.62, p = .002; 1-RP slope/T1112: Rho = −.50, p = .019).

FTT and RP comparisons

Spearman’s Rho correlation analyses tested if the levels of performance of the two procedural memory tasks are related to one another. We correlated the individuals’ performance levels between the two tasks, i.e. typing speed for non-repeating and repeating sequences for FTT, and time-on-target for RP. First, we tested if the baseline performance levels at the beginning of each task are related to each other, and we found that for FTT non-repeating sequences NRS45 (i.e. no procedural learning), faster typing speed correlated moderately with better performance (longer time-on-target) at the beginning of the RP task (T1112) (Rho = .443, p = .02).

For repeating sequences, we compared the averaged data points of both tasks during procedural learning, and found several moderate correlations (Supplementary Table 3). Specifically, we found that a higher level of RP performance at T1516 in the first task run correlated with higher FTT performance levels, over several data points, in the first task run (TESTING), and with the overall peak performance at RS343536 during RETEST (see Figure 3). Similarly, the level of FTT performance at RS161718 correlated with higher RP performance levels, at several data points, in the first and second RP test sessions (cognitive phase). When comparing trials without further improvement (after reaching the peak), i.e. procedural memory, of both tasks (FTT: RS222324 to RS282930 and RS373839 to RS43445; with RP: T4748), no significant correlations between the two tasks were observed (all p’s > .05).

Discussion

Procedural memory of the Finger Tapping Task

The goal of this study was to present the FTT as a valid task that is able to assess the procedural learning curves, traditionally described as nonlinear changes of performance with practice (see power law of practice, Fitts & Posner, 1967), and to use the FTT to differentiate the three phases of procedural learning and memory. There are a number of benefits unique to the FTT. These include less total time required to complete than the SRTT (Hotermans, Peigneux, Maertens de Noordhout, Moonen, & Maquet, 2006; Walker et al., 2002) due to the FTT’s simpler sequence, shorter training period for each trial block, and a flexible interval between sessions. Also, the task features a longer rest period, which minimises muscle tension and exhaustion. Moreover, because the FTT uses letter instead of numeric sequences, it avoids recruiting procedural knowledge of number-finger-matching, a significant confounding variable in subjects who play musical instruments (Musafia, 1971; Smith & McDowall, 2004). The non-repeating letter sequences can also be used as a control condition to better assess whether improved finger tapping performance is due to procedural learning, rather than to a general improvement in typing speed due to practice. The RETEST session of the FTT can be used to ensure that an autonomous procedural memory has formed, which is indicated by typing speeds of learned sequences that are comparable to those achieved before the break, i.e. whether the procedural memory can be retrieved. In addition, the RETEST session can demonstrate the plateauing performance in the autonomous phase is not a result from the “ceiling effect” imposed by task constraints (Norman & Bobrow, 1976).

As anticipated, FTT performance followed the expected pattern, which can be differentiated into phases by using a statistical approach without collecting many psychophysiological variables (Beaunieux et al., 2006). A dramatic increase in performance occurred at the beginning of the RS trials – consistent with an initial cognitive phase learning curve. This was followed by steady improvements as the RS was rehearsed – consistent with an associative phase, after which an autonomous phase of steady high-level performance was maintained. After a rest period, there was a second dramatic improvement in performance – similar to the initial cognitive phase learning curve, which reached a new high level that was then sustained. Thus, the continuous high-level performance during TESTING cannot be a consequence of the ceiling effect, which would negate further improvements during RETEST. Statistical analysis of the FTT dataset intercepts identified this period of sustained performance height as another separate phase, consistent with a second autonomous phase. As expected, the performance of the last NRS trails was not significantly different from the beginning NRS trails.

Phase intercepts (an estimate of the performance level at the beginning of each phase) were significantly different between cognitive and associative phases, as well as between associative and autonomous phases, evidencing a significant improvement in performance with task progression. In addition, our slope analysis showed significant differences between cognitive and associative phases, and between cognitive and autonomous phases. However, a significant difference was not evident between associative and autonomous phases. These findings further characterise the first phase of procedural learning as a cognitive stage involving working memory and intervention of cognitive functions; the associative phase is thus considered as an intermediate stage with reduced cognitive involvement, while the autonomous phase is more appropriately described as an implicit memory stage which is independent of the executive functions (Hubert et al., 2007). Therefore, we propose that the term “procedural learning” is applicable for describing the phases of memory formation (cognitive and associative phases), whereas “procedural memory” is the proper term for the autonomous phase.

The slopes calculated from the initial learning curve and the learning curve following the rest period were not significantly different, suggesting that this second post-rest curve would be best characterised as a second cognitive phase. However, the results of a study by Hotermans and colleagues (2006) indicated that rest periods can support consolidation-based stabilisation. A review of a broad range of research in memory consolidation indicates that procedural memory consolidation occurs during rest periods between sessions of motor or cognitive learning (Hauptmann, Reinhart, Brandt, & Karni, 2005; Walker, 2005). There appear to be two forms of consolidation: stabilisation (retention and maintenance of learned material), and enhancement, a phenomenon in which delayed learning occurs in the form of improvements in performance without added practicing (Walker, 2005). It was originally believed that learning enhancement required a period of sleep, whereas stabilisation occurred after an interval of rest (Hotermans et al., 2006; Walker et al., 2002). However, when a skill is acquired implicitly, the enhancement portion of memory consolidation appears to depend solely on amount of time elapsed since practice began (Robertson, Pascual-Leone, & Miall, 2004). The cognitive phase therefore constitutes a measurable leap of improvement in motor skill performance – the initial learning curve. After a short, wakeful rest period, allowing for memory consolidation, an additional boost in performance can occur, resulting in even higher performance levels (Hotermans et al., 2006), similar to distributed learning conditions typically utilised for more complex skill training, as in athletic disciplines (Dail & Christina, 2004).

Although we did not conduct a manipulation of wakeful rest to further support the assumption of a transient consolidation period enhancing procedural memories, we did observe a boost in performance after wakeful rest. In accordance with Hotermans et al. (2006), this boost may be thought of as another initial learning process that merely begins at a more advanced level, to which the subject has temporary access. If, as proposed by Beaunieux et al. (2006), the cognitive phase of procedural learning is accomplished via explicit functions, and if, with further practice, explicit and implicit processes begin to operate in parallel during the associative phase, thereby enabling continued high performance levels without tiring (because of emerging automatisation of non-conscious skills), it can be speculated that rest- or sleep-dependent consolidation processes can further improve performance (Robertson et al., 2004).

As we anticipated, speed did not differ between NRS trials at the beginning and end of the FTT. This result indicates that NRS trials constitute a useful control condition, and confirm that the performance improvements over the course of the RS trials reflected actual procedural learning of the letter sequence. NRS trials bracketing the RS trials may therefore serve a useful purpose in future studies using the FTT by controlling for subject-specific confounds and providing a baseline level of typing performance with which to compare RS learning.

Clinical relevance

The storage and retrieval of explicit memory is associated with the medial temporal lobe (MTL), a structure that participates in a network that is able to operate independently from implicit memory processes (Squire, 1992). Studies of patients with large MTL lesions have repeatedly indicated that explicit acquisition of any new information is extremely limited (Reed & Squire, 1998; Sanchez & Reber, 2013). The amygdala is believed to trigger explicit learning of semantic and episodic (factual and temporal) information (Phelps & LeDoux, 2005), which the hippocampus then consolidates into memories (Eichenbaum, 2001; Eichenbaum & Cohen, 2001; Gabrieli et al., 1997). Implicit learning is associated with a network of brain regions distinct from the MTL, which are involved in motivation, coordination, and visual and auditory processing (Buckner et al., 1995; McClelland, McNaughton, & O’Reilly, 1995). The implicit acquisition of procedural memory is associated with motor circuitry constituting a network of brain regions that include the premotor cortex, supplementary motor area, cerebellum, and prefrontal regions supporting executive functions, all of which participate in a larger neural circuit with the basal ganglia (Squire, 2004).

However, research indicates that many forms of learning that normally recruit explicit processes involving the MTL are replaced by implicit processes via automation, albeit at a slower rate (Bayley & Squire, 2002). SRTTs, such as the FTT, appear to be particularly versatile in this respect, in that they can differentiate between phases of procedural learning and memory, which are differently associated with cognitive and implicit procedural functions. Parkinson’s disease patients with deficits in explicit memory, attention and executive functions showed reduced motor sequence learning in a SRTT task, but were able to acquire the procedural motor skill after practice (Muslimovic et al., 2007). Similarly, improvements in SRTT performance have been demonstrated by patients with explicit memory impairments (Knopman & Nissen, 1987; Nissen & Bullemer, 1987). These findings suggest that while improvements can be seen in the SRTT despite explicit learning deficits, and although the patterns of learning differ from those with intact learning processes; once the sequence-specific knowledge is acquired, the procedural memory appears to be independent of impaired executive function.

Procedural memory in the Rotary Pursuit task

In order to provide more insight into procedural memory processing, we compared the FTT with the RP task, a measure of procedural learning with a much longer history of use than the SRTT (Vakil et al., 2017), from which the FTT was derived. Similar to procedural learning and memory in the FTT, we also identified three phases of procedural memory in the RP task using the same definition and slope intercept analysis for phase definition as in the FTT. These phases followed the predicted order of the phases model of procedural memory (Anderson, 1982). During RP, comparatively more time was spent in the cognitive and associative learning phases. Only a brief autonomous phase was established at the end of the forth session, following the final associative phase. As with the FTT, our results indicated overall performance improvement throughout the duration of the task.

In analysing the correlations between FTT and RP, we found that faster typing speed (FTT) and longer time-on-target (RP) were related at the beginning, before procedural learning had started, indicating that individuals have corresponding baseline performance levels across tasks, notwithstanding the different requirements of the actual motor skill performed. During the cognitive and associative phases of procedural learning, performance levels were indeed related among several averaged trials. In particular, the peak performance level in the first session of the RP task (T1516) was related to FTT performance levels in the associative phase as well as to peak performance reached after consolidation. However, when the procedural memory was formed (autonomous phase), levels of performance were not related between tasks. Thus, despite the genuine similarities in procedural learning between tasks, with steepest gains in performance occurring in the initial learning curve of the cognitive phase followed by moderate, continuous gains at higher performance levels in the associative phase, the relationship between the RP and FTT dissolved in the following phases of procedural memory. This result was likely due to individuals’ diverse learning curves for each task.

There were a number of significant differences between the RP and FTT tasks. The FTT shows less total improvement in performance overall, but isolated, dramatic improvements are more evident. The cognitive and associative phases in RP are also greatly dilated compared to the analogous FTT phases. Additionally, the FTT has only one rest period in which consolidation may take place, while RP has three. It could be argued that the cognitive phase in the FTT is dependent on the sequence being simple and potentially explicitly transparent to the participants; however, during RP comparatively more time was spent in the cognitive phase than in the FTT. Thus, cognitive guidance is likely a general characteristic of procedural learning and probably generalisable to other PM tasks, while sequence complexity and task difficulty influence the length and slope of improvement in the cognitive phase. Our results further indicate that initial effort and cognitive engagement involved in learning is most similar at the beginning of any novel task, and becomes progressively divergent as each task’s idiosyncratic skillset produces more specialised learning. As distributed learning, while impacting implicit learning in the associative and autonomous phases, does not appear to have any impact on the involvement of explicit processes, this explanation would also be consistent with the three phases model (Beaunieux et al., 2006).

Limitations

Although we were able to distinguish between different phases of procedural learning and memory, the small sample size is a limitation for interpretation of the individual variations in phase onsets and learning curves, e.g. as a function of adult age. Consequently, future studies with larger samples, as well as another round of data collection, are needed to confirm and replicate our observations about learning curves and phases to build a foundation for studies on procedural memory during development, healthy aging, and disease. Additional neuropsychological assessments and functional neuroimaging studies are helpful to identify the exact implicit and explicit functions involved in the different phases of FTT procedural memory acquisition. Future studies could provide a convergent test of the proposed distinction between phases by introducing a secondary task component. We would predict, for example, that the cognitive phase would be more vulnerable to secondary-task interference than later phases. Furthermore, our correlational results comparing FTT and RP performance imply that procedural memory is a complex concept and strongly depends on the kind of motor skills involved. Consequently, learning slopes may be difficult to compare between different procedural memory tasks. For example, the expanded duration of cognitive and associative phases in RP, when compared to the FTT, may be due to the greater difficulty of the RP task. Also, the greater number of rest periods during RP could have affected performance by boosting learning through consolidation, resulting in continuous improvement over several task runs. However, if more RP trials were given, the autonomous phase might be more noticeable at the end of the RP task. The small sample size of this study also may have obscured significant jumps in RP performance after consolidation phases, marking procedural learning beyond active, conscious practicing of the motor skill. In other words, there may indeed be a meaningful consolidation effect, which is rendered invisible by the distributed nature of RP. The results of Beaunieux and colleagues (2006), which found that distributed cognitive procedural learning resulted in an expanded associative phase and a more gradual automation of the task, supports this proposition. However, the many differences between the FTT and RP resulting from the RP’s higher level of difficulty (e.g. greater complexity and cognitive recruitment, whole body movement coordination, higher energy and motor demands, longer time spent on task) are also likely to share responsibility for these dissimilarities.

Conclusions

Originally, Beaunieux and colleagues (2006) empirically confirmed the existence of three separate phases, and the conceptualisation of the three phases as a combination of explicit and implicit contributions to the procedural memory formation. Our study provides further support to Anderson’s three phases model (Anderson, 1982) using an analogous and easy-to-use task. In particular, the FTT exploits the capacity of the SRTT to differentiate explicit and implicit memory processes (Sanchez et al., 2010) through the identification of different learning phases, and is a measure capable of distinguishing the cognitive and associative phases of procedural memory as the learning portion of the procedural memory process, and the autonomous phase as a true, functionally autonomous, procedural memory. Moreover, because of the flexible span of time between trials and simplified sequence, the FTT can be easily incorporated with other neuropsychological tasks. Additionally, it can be easily adapted to brain imaging studies to accommodate subjects with motor impairments such as Parkinson’s disease (Jankovic, 2017).