Non-human primates can flexibly learn serial sequences and reorder context-dependent object sequences

Abstract

Intelligent behavior involves mentally arranging learned information in novel ways and is particularly well developed in humans. While nonhuman primates (NHP) will learn to arrange new items in serial order and re-arrange neighboring items within that order, it has remained contentious whether they are capable to re-assign items more flexibly to non-adjacent serial positions. Such mental re-indexing is facilitated by inferring the sequential structure of experiences as opposed to learning serial chains of item-item associations. Here, we tested the ability for flexible mental re-indexing in rhesus macaques. Subjects learned to choose five objects in a predetermined sequential order. A change of the background context indicated when the object order changed, probing the subjects to mentally re-arrange objects to non-adjacent positions of the learned serial structure. Subjects successfully used the context cue to pro-actively re-index items to new, non-adjacent positions. Mental re-indexing was more likely when the initial order had been learned at a higher level, improved with more experience of the re-indexing rule and correlated with working memory performance in a delayed match-to-sample task. These findings suggest that NHPs inferred the sequential structure of experiences beyond a chaining of item-item associations and mentally re-arrange items within that structure. The pattern of results indicates that NHPs form non-spatial cognitive maps of their experiences, which is a hallmark for flexible mental operations in many serially ordered behaviors including communication, counting or foraging.

Fig 1. Learning 5-object sequences.

(A) Each trial presented six objects. Monkeys learned to touch five objects in a pre-determined order A-B-C-D-E and avoid a distractor object. A correct choice led to visual feedback (yellow halo) and incremented the slider position of a progress bar on top of the screen. Monkeys received fluid reward upon completion of the sequence. (B) Each sequence was presented in 15 separate trials on the same ‘context 1’. Then the context changed and the object position 2 (object B) and 4 (object D) swapped (red font). After 5–6 “context 1 – context 2” pairs subjects performed a delayed match-to-sample (DMTS) task for 120 trials. Then, sequence learning was performed again with approximately 3 new and approximately 3 repeat-sequence pairs from earlier in the session. Sequences were presented on rock- or sand- themed contexts. (C) Prop. of correct choices for each ordinal position in context 1 averaged across trials 12–15. Darker dashed line indicates forward-looking chance level, i.e., assuming subjects do not consider reselecting previously correct objects. Lighter dashed line is absolute chance levels (1/6). Error bars are 95% conf. int. (D) Avg. rate of completing the 5-object sequences. Symbols mark the first trial (±SE) at which participants completed a sequence on average in 80% of all trials, with this level of performance sustained for at least three consecutive subsequent trials. (E) Avg. completion rate for the top third of ‘good’ sequences with lowest error rate (<5.33 errors) and for the bottom third of ‘poorer’ performed sequences (>6.6 errors). Red and blue triangles represent the mean trials to reach completion (±95% CI) for good and poor blocks, respectively, considering only blocks that reached completion threshold. The data underlying this figure can be found in the S1 Data file.

Fig 2. Subjects swap objects between non-adjacent ordinal positions.

(A) Context 2 swapped the ordinal position of object B and D. (B) Probability of choosing objects immediately after erroneously choosing object B in context 2: for A: 0.007 ± 0.003 (95% CI); B: 0.004 ± 0.002; C: 0.356 ± 0.020; D: 0.450 ± 0.022; E: 0.130 ± 0.014; Distractor: 0.054 ± 0.010. Three stars denote p < 0.001 difference (Welch’s t test with Bonferroni correction). (C) Choosing object D after an error on B (‘retro-active swapping’) was more likely when the context 1 sequence was learned better (blue: one third of good performed sequences with low (≤5.33) errors per trial) than when they were learned poorly (red: one third of sequences with highest number of (>6.6) errors per trial). (D) Choice probability for the first touched object after correctly choosing A in context 1 (upper), context 2 (lower) in trials prior to reaching the 80% completion rate. (E) The difference of the upper and lower panels in D. Subject more likely chose D over C and B in context 2. (F) Choice probability for the first touched object after correctly choosing D in context 2. Stars denote significance levels (Welch’s t test with Bonferroni correction). Error bars are SEs. The data underlying this figure can be found in the S1 Data file.

Fig 3. Bayesian analysis quantifies the choice strategies of individual subjects during sequence re-ordering.

(A–D) Posterior estimations of choice probability of choosing object B, C and D after choosing object A in context 2. (E–H) Estimated posterior distribution differences of the sequence of choices of object D to B and D to C. All subjects showed above 0 for the choice sequence D→B, two subjects showed above 0 for the choice sequence D→C. (I) Proportion of the use of different strategies in context 2 at the second ordinal position across sessions. The serial inference strategy correspond to P(C) > P(D) and P(C) > P(B); the swapping (position coding) strategy corresponds to: P(D) > P(B) and P(D) > P(C); other strategies were not evident in any session but could have included: P(B) > P(C) and P(B) > P(D) and P(C) > P(D). Three subjects (subjects J, K, and S) showed predominantly the swapping strategy (position coding), while one subjects (subject B) used the serial inference strategy (Subject J: position coding: 62.4%, serial inference: 37.6%; Subject K: position coding: 99.6%, serial inference: 0.4%; Subject S: position coding: 100.0%, serial inference: 0.0%; Subject B: position coding: 0.1%, serial inference: 99.9%). Other strategies were not evident in any subject. The data underlying this figure can be found in the S1 Data file.

Fig 4. Subjects pro-actively re-index object order prior to choice feedback when sequences were learned well.

(A) Choice probabilities for the first choice in the last trial of context 1 (after choosing object A). (B) Choice probabilities for the first choice of the first trial in context 2 (after choosing object A). Object D was chosen more frequently than B. (C) Difference in choice probabilities between the first choice of the first trial of context 2 and the last trial of context 1. Stars denote p < 0.05 (*), p < 0.01 (**), and p < 0.001 (***) (Welch’s t test with FDR correction). (D–F) Same format as A–C for the third of blocks with the best performance in context 1 (<5.33 errors per trial, top two-thirds of blocks). (G–I) Same format as A–C for the third of blocks with the poorest performance in context 1 (>6.6 errors per trial, bottom third of blocks). (J) Trial-by-trial differences of the ‘first choice’ choice probabilities after subjects chose correctly object A context 2. The blue line shows the difference of the likelihood to choose D rather than B (‘D–B’), while the other lines show the difference of the likelihood to choose C rather than B (grey: ‘C–B’) and to choose rather than C (yellow: ‘D–C’). Object B was less likely chosen than C and D as first choice in the first trial of context 2, and object C was less likely chosen than D starting from the seventh trial. (K) Choice probabilities of the seventh trial in context 2 across all blocks. Object D was chosen more frequently than object C. The data underlying this figure can be found in the S1 Data file.

Fig 5. Memory of object sequences improves pro-active and retro-active swapping performance.

(A) Completion rate of sequences before the working memory task (New Early), and of new and repeated sequences performed after the working memory task (New Late, Repeat). (B) Pro-active swapping in context 2 (values above 0 reflect that D is more likely being chosen than B after choosing A): Anticipating that D is in second position in context 2 is similarly evident in all conditions. (C) Retro-active swapping after erroneously choosing object B in context 2 (values above 0 reflects that D is more likely being chosen than C after choosing B): Difference of choosing object D vs. C. (D) Better performance in context 1 (blue vs. red: higher vs. lower completion rate) is associated with higher likelihood of pro-actively swapping object D and B in context 2. (Poor performed sequences: M = 0.1471, 95% CI [0.1201, 0.1740]; well performed sequences: M: 0.2102, 95% CI [0.1964, 0.2240]; p: 5.2 × 10⁻⁵). (E) Regression of the completion rate in context 1 and the likelihood to choose object D immediately after object A in context 2 (β: 0.2136; intercept: 0.0215; R²: 0.0214; p: 3.67 × 10⁻⁸; Cohen’s f²: 0.0220). The different colors represent individual subjects. The dashed black line is the regression calculated across all subjects’ data points. (F) Same format as D for retro-active swapping, i.e., for choosing object D in context 2 at the second ordinal position after erroneously choosing B (poorly performed sequences: M: −0.0194, 95% CI [−0.0999, 0.0611]; well performed sequences: M: 0.1202, 95% CI [0.0809, 0.1594]; t test, p: 0.0024). (G) Regression of context 1 completion rate and retro-active swapping in context 2 (β: 0.3502; intercept: −0.1968; R²: 0.0079; p: 0.0017; Cohen’s f²: 0.0079). The colors are fits to data from different subjects; the dashed black regression line is based on data from all subjects combined. (H) Choosing object D proactively after object A in context 2 was more likely than choosing object B across 124 valid sessions (M: 0.20, 95% CI [0.18, 0.22]; p < 1 × 10⁻¹⁰). The effect increased in strength over sessions (regression β value: 0.0020; p: 0.0030; Cohen’s f²: 0.0773). (I) Choosing object D after an erroneous choice of B in context 2 was more likely than choosing object C on average across 121 valid sessions (M: 0.10; 95% CI [0.06, 0.14], one-sample t test, p = 1.8 × 10⁻⁵). Regression slope (red line) was not significant (β value: 0.0015; p: 0.2559; Cohen’s f²: 0.0110). The data underlying this figure can be found in the S1 Data file.

Fig 6. Working memory performance correlates with pro-active re-indexing performance.

(A) Delayed match-to-sample (DMTS) paradigm. (B) Across sessions, the DMTS accuracy did not correlate with the completion likelihood of sequences. (C) DMTS accuracy correlated with pro-active swapping, i.e., with choosing object D after A in context 2. Black line denotes avg. Linear Mixed Effect model; colors show individual subjects. (D) Contexts facilitated sequence learning: Completion rates were higher for sequences repeated with the same context (n = 669, yellow) than a different context (n = 427, blue): same contexts: Mean ± 95% CI: 0.0868 ± 0.0104; different contexts: 0.0662 ± 0.0111l Welch’s t test: p = 0.0083. (E, F) Same and different contexts in repeated sequences resulted in similar pro-active swapping performance (E: same context: Mean ± 95% CI: 0.0374 ± 0.0279; Different context: −0.0083 ± 0.0286; Welch’s t test: p = 0.3556), and similar retro-active swapping (F: same context: Mean ± 95% CI: 0.0374 ± 0.0279; different context: 0.0083 ± 0.0286; Welch’s t test: p = 0.1225). The data underlying this figure can be found in the S1 Data file.