Development of relational memory processes in monkeys.

The present study tested whether relational memory processes, as measured by the transverse patterning problem, are late-developing in nonhuman primates as they are in humans. Eighteen macaques ranging from 3 to 36 months of age, were trained to solve a set of visual discriminations that formed the transverse patterning problem. Subjects were trained at 3, 4-6, 12, 15-24 or 36 months of age to solve three discriminations as follows: 1) A+ vs. B-; 2) B+ vs. C-; 3) C+ vs. A. When trained concurrently, subjects must adopt a relational strategy to perform accurately on all three problems. All 36 month old monkeys reached the criterion of 90% correct, but only one 24-month-old and one 15-month-old did, initially. Three-month-old infants performed at chance on all problems. Six and 12-month-olds performed at 75-80% correct but used a 'linear' or elemental solution (e.g. A>B>C), which only yields correct performance on two problems. Retraining the younger subjects at 12, 24 or 36 months yielded a quantitative improvement on speed of learning, and a qualitative improvement in 24-36 month old monkeys for learning strategy. The results suggest that nonspatial relational memory develops late in macaques (as in humans), maturing between 15 and 24 months of age. Copyright Â

Introduction

It is widely agreed that memory is not a unitary process, but can be dissociated into multiple, somewhat overlapping, neural circuits. One way of revealing these different circuits has been the use of selective damage to various structures in the brain. Another is to study the developmental changes in cognitive processes they subserve.

A key branch of work in our laboratories has been to outline cognitive development in nonhuman primates to determine when the brain areas that support specific memory abilities mature. Work from several groups including our own has yielded evidence that the memory abilities supported by the medial temporal lobe structures do not show a single pattern of development. In monkeys, for example, object recognition memory and the ability to learn simple discriminations is present in the first months of life and is differentially sensitive to medial temporal lobe damage, with that sensitivity increasing as animals approach adulthood. By contrast, tasks requiring greater cognitive demands (e.g., spatial navigation/relational memory) emerge between 12 and 24 months and mature over many years (see for discussion Zeamer et al., 2010a, Bachevalier, 2013) and show sensitivity to hippocampal or other medial temporal cortical damage whether it occurred in infancy or in adulthood (Alvarado and Bachevalier, 2005a, Alvarado and Bachevalier, 2005b; Bachevalier and Nemanic, 2008, Blue et al., 2013, Glavis-Bloom et al., 2013, Nemanic et al., 2004, Zeamer et al., 2015), although performance deficits may only appear in adolescence or adulthood. This finding suggests either progressive development of hippocampal subfields (Jabès et al., 2010), and/or prolonged maturation of hippocampal connections (e.g. to/from prefrontal cortex). This protracted maturation is similar to that seen in young humans, with recognition (semantic) memory appearing early, but episodic, spatial and relational memory maturing over the first 5–7 years of life (see Bachevalier and Vargha-Khadem, 2005, Overman et al., 1996a, Overman et al., 1996b, Overman et al., 2013, Rudy et al., 1993).

In recent studies we showed that in monkeys tested at 6–8 months, 18 months, and 5–6 years of age, delay-dependent recognition memory and spatial recognition memory emerged at the 18 months of age, whereas spatial relational memory (in this instance, the ability to recognize changes in the visual arrangement of stimulus arrays) was evident only at the 5–6 year testing age (Blue et al., 2013, Zeamer et al., 2010b). Similarly, neonatal hippocampal lesions did not impact spatial recognition or delay-dependent recognition until 18 months, when presumably this structure was functional in control monkeys (Blue et al., 2013, Zeamer et al., 2010b). Interestingly, neonatal hippocampal damage impacted performance on spatial relational recognition memory when tested in adulthood (Blue et al., 2013). These findings suggest that in normally-developing monkeys, hippocampal contributions to memory become necessary between 6 and 18 months, but reach full maturation sometime after. Having also shown that nonspatial relational memory is also sensitive to neonatal hippocampal damage in adult monkeys (Alvarado et al., 2002) as well as in adult monkeys with hippocampal damage (Alvarado and Bachevalier, 2005a), we were interested to see whether nonspatial relational memory shows a similar protracted developmental profile. Thus, to explore the development of nonspatial relational memory in young rhesus macaques, we compared the ability of young rhesus monkeys to learn the transverse patterning problem during the first 3 years of life.

As first described by Spence (1952), the transverse patterning problem comprises 3 concurrent discriminations with overlapping stimuli (see Fig. 1). The three problems take the following form: Problem 1 A+ vs. B− (where + and − indicate the correct and incorrect choices respectively), Problem 2 B+ vs. C−, and Problem 3C+ vs. A−. Based on the individual stimuli, solution of this problem seems impossible as it violates the law of transitivity [e.g. “if A > B, and B > C, then A > C”]. Furthermore, each stimulus is rewarded 50% of the time. In this task, the meaning of each stimulus depends upon which other stimulus is also present. By using a relational or configural solution, the problem can be solved by responding to the unique relationship between each pair [e.g. “If A & B, pick A”] or a directional response to the configural cue [if , go left; if go right]. It is important to emphasize that only when all three discriminations are performed within the same session, is a configural or relational solution necessary for performance1. By contrast, any two problems (plus a third that maintains a linear order, e.g. C+ vs. D−) could be solved using a simple elemental solution, A > B > C > D, in which the associative strength accrued by the two anchor stimuli (A = 100% and D = 0%) allow accurate performance on each discrimination, without requiring relational encoding, although variations of this problem (e.g. “transitive inference”) have been widely used to study the development of relational learning in children, it is important to note that in both animals and humans (Frank et al., 2005), performance is possible without the use of logical inference.

Fig. 1

Transverse Patterning Problem: Example stimuli forming the transverse patterning problem. Three objects, a ball (“A”), a duck (“B”) and a bow (“C”) in the example shown, can be presented as discrimination pairs. Lower panel, three example trials (30 trials per session): Pairs were placed over food wells, with the correct object covering a food reward. Choosing the correct item (indicated in the figure by a “+”) reveals the reward. Discrimination pairs are presented with the correct object equally often on the left or right side for 30 trials in a session.

Numerous groups have explored learning of transverse patterning in a variety of species as well as after MTL damage in animals (Alvarado and Rudy, 1995, Alvarado et al., 2002, Alvarado and Bachevalier, 2005a, Alvarado and Bachevalier, 2005b, Driscoll et al., 2005, Rondi-Reig et al., 2001) and humans (Hanlon et al., 2003, Hanlon et al., 2011, Rickard and Grafman, 1998, Reed and Squire, 1999). Our own work has suggested that in adult nonhuman primates or rodents, performance on this task is sensitive to medial temporal lobe damage, in particular to hippocampal damage, although variations of the task have suggested a role for other MTL areas (Alvarado and Bachevalier, 2005b, Saksida et al., 2007).

In keeping with the above results and what is known about the development of medial temporal lobe structures in mammals, performance on transverse patterning in human children is late-developing, with children under 5 years of age unable to solve the task (Rudy et al., 1993). Thus, if protracted MTL maturation is the basis for delayed proficiency on this task, and if MTL functional maturation lasts the first 18 postnatal months in rhesus macaques, then we would expect young monkeys to improve acquisition and performance of this task with age and particularly achieve a high level around 18 months (based on observations from Blue et al., 2013 and Zeamer et al., 2010b). Similarly, we would expect to see a switch in strategy during this period from elemental to relational solution that will manifest in equally high performance on all three problems.

To that end, we compared 18 young macaques ranging in age from 3 to 36 months at the start of the study. The specific age groups of 3, 6, 12, 24 and 36 months were chosen based on previous work in our group that has looked at the development of various memory abilities at these ages. In addition we had three naive available subjects, two that started training at 4 months old and one that started at 15 months. They were trained on the transverse patterning problem and compared in the ability to solve the 3 concurrent discriminations and the strategy used to perform the task. In particular, we were interested to learn at what age the monkeys were able to perform equally well on all 3 problems, indicating an ability to use relational memory. Infants failing to learn at a younger age were retrained at an older age to directly observe developmental improvements. Lastly, once they successfully learned the task using one set of objects, they were transferred to a new set of transverse patterning discriminations to assess whether they had indeed learned to use a relational strategy.

Materials and methods

Subjects

Subjects were 18 rhesus monkeys (Macaca mulatta), 9 male and 9 female, experimentally naïve before assignment to this study. Fifteen of them were born in the NIH Veterinary Resources Branch and were raised in the primate nursery of the Laboratory of Neuropsychology (LN), NIMH, and three monkeys were brought to the LN at the age of 3 years. The monkeys were assigned to groups according to the age at which they began behavioral training. As listed in Table 1, the initial age groups were as follows: 3 months (n = 4), 4–6 months (n = 4), 12 months (n = 3),15–24 months (n = 4), 36 months (n = 3). Additionally, a number of subjects who failed to learn at a younger age were either retrained at 12, 18, 24 or 36 months, or were trained on a new transverse patterning set (transfer) if they had learned the task. Two subjects were assigned to another project between training ages (S1 & S2) and so were not tested further, and one subject (S7) became ill for unrelated reasons and so was released from the study after his first retraining.

Table 1

Training, retraining and transfer testing ages for each subject. Some subjects were reassigned to other projects and so may not have been retrained or transferred (see text). All subjects reaching criterion at any age proceeded to the transfer test.

Subjects testing age and participation
Testing Age in Months
Subject	TrainingSet 1	RetrainingSet 1	RetrainingSet 1	TransferSet 2
S1	3	–	–	–
S2	3	–	–	–
S3	3	12	–	–
S4	3	15	–	16
S5	4	12	24	24
S6	4	12	24	24
S7	6	12	–	–
S8	6	–	–	–
S9	12	24	–	24
S10	12	24	–	–
S11	12	24	–	24
S12	15	18	–	18
S13	24	36	–	36
S14	24	36	–	36
S15	24		–	24
S16	36	–	–	36
S17	36	–	–	36
S18	36	–	–	36

Apparatus

Training took place in a Wisconsin General Testing Apparatus, appropriately sized to the age of the monkey. The testing tray had two food wells, 10 cm apart on center in the middle of the tray within reach of the subject. Sets of three testing stimuli were formed from easily discriminable junk objects (designated by letters A–L). Set 1 (items A–C, see Fig. 1 for example) was used for the initial training on the task and then for the re-training phase for those animals who did not reach criterion during training. Set 2 (items D-F) was used after animal reached criterion on Set 1 (either during training or retraining) and assess whether the animals could transfer the relational rule to a new problem. Within each set, three discrimination problems were created, for example in Set 1: Problem 1, A+ vs. B−; Problem 2, B+ vs. C−; and Problem 3, C+ vs. A−, where (+) designates reward and (−) designates no reward. For each discrimination problem, the pairs of junk objects were placed over the food wells, with each object appearing equally often over the left or right well. The correct object for each pair hid a food reward which the animal retrieved after displacing the correct object.

Behavioral training

Monkeys were trained 5 days a week for 30 trials per day to retrieve a food reward (e.g. peanut, raisin, grape, banana pellet etc.). Food rewards were selected for each subject to be highly motivating for behavior, thus different subjects worked for their preferred reward. Because these were growing monkeys, there was no food restriction during testing, rather they received their daily feed following testing to maximize motivation. Subjects were weighed weekly and monitored for normal growth.

Pretraining

Subjects were acclimated to the testing apparatus and shaped to displacing objects to obtain a hidden food reward. Once they reliably displaced objects, they moved to the transverse patterning pretraining.

Transverse patterning pre-training (Phases 1–2)

Training took place progressively, as adapted from procedures described in Alvarado and Rudy (1992) and Alvarado et al. (2002). As these were young, naïve monkeys, we began with a single set of objects (Phase 1: Problem 1. A+ vs. B−) for a few days to give them a chance to experience the problem contingencies. After a few days, they were introduced with trials of Problem 2 interleaved in block fashion with continued presentations of Problem 1 (Phase 2: Problem 1. A+ vs. B−; Problem 2. B+ vs. C−). Again, this training continued for a brief time so as not to push the animal towards a specific (e.g. elemental) strategy that could impact their ability to solve the full set. Indeed, it is important to note that the first two Phases do not require a relational solution for high level performance.

Transverse patterning training (Phase 3)

The full transverse patterning problem was achieved in Phase 3 with the addition of Problem 3 (C+ vs. A−). At this point, a relational or configural solution is required to perform better than chance on all three discrimination problems. Training was as follows: the first session of Phase 3 begins with 5 trials of Problem 1, followed by 20 trials of Problem 3, and then 5 trials of Problem 2. The second session of this phase presents each problem in blocks of 5 trials (1, 2, 3, 1, 2, 3). For all subsequent sessions the problems were intermixed with 10 presentations of each problem, presented pseudorandomly and switching every first, second or third trial (i.e., no more than three sequential presentations of a single problem). Training continued until the subject reached a criterion of 90% correct overall, and 80% or better on each individual problem, or until a maximum of 2010 trials was reached. Those who reached criterion proceeded to Transfer training (see below). Those failing to reach criterion were put on rest until they reached the next yearly age when they were retrained on Set 1.

Retraining

As shown in detail in Table 1, subjects who failed to reach criterion on Phase 3 for the initial training Set 1 were retrained on that set at an older age. That is, those initially trained at 3–4 months who failed to learn the task were retrained at 12 months of age (except S4 who received that retraining at 15 months). For the two subjects initially trained at 6 months, S7, was also retrained at 12, and S8 was assigned to another study in the interval and so did not participate further. Those failing initially, or after retraining at 12 months were trained again at 24 months (except S7, who became ill after his 12 month retraining and was released from the study). Subject S12 who failed to learn when trained at 15 months was retrained at 18 months of age. Lastly, those failing at 24 were retrained at 36 months. This retraining allowed us to examine the effects of practice (e.g. trained at 12 months vs. retrained at 12 months), and ensure that early failure did not result from reasons other than immaturity. If they succeeded on retraining, they then proceeded on to Transfer training Set 2.

Transfer

Those subjects who successfully learned Set 1, Phase 3 were re-tested on a novel set of stimuli (Set 2) to see whether they had learned a rule/strategy, or if the performance was set-specific. Training on Set 2 followed the same protocol as for Set 1.

Data analysis

All statistical analyses were conducted using IBM SPSS Statistics for Macintosh, Version 22.0. Data were analyzed using analyses of variance (ANOVAs) with repeated measures across Problem or Rank where appropriate. Where data were normally distributed, one-way ANOVAs were used to evaluate learning ability at each age, comparing Trials and Errors to criterion by Age. If data were skewed, appropriate nonparametric comparisons were made (Kruskal-Wallis). Performance accuracy was compared as an overall percent correct across each Age. The strategy used to perform the task was determined by looking at performance across each problem. For repeated measures ANOVAs, Hyunh-Feldt corrections were used when data violated Mauchly’s Sphericity tests. Lastly, following an initial examination of the data, it was evident that the two 4-month-old subjects performed quite differently from the 3-month group, and similarly to the 6-month group, so we combined them. Similarly, the one 15-month-old subject was combined with the 24-month group (see Table 2 and Fig. 2). Thus for all analyses, we maintained this modified grouping.

Table 2

Trials and Errors to Criterion, and final Performance for Phase 3 Acquisition. Averages shown by age, however performance patterns suggested the following groupings for the analyses: 3 months; 4–6 months; 12 months; 15–24 months; 36 months.

Acquisition of Phase 3
Subject	Age	TTC	ETC	Performance
S1	3	2010	924.5	51.67
S2	3	2010	1029.67	55.00
S3	3	2010	870	46.67
S4	3	2010	1152.5	60.00
Avg.		2010	994.12	53.33
S5	4	2010	623	81.67
S6	4	2010	598	81.67
Avg.		2010	610.5	81.67
S7	6	2010	671.5	81.67
S8	6	2010	858	73.33
Avg.		2010	764.75	77.5
S9	12	2010	842	85.00
S10	12	2010	524.5	75.00
S11	12	2010	657	76.67
Avg.		2010	674.5	78.89
S12	15	780	224	93.33
S13	24	2010	717.5	75.00
S14	24	2010	491	76.67
S15	24	1350	424	93.33
Avg.		1790	544	81.67
S16	36	1440	552	90.00
S17	36	750	242	96.67
S18	36	1140	412	91.67
Avg.		1840	402	92.78

Fig. 2

Training: Ranked Performance in Phase 3. Group averaged performance (±SEM) on each discrimination is shown, ranked by problem with Best (black bars), Intermediate (grey bars) and Worst (white bars) performance level. Dashed lines indicate chance performance (50%), minimum criterion requirement per problem (80%) or criterion level averaged across the three problems (90%).

Results

Phase 3 acquisition

Table 2 shows the trials and errors to criterion for all subjects in Phase 3 for each animal as well as their achieved performance level over the last 4 days of training, averaged across the three problems (criterion was 90% correct). As shown in Table 2, all of the 36-month-old subjects reached criterion on Phase 3. Those at 24 months of age showed a mix of success, with the 15-month-old and one 24 month-old reaching criterion. By contrast, no subject under 15 months of age at the start of training was able to reach criterion. Because of the ceiling effect for trials to criterion for the younger subjects, nonparametric comparisons for Trials to criterion were made using a Kruskal-Wallis H test, which showed a statistically significant effect of Age, χ2(4) = 11.89, p = 0.018, with a mean rank of 3 for the 36-month group, 7.5 for the 24-month group, and 12 for the remaining younger groups. Pairwise comparisons however, revealed a difference between the 36 month group with the 6 and 3-month groups only, χ2(1) = 11.89, p = 0.018 An analysis of variance (ANOVA) comparing the effects of Age on Errors to criterion revealed a main effect of Age for Errors to criterion [F(4,13) = 8.79, p = 0.001]. Post hoc comparisons (Tukey) revealed that for Errors to criterion, the 24 and 36 month old subjects differed from the three younger groups (p’s < 0.001) but not from each other.

Phase 3 performance

As shown in Table 2, performance at the end of acquisition for the 36-month-olds was identical to fully mature adults (Alvarado and Bachevalier, 2005a, Alvarado and Bachevalier, 2005b), with an average performace of 90% across the three problems. Half of the 15–24-month group reached criterion, and the remaining subjects performed at a high level. The 12 and 4–6 month groups performed less strongly, but still at about 80% on average. By contrast, the youngest group, starting at 3 months of age barely exceeded chance levels, which is consistent with their responding to the reward contingencies of each individual stimulus (i.e., 50%). Analyzing the effects of age on Performance levels, a one-way ANOVA revealed a significant effect of Age on overall performance level achieved for each group [F(4,13) = 19.86, p = 0.0001]. Post hoc comparisons (Tukey) revealed that the 3-month group differed from the remaining groups (all p’s < 0.001), but there were no other reliable differences.

Performance strategy analysis

For the older animals who did not reach criterion, it was important to determine whether their performance reflected a) general immaturity (e.g., a generalized tendency towards errors), as opposed to a lack of relational memory ability b) a tendency to respond to the elemental reward contingencies of the stimuli (50%), or c) the use of a strategy that we have observed in adults unable to solve the task. This “linear” strategy involves the animal selecting an anchor object, for example A, that becomes the preferred object of choice, and another, for example C, to which it rarely responds. So, for example, if the subjects were using an elemental strategy, they could perform well on two problems (>90%) and below criterion on the third (60%) while still maintaining 80% averaged performance. By contrast, those using a relational strategy should perform fairly equally and at a high level across the three problems, as this strategy reduces interference across discriminations. To explore this further, we present the data by ranked problem in Fig. 2. That is, for each animal we ordered the performance by their Best, Intermediate, and Worst problem performance, rather than by specific problem, as preference differed across subjects. Thus, if Subject 1 performed best on Problem 3, and next best on Problem 1, his scores would be ranked 3,1, 2. Subject 5 might have the problems ordered as 1, 2, 3. Presenting performance in this way may reveal a clearer picture of the strategy used to solve the task (see Discussion in Alvarado et al., 2002).

As shown in Fig. 2, there was an age-related shift in performance across the problems. At 36 months, the adult pattern of relational performance is apparent in all subjects. Starting at least by 15 months of age, there is a shift in both ability to reach criterion on the task and the strategy used to solve it. Although not all subjects demonstrated the use of a relational solution, monkeys between 15 and 24 months of age were capable of solving the problem in an adult-like way, though some still chose the elemental solution. By contrast, the 12 month old group demonstrated the elemental pattern of discrimination, rather than a less accurate use of a relational strategy. Those in the 4–6 month group were relatively even, across the three problems, but perform at a low enough level to suggest that they were not consistantly choosing a preferred problem and/or are not using a relational solution in an adult-like way. Lastly, subjects in the 3 month old group barely exceeded chance at the end of training; their choices directly reflecting the associative strength of the individual objects (50% reinforcement overall). Despite the apparent performance difference, an Age X Rank ANOVA with repeated measures revealed, the aforementioned effect of Age [F(4,13) = 19.86, p = 0.001], and of Rank [F (1.84,26) = 38.3, p = 0.001], but no interaction. Post hoc (Tukey) tests of the Age effect revealed that the 3 month group differed from all others (p’s < 0.05), but the other groups did not.

Subjects who solved the task were tested on a new set of discriminations to see how well the relational rule transferred to new items (see Transfer, below). Those who failed to solve it were retrained 6–12 months later (see Retraining). Two subjects in the 3 month group (S1 & S2) had been reassigned to other projects in the interim and so did not participate further in this study.

Retraining

Table 3 shows the trials and errors to criterion and overall performance for Phase 3 retraining for the subjects who failed to learn Phase 3 initially. Fig. 3 shows their ranked performance by problem. The effects of prior training were revealed in the animals retrained at 36 and 15–24 months. All subjects in those groups reached criterion in fewer trials and with fewer errors than naïve animals trained at those ages, suggesting that certain aspects of their performance reflected inexperience and immaturity. By contrast, only one of the 12 month old subjects reached criterion, despite prior experience. Two-way ANOVA’s for the effects of Age X Training X Trials or Errors to criterion confirmed improved performance by Age [F(2,16) = 9.387, p = 0.002] and by Training [F(1,16) = 7.783, p = 0.013], but no interaction for Trials, and a main effect of Age [F(2,16) = 8.964, p = 0.002] and of Training [F(1,16) = 7.830, p = 0.013], but no interaction for Errors. Post hoc (Tukey) examination of the Age effect revealed that the 12 month group differred from both the 24 (p = 0.01 Trials; p = 0.008 Errors) and 36 (p = 0.004 Trials; p = 0.006 Errors) month groups, which did not differ from each other.

Table 3

Trials, Errors and Phase 3 Performance at criterion when retrained on Set 1 at a later age.

Retraining: Acquisition of Set 1
Subject	Age	TTC	ETC	Performance
S3	12	2010	782	65.00
S5	12	2010	621	75.00
S6	12	2010	643	78.33
S7	12	780	249	90.00
Avg.		1702.5	574	77
S4	15	810	282	93.33
S5	24	540	190	93.33
S6	24	660	232	96.67
S9	24	1170	338	90.00
S10	24	720	211.5	88.33
S11	24	1200	361	91.67
Avg.		850	269.1	92.2
S13	36	270	47.5	95.00
S14	36	720	180.5	90.00
Avg.		495	114	92.5

Fig. 3

Retraining: Ranked Performance in Phase 3 by age at retraining. Note differing pattern of performance between 12-month-old subjects and the two older groups. This pattern is similar to that shown during Training phase for the same ages. All conventions as in Fig. 2.

However, despite the advantage in acquisition, performance seemed to be more dependent on age than experience, as illustrated in Fig. 3. The pattern of ranked performance following retraining was quite similar at each age to those shown in Fig. 2. That is, 36 month old subjects used a relational solution, performing equally well across all three problems. All subjects in the 15–24 month group consistently used a relational solution, with all subjects performing at better than 80% on average across the three problems. By contrast, only one monkey retrained at 12 months performed at this level; the remaining subjects in this group performed similarly to naïve 12-month-old monkeys, performing well on two problems and poorly on the third which is typical of a ‘linear’ elemental strategy. Indeed, a three-way ANOVA with repeated measures comparing the effects of Age X Experience (i.e., training vs. retraining) X Rank indicated a main effect of Age [F(2,16) = 6.753, p = 0.007] but no effect of Experience. There was the expected effect of Rank (because the problems were ordered) [F(2,32) = 50.093, p = 0.001], and a Rank X Age interaction [F (3.75,32) = 3.93, p = 0.012]. No other interactions were reliable.

To further explore the interaction of Age and Rank, we focussed our analysis on the Worst performance, as it is this level which determines whether the subject is using a relational or linear strategy. A one-way ANOVA comparing the effects of Age x Worst performance level for the three older groups, regardless of Experience confirmed the effect of Age [F(2,19) = 16.325, p = 0.001]. Post hoc comparisons (Tukey) across the three ages confirmed that the performance of the 12 month group differed significantly from the two older groups (p’s = 0.001). The two older groups did not differ from each other (p = 0.614).

Those subjects who learned either in the training or retraining phases were then immediately transferred to a new set of transverse patterning discriminations to assess how well the rule transferred (note, unfortunately S7 became ill at this point and was released from the study). Table 4 and Fig. 4 show the results of the transfer test. All subjects (ages 24 and 36) reached criterion at equal or fewer trials and errors to criterion as their successful Set 1 acquisition. More importantly, all subjects showed the adult pattern of relational strategy use on the new set (Fig. 4). The improvement in performance on Set 2 was confirmed by a Set x Age ANOVA with repeated measures for the factor Set. For Trials to criterion, the results showed a main effect of Set [F (1,8) = 9.32, p = 0.016] that did not vary across age. Similarly, subjects committed fewer errors on Set 2, showing a main effect of Set [F (1,8) = 8.93, p = 0.017].

Table 4

Trials, Errors and Performance for Phase 3 on the Set 2 Transfer test.

Transfer to Set 2
Subject	Age	TTC	ETC	Performance
S4	15	900	342.5	93.33
S5	24	300	79	96.67
S6	24	690	206.5	91.67
S9	24	720	192	91.67
S11	24	510	150	91.67
S12	18	240	62.5	90.00
S15	24	450	105	91.67
Avg.		544.3	162.5	92.4
S13	36	180	27.5	90.00
S14	36	360	103	96.67
S16	36	390	97.5	93.33
S17	36	690	174	91.67
S18	36	450	153.1	90.00
Avg.		414	111.02	92

Fig. 4

Transfer: Phase 3 Ranked Performance on Set 2 for subjects who reached criterion during Training or Retraining. All conventions as in Fig. 2.

Overall performance analysis

Because the performance results for animals aged 12–36 months were so consistent regardless of amount of training, we averaged their performance scores obtained on Phase 3 during Training (Fig. 1), Retraining (Fig. 2), and Transfer (Fig. 4) across the 3 problems and made an overall comparison of Age x Performance on the Worst problem. This analysis yielded an effect of Age [F(2,31) = 25.595, p = 0.001. Post hoc (Tukey) tests confirmed that subjects trained at 12 months performed worse that both the 24 month group (p = 0.001), and the 36 month group (p = 0.001), whereas the 24 month group did not differ from the 36 month group (p = 0.889).

Discussion

The results showed a developmental progression in the ability to solve the transverse patterning task, as well as an interesting shift in the use of alternative strategies. The performance patterns of the youngest groups (3 months old) clearly indicated that their choices were largely tied to the associative strength (reward history) of the individual stimuli, which was 50%. However, many who began training at 4 or 6 six months, adopted a strategy which improved their success, but did not allow them to solve all three problems. This strategy is one we have seen in adult monkeys with hippocampal damage, which will respond as if the problems were linear (e.g. A > B > C) and choose accordingly (Alvarado et al., 2002). Thus, they perform at criterion levels on two problems, but not the third. This same pattern was demonstrated by the 12 month group, with improved success on the two problems, as well as by those subjects that failed to reach criterion in the 15–24 month group. However, it was also the 15–24 month age at which some subjects achieved criterion on the transverse patterning task, and were able to transfer the rule to a new set of stimuli. By 36 months of age, all subjects were able to reach criterion, performing well across the three problems. Interestingly, even when retrained at older ages, only one 12-month-old subject reached criterion, whereas all subjects retrained at 15–24 and 36 months performed at criterion levels and used a relational strategy.

Successful performance on the transverse patterning problem requires relational memory abilities, although there can be alternative strategies that support performance, such as in the rock-paper-scissors game or other semantic cues (Moses et al., 2008). The results of the present study show clearly that the relational learning as measured by the transverse patterning problem is late-developing. Unlike simple concurrent discriminations, which can be performed at a younger age, or the ability to use a nonmatching rule, which is present relatively early, but matures over the first year of life (see Bachevalier, 2013, Bachevalier and Vargha-Khadem, 2005 for review), the capability to use a relational rule does not seem to emerge before 12 months of age at the earliest, with animals older than 15 months of age able to solve the transverse patterning task. However, even at 24 months of age, not all naïve subjects used a relational strategy, whereas all subjects who had prior experience with the task and were retrained at 15 or 24 months did. Their performance improved with age such that at 36 months they learned at a mature level, similar to adults in our previous studies (Alvarado et al., 2002; Alvarado and Bachevalier, 2005). The only other published developmental study using transverse patterning trained similarly to ours was in children. Rudy and colleagues (1993) showed that the ability to learn the transverse patterning discriminations is developmentally delayed in humans, who are unable to solve the task until approximately 5 yr of age (Rudy et al., 1993). Interestingly, both our and Rudy’s results parallel the ability to perform spatial navigation tasks in children (Overman et al., 1996b) and spatial relational tasks in nursery-reared monkeys (Blue et al., 2013).

Given the results from lesion and developmental neuroanatomical studies in the literature, we can speculate as to the neural basis for the prolonged maturation of relational memory abilities. Evidence that this developmental delay reflects maturational processes within the medial temporal lobe in monkeys is provided by studies in which performance on this task is severely impaired by hippocampal lesions. For example, performance was impaired in adult monkeys with either neonatal damage to the hippocampal region (Alvarado et al., 2002), or neurotoxic damage to the hippocampal formation, perirhinal cortex or area TH/TF sustained in adulthood (Alvarado and Bachevalier, 2005a, Alvarado and Bachevalier, 2005b). Interestingly, the animals in each of those studies performed similarly to the 12 month old groups in the present study, regardless of the age at which the lesion occurred. That is, they adopted a performance strategy that treated the individual stimuli in a linear heirarchy (i.e., A > B > C). Similar results have been shown in adult amnesic humans with hippocampal or temporal lobe damage (Rickard and Grafman, 1998, Reed and Squire, 1999), and in rats with neurotoxic damage to the hippocampal formation (Alvarado and Rudy, 1995).

Although the specific structure within the medial temporal lobe supporting the development of relational memory is still under investigation, the fact that performance on tasks requiring perirhinal cortex develop early (Zeamer et al., 2015), and that the effects of early hippocampal damage to memory emerge late, suggests that the likelier candidates would be hippocampus and/or TH/TF in the medial temporal lobe. Furthermore, the finding that relational memory abilities in the spatial or nonspatial domain develop later than those for spatial recognition memory, suggests that additional late-developing structures, or protracted maturation of connections to other structures, contribute to this prolonged maturation. For example, Blue et al. (2013) traced the development of spatial memory, testing at 8, 18 and 60 months of age and found that young monkeys recognized a spatial location by 18 months, but memory for object-place relations was only present at the 60 month testing age (they were not tested on spatial tasks between 18 and 60 months). Interestingly, neonatal hippocampal damage delayed the emergence of spatial recognition, and prevented emergence of object-place relations. This finding might explain some of the differences observed among variations in the task and the effects of specific MTL damage. For example, in our hands, using 3 dimensional objects, damage to the hippocampus impacts performance on transverse patterning, however so do lesions of the parahippocampal or perirhinal cortices (Alvarado and Bachevalier, 2005a, Alvarado and Bachevalier, 2005b). Using different methods and stimuli that emphasize a strictly configural solution with perceptually complex or ambiguous stimuli (e.g. Saksida et al., 2007; see Alvarado and Bachevalier, 2005b for similar discussion) are less (or not at all) affected by hippocampal damage, but are impaired by perirhinal cortex damage.

Similarly, in human participants, there is evidence of hippocampal activation as measured by magnetoencephalography (MEG) during performance of transverse patterning (Hanlon et al., 2003, Moses et al., 2009). Interestingly, hippocampal activity decreased and MTL cortical activity increased when semantic meaningfulness of the cues increased. Furthermore, and pertinent to the present findings, MEG showed that patterns of stronger right hippocampal lateralization, which was the mature pattern of activity, correlated highly with accurate performance on transverse patterning, particularly in younger subjects (11–14 years old) as compared to older teens (15–18) or adults (Hopf et al., 2013).

Given the established role for the hippocampus in the spatial domain, the fact that both spatial and nonspatial relational memory abilities are both impacted by hippocampal damage and yet mature somewhat later than other spatial memory processes suggests a possible role for other later developing structures in the relational memory network. For example, the prefrontal cortex, which is late developing in monkeys and humans (e.g. Fuster, 2002, Goldman-Rakic, 1987, Malkova et al., 2014, Nejime et al., 2015, Tsujimoto, 2008) may contribute to relational memory in the spatial or nonspatial domain. Indeed, electrophysiological studies in macaques showed task specific firing of prefrontal neurons in the dorsolateral and medial prefrontal regions during performance of a transverse patterning task (Nejime et al., 2015), and human subjects showed increases in prefrontal activity with increased ‘meaningfulness’ of stimuli (Moses et al., 2009).

The full map of the relational memory circuit remains to be completed, however the present results, taken in context with our previous work, certainly points to late developing relational memory abilities in monkeys as in humans, and that this late development depends in part on protracted maturation of the medial temporal lobe, and likely on later contributions of the hippocampal-prefrontal network.

Conflict of interest

The authors declare no conflict of interest.