Dual n-back working memory training evinces superior transfer effects compared to the method of loci

Working memory (WM) training is a prevalent intervention for multiple cognitive deficits, however, the transfer effects to other cognitive tasks from gains in WM induced by different training techniques still remains controversial. Therefore, the current study recruited three groups of young adults to investigate the memory training transference, with N-back group (NBG) (n = 50) training on dual n-back task, Memory Palace group (MPG) (n = 50) on method of loci, and a blank control group (BCG) (n = 48) receiving no training. Our results showed that both training groups separately improved WM capacity on respective trained task. For untrained tasks, both training groups enhanced performance on digit-span task, while on change detection task, significant improvement was only observed in NBG. In conclusion, while both techniques can be used as effective training methods to improve WM, the dual n-back task training method, perhaps has a more prominent transfer effect than that of method of loci.


Results
Training performance. Both NBG and MPG showed a linear improvement across memory training on their respective training task (Fig. 1). The average daily maximum performance of NBG improved from n = 2.21 (SD = 0.798) at day 1 to n = 5.71 (SD = 1.336) at day 20 (t (47) = 18.358, p < 0.001). In the MPG, the mean daily maximum performance of words memory also increased gradually from the beginning of memory test (day 11) (M = 3.34, SD = 0.984) to the end (day 20) (M = 10.87, SD = 2.763) (t (46) = 19.171, p < 0.001).

Discussion
In the present study, we applied memory training techniques with n-back task for NBG and method of loci for MPG to improve WM performance. Two trained tasks, dual n-back task and words memory task, and other two non-trained tasks, digit-span task and change detection task, were measured at both pre-test and post-test. Both training groups resulted in greater training improvement, with NBG displaying significant performance improvement of n-back task, and MPG in words memory task. For untrained tasks, both training groups yielded significant transfer effect to digit-span scores. However, for change detection task, only NBG showed significant improvement on response speed, while MPG did not show significant enhancement.
In accordance with previous evidence suggesting that n-back training can increase WM capacity and task performance 18,19,26,[29][30][31] , our study also observed great improvement on performance of each day training as well as on performance of dual n-back task after training in NBG. Moreover, the significant performance improvement on dual n-back task was observed uniquely in NBG rather than MPG or BCG for higher task loads (4-back and 6-back). For 2-back task load, significant improvement in performance was realized in both NBG and BCG. Perhaps this may be due to the learning effect of easy tasks, such as n-back task with lower loads in BCG. However, there was no significant effect on performance in 2-back task load in MPG, implying that there was no learning effect. A possible explanation for this is that perhaps the memory process of squares and consonants was intervened by the utilization of method of loci, which is successful in enhancing episodic memory with items such as words and names 36,48 . Considering task difficulty of 6-back in n-back task, 2-to 4-back trained task loads may be more suitable to obtain more accurate measures of the effect of different training techniques.
For words memory task, training effect was found uniquely in MPG on performance of abstract words rather than concrete words. On one hand, it is well known that the processing of abstract words is much more difficult than that of concrete words [57][58][59][60][61][62][63] . On the other hand, the utility of method of loci has been reported to facilitate the transformation of abstract information into concrete information 25 , which can be further processed more easily by memory-related neural system. It is, therefore, not surprising that the WM capacity for abstract words increased in response to training of method of loci. Additionally, all groups resulted in improvement on concrete words memory in post-test as compared to those in pre-test. The observed performance increase in concrete words in all groups could also be attribute to learning effect, because, in comparison to abstract words, concrete words indeed need relatively limited available contextual information to encode [57][58][59][60] .
For digit-span task, significant improvements were observed in both NBG and MPG, which is consistent with our initial expectation. Forward digit-span test is often used to assess near transfer effects, i.e. short-term memory or WM capacity, in WM training 9,21,22,53-55 . As our results demonstrate, both n-back training and method of loci could effectively promote capacity of short-term memory for digits. Concerning short-term memory, it is particularly worth mentioning that temporary storage is essential for the implementation of dual n-back task, and perhaps, the short-term memory limited capacity can be increased either on the verbal or spatial domain through longtime training 26 . Most studies have documented significant transfer effect of n-back training gains to digit-span scores 18,30,54 . Similarly, our study observed this improvement not only in NBG but also in MPG. In digit-span task, the to-be-remembered items (digits) are presented in a continuous stream, while the implementation of method of loci also focuses on sequence memory, with items put in sequential landmarks when encoding as well as when retrieving them in order during recall phase. Therefore, the significant improvement on short-term memory capacity for digits in MPG may benefit from the repeated training on similar form of sequential memory.
As for the processing speed improvement in change detection task, we only observed a significant improvement on higher load in NBG, but not in MPG. This was somewhat unexpected, given the initially similar spatial memory environment with the method of loci, and hence contrary to our hypothesis. Some studies have reported that in order to acquire substantial transfer effect, the training paradigm and untrained task must share relevant information-processing components and be involved in similar neural substrate 64 . Coincidentally, the central mechanism of process-based training technique aims at producing more stable effects in functions that engage a common neural circuitry 65 . Previous evidence demonstrate that the performance of n-back task recruits the fronto-parietal executive control network 28,48,66,67 , the same neural network also involved in change detection task 56 . More so, recent meta-analytical evidence further indicates that more substantial transfer may only occur when the trained task and transfer task share the same task paradigm 22,68 . In the n-back task, the to-be-remembered items are presented in a continuous stream, and participants must make quick responses to repeated targets according to the task loads. RT represents processing speed, which is one of processing capacities that process-based training hope to enhance 15,27 . Training may increase the efficiency of this process by accelerating responses to repeated target trials. So training-induced improvement in potential ability may be transferred to other ability to detect repeated objects in change detection task that involve similar task demand with n-back task. Demonstrably, these two kind of tasks have much in common either on task paradigm or neural substrate they are involved in.
Conversely, the method of loci has most often been associated with the episodic memory domain, where performance is enhanced by facilitating encoding and retrieval of information of unrelated word pairs 48 . Therefore, the lack of positive transfer to reduced RT may be due to the specificity of this mnemonic strategy with limited applicability to other ability-related domains including processing speed. Rebok et al. 47 also summarized in their paper that training outcomes are highly specific to the ability of mnemonic strategies that have been trained. Moreover, the limitation effect of mnemonic strategy training has been proved to be linked to age with less www.nature.com/scientificreports/ training-induced changes in older adults than in younger ones 46,69 . Precisely, the elderly adults usually encounter great difficulty in applying mnemonic strategy to their daily lives 15 . Generally, our results are consistent with our expectation that n-back training produces relatively prominent transfer effects compared to method of loci on the untrained tasks. However, these findings on one hand, may be partly due to the possibility that the two untrained tasks, one selected as the measure of WM capacity and the other as the measure of processing speed, are more involved in the common processing mechanism, which happens to be the central component that the process-based WM training focuses on. More so, in our study, the two untrained measure tasks seem to be more similar to dual n-back task on paradigm form. If the preceding arguments are valid, this may also reveal one major limitation of our study, that the measures of outcomes might be biased on a single training technique. Therefore, the future studies should bring into consideration the possible influence of biased measure task so as to clearly ascertain whether the training techniques are successfully improving the WM performance. More feasibly, comprehensive assessments covering both training domains should be employed optimally to investigate the training effects of the two memory techniques in the future. On the other hand, mnemonic strategy of method of loci may indeed have a limitation transferring to other situations sharing relevant ability. Although previous findings on memory training in older adults suggest that strategy-based training may be less effective as compared to process-based training technique 15 . Therefore, more evidence including broader age domains may be needed before a general conclusion of the effectiveness between these two training techniques is drawn. Importantly, in the present study, only two most frequently applied training techniques were selected in order to comprehensively compare the effect of the process-based training to that of the strategy-based training. Hence, we also recommend that a variety of training methods, deriving from both process-based and strategy-based training techniques, should be recruited to training experiments in future studies.
In conclusion, the current study recruited three group of participants to investigate memory training effect, and both training groups resulted in great improvement in WM capacity. In particular, n-back training yielded a more prominent transfer of training gains to untrained tasks than training of method of loci. Therefore, it may be recommendable to adopt process-based training for participants with multiple cognitive deficits or persons requiring enhancement of cognitive functions.

Participants.
To determine the sample size, we considered a medium effect size (Cohen's d = 0.52) suggested by Melby-Lervåg and Hulme 70 in their meta-analytic study on the effectiveness of working memory training. Notably, the Cohen's d = 0.52 equates to Cohen's f = 0.26, according to online conversion calculator 71 (Psychmetrica, https ://www.psych ometr ica.de/effec t_size.html). Then we performed a power analysis using G*Power 72 , with alpha = 0.05, power = 0.95, and effect size = 0.26 in ANOVA (between-within interaction). From our analysis, the required number of participants was 63. However, to further ascertain the appropriate sample size, we reviewed literature on the related work. Precisely, a meta-analytic study by Salmi et al. 14 indicate that most working memory training studies recruit between 5 and 50 participants per training group. Therefore, we chose a threshold of 50 participants per group in our study.
Accordingly, one hundred and sixty-three (163), healthy right-handed participants in total (75 female; mean age: 21.22; age range: 18-26 years) were initially recruited from the University of Electronic Science and Technology of China (UESTC). All participants reported normal (or corrected-to-normal) vision, had no history of mental disorders, and gave written informed consent. After undertaking the test of Wechsler Adult Intelligence Scale-Revised in China (WAIS-RC) 73 , and the exclusion of fifteen participants who reported that they may not have enough time to complete training every day, the remaining participants were assigned randomly and evenly across three different groups: 50 (22 female; mean age 21.06 ± 1.932 SD years), 50 (25 female; mean age 21.1 ± 1.961 SD years) and 48 (24 female; mean age 21.38 ± 1.566 SD years) for NBG, MPG, and BCG respectively. Figure 4 shows the flow of the study. Three groups did not show significant difference in demographic characteristics consisting of gender, age and scores of WAIS-RC (all p > 0.2) ( Table 2). This experiment was approved by the local committee for the Protection of Human Subjects of the UESTC and was conducted in accordance with the declaration of Helsinki.
Procedure. All participants were tested with four cognitive tasks (described below) lasting for about two hours at both pre-test and post-test. The training session lasted for 20 days, with four consecutive weeks and five days per week. The training groups spent 30 min each day in the laboratory to complete the corresponding working memory training task (described below), whereas the BCG did not get any training during the 20 days. All participants in both training groups completed more than the required 80% of the training days, with 96% participants completing 19-day and 94% completing 20-day training. All training processes were under supervision of experimenters, who took charge of solving problems of experimental procedures, checking training data every day, and informing participants to make up for the missed training time if necessary.
Memory training paradigms. The dual n-back training task. We used a classic WM training paradigm employed by Jaeggi et al. 18 , see Supplementary Material, Fig. S1A. Two kind of stimuli were displayed in the experiment, with squares (visual stimuli) presented sequentially at eight different locations of the computer screen and consonants (auditory stimuli, eight in total), one at a time in sequence, presented simultaneously through headsets. When one of the presented stimuli matched the one presented n positions back in the sequence, a response should be made, with pressing letter "A" on the keyboard for visual targets and letter "L" for auditory targets. Six visual and six auditory targets appeared randomly in each block, with only two appearing simultaneously in both streams of stimuli. So ten targets in total were presented in each block, consisting of 20 + n trials, The method of loci. We applied the paradigm of method of loci 25. as a strategy-based training technique. This training process included two sessions with 10 days each. The first 10 days were used to introduce the method of loci to participants, and guide participants to remember loci routes with several landmarks, within which they were trained to memorize random words associated with landmarks. Five loci routes of samples were remembered and applied skillfully, and a new loci route based on individual experience that was more suitable for oneself to memorize words was established to be used in the subsequent stage. The second session was to memorize random words using an adaptive vocabulary memory software developed by our laboratory. The initial memory level started with 5 random words, and the duration of memory time limited to 90 s. Each additional level was increased by 5 words, and correspondingly the memory time increased by 90 s, and so on. In the recall stage, participants were required to type the words they had just remembered into the test box in sequence, separated by spaces. If all words entered were absolutely correct and their sequence accurate, the next level began, otherwise the participants continued with the current level. The program recorded automatically the highest daily level achieved by each subject. More than 5000 nouns with the same word frequency selected from online corpus (http://corpu s.zhong huayu wen.org/Resou rces.aspx) were used in the experiment.
Trained memory tasks. Dual n-back task. For this task, we used the same material as the dual n-back training task. The only difference is that the levels of load were fixed to three conditions, i.e. n = 2, 4 and 6. There were three sessions presented in the task, each comprising of three blocks of different loads with sequence in 2-4-6 back, 4-6-2 back or 6-2-4 back.
Words memory task. This paradigm was used to test the performance of words memory 25,52 . The task consisted of two sessions, words encoding session and words recognition session, see Supplementary Material, Fig. S1B. In the first session, a random list of 72 two-character words, including 36 concrete words and 36 abstract words, was presented at a rate of 2 s per word. After continual presentation of 6 words, 20 s were given to participants to recall words presented before or take a temporal rest. Five minutes later after the end of the words memory session, a list of 144 words, consisting of 72 words presented before and 72 new words (half concrete, half abstract words), was presented to participants in a random order. During words recognition session, participants were required to indicate whether the presented word was old or new as soon as possible by pressing one of two buttons on the standard keyboard.  www.nature.com/scientificreports/ Non-trained memory tasks. Digit-span task. The digit-span task (forward version) came from the WAIS-RC 73 . The length of digit-span was 3-12 bit. Participants were asked to verbally repeat lists of numbers in the same order at the end of each auditory presentation. If the participant failed twice at the same number list, the last number list repeated successfully will be recorded as final score.
Change detection task. The paradigm was modified from previous design 56 , see Supplementary Material, Fig. S1C. The disks of 9 colors without repetition were selected as attended items, which might appear randomly in 10 possible locations on an invisible 2 × 5 matrix in both left and right visual fields, with 2, 4 or 6 disks per hemi-field. A memory array consisting of the same number of disks in both visual fields were presented to participants after a prior presentation of an arrow, which instructed participants to memorize the items in the corresponding visual field. After a delay of 1600 ms, the test array were presented, where participants were asked to make a judgement on whether the locations of the disks in the attended hemi-field were the same or different from those in the memory array regardless of object colors. Location repeat condition meant none change of items locations from memory array to test array, while location change condition indicated that one item changed its location in the test array.
Data analysis. Due to problems such as program errors and data transmission during the experiment, some data was damaged and rendered unusable. Therefore, not every cognitive task had 148 pieces of complete data: 139 data for dual n-back task, 147 data for words memory task, 148 data for digit-span task and 142 data for change detection task. All data were analyzed with SPSS Statistics version 21.0 (Inc., an IBM Company). To compare the baseline performance on cognitive tasks among three groups, we employed one-way analyses of variance (ANOVA). To elucidate training efficacy, three-way and two-way repeated measures ANOVA were conducted separately in each cognitive task with group (NBG vs. MPG vs. BCG) as a between-subjects factor and time (pre-test vs. post-test) as a within-subjects factor. For all ANOVA analyses, Greenhouse-Geisser corrections were used for non-sphericity data as needed and Bonferroni corrections were applied for post-hoc multiple comparisons. Partial eta square ( η 2 p ) was treated as effect size and p < 0.05 was considered as statistically significant.