Positional encoding in cotton-top tamarins (Saguinus oedipus)

Strategies used in artificial grammar learning can shed light into the abilities of different species to extract regularities from the environment. In the A(X)nB rule, A and B items are linked but assigned to different positional categories and separated by distractor items. Open questions are how widespread is the ability to extract positional regularities from A(X)nB patterns, which strategies are used to encode positional regularities and whether individuals exhibit preferences for absolute or relative position encoding. We used visual arrays to investigate whether cotton-top tamarins (Saguinus oedipus) can learn this rule and which strategies they use. After training on a subset of exemplars, half of the tested monkeys successfully generalized to novel combinations. These tamarins discriminated between categories of tokens with different properties (A, B, X) and detected a positional relationship between non-adjacent items even in the presence of novel distractors. Generalization, though, was incomplete, since we observed a failure with items that during training had always been presented in reinforced arrays. The pattern of errors revealed that successful subjects used visual similarity with training stimuli to solve the task, and that tamarins extracted the relative position of As and Bs rather than their absolute position, similarly to what observed in other species. Relative position encoding appears to be the default strategy in different tasks and taxa.


Introduction 58
Extracting regularities is necessary to make sense of the numerous stimuli available in the We tested four adult cotton-top tamarins, one male (RK) and three females (RB, SH and 171 EM), housed at Harvard University. All subjects were born in captivity and socially housed, 172 with separate home cages for each breeding pair and their offspring. Subjects were 173 maintained on a diet of monkey chow, fruit, seeds, and mealworms, together with free access 174 to water. Subjects voluntarily left their home cages, lured out by a piece of raisin. 175 They could access two pulling tools presented on an acrylic apparatus (40x50x6 cm) through 179 two small holes. Each tool consisted of a pulling stem and a card covering a tray at the end of 180 the stem. When subjects pulled one of the stems the tool advanced, the card flipped back, 181 presenting either the food reward (a small piece of a Froot Loop© cereal) or nothing at all. 182 Stimuli were presented on a plastic laminated sheet (11.5x7.5 cm) with different, linearly 183 arranged shapes corresponding to the consistent or inconsistent arrays. 184 Consistent arrays followed the A(X) n B rule, for example A 1 X 1 X 1 B 2 . In arrays not consistent 192 with the target rule, the position of A and B tokens was swapped, as in B 2 X 1 X 1 A 1 , or either A 193 or B tokens were not used. X tokens, irrelevant in determining the consistency of the stimuli 194 with the target rule, could vary in size and number, extending or reducing the distance of the 195 dependency between other tokens. 196 Possible arrangements of the A and B tokens are shown in Table 1: 12 patterns are consistent  197 with the target rule A(X) n B and 37 are not consistent with it. Arrangements not consistent 198 with the rule include transpositions, in which Bs are located to the left of As, and 199 substitutions, in which two tokens of the same A (or B) category are present. During the 200 initial training, we used only 6 of the grammatical arrangements and 7 of the ungrammatical 201 arrangements, 2 or 4 X 1 and X 2 tokens, saving the other configurations of stimuli for the tests. 202 The 13 patterns employed in the training are shaded grey in Table 1

Identical Bs
A 2 (X) n B 3 B 3 (X) n A 1 B 3 (X) n B 1 A 2 (X) n B 4 B 3 (X) n A 2 B 3 (X) n B 2 A 3 (X) n B 1 B 3 (X) n A 3 B 3 (X) n B 4 A 3 (X) n B 2 B 4 (X) n A 1 A 3 (X) n B 3 B 4 (X) n A 2 A 3 (X) n B 4 B 4 (X) n A 3 206  Table 1). In Test 2 we introduced four novel X tokens 212 in the central position (we used for example stimuli such as A 1 (X 3 ) n B 2 and A 2 (X 4 ) n A 2 ). In 213 Test 3 and Test 4 we introduced new X tokens in novel positions either internally (as 214 A 3 X 6 (X 1 ) 2 B 4 X 7 and A 3 X 6 (X 1 ) 2 X 7 A 2 in Test 3) or at the edges of the arrays (as 215 determining the rule-consistency of the stimuli and were used to evaluate the information 217 encoded by tamarins during the training. 218 219

Procedure 220
Before starting the training on the A(X) n B rule, each subject was familiarized with the 221 apparatus and the experimental procedure. Both training and test sessions involved the 222 following procedure. The target subject was lured out of its home cage with a piece of food 223 into a transport box and then moved individually to the experimental room for an 224 experimental session. Prior to a trial, and out of view from the subject, the experimenter 225 prepared the appropriate stimuli and reward. The overall sequence of different stimulus 226 pairings, along with the right or left position of each card, was randomized and 227 counterbalanced across trials and within sessions. For each session, consistent cards were 228 equally distributed between the right and left sides of the apparatus and no more than two 229 consistent cards were presented on the same side consecutively. 230 Each test session consisted of 2-6 warm-up trials followed by 12 test trials interspersed with 4 250 trials with stimuli already presented during the previous training. When responsive, each 251 subject ran two experimental sessions per day. To guarantee an appropriate level of 252 motivation, the inter-session interval within a day was at least three hours. The difference 253 between training and test sessions consisted in the novel material presented during test 254

sessions. 255
Monkeys' responses were coded in terms of which array (consistent or inconsistent) was 256 selected on each trial. To move from the initial training stage (Training 1) to the tests, we 257 required subjects to reach a criterion of 40/48 correct trials or better. This corresponds to the 258 cut-off value for a binomial test with alpha=0.05, consisting of 10 out of 12 correct trials over 259 4 consecutive sessions or better. 260 To determine whether subjects could discriminate between novel consistent and inconsistent 261 stimuli, showing generalization, we analyzed (a) the scores of the first 48 trials (4 sessions) 262 and (b) the scores of the first 96 test trials (8 sessions). We ran the analysis on 8 sessions to 263 increase the number of trials and investigate the responses to specific violations. 264 265

Experimental schedule 266
The experimental schedule went through the following stages, summarized in Table 2 Table 1). 276 Test 1 explored the extent to which tamarins generalized the distinction between consistent 277 and inconsistent stimuli to new spatial arrangements of the tokens experienced during the 278 training. We hypothesized that if tamarins had encoded the positional regularity of A and B 279 tokens ("A to the left of B") they should, in the absence of further training, choose consistent 280 combinations (i.e. A(X) n B) more often than inconsistent combinations. 281 Test 2 was identical to Test 1, with the exception that experimental trials were composed with 282 four novel Xs. Test 2 explored whether tamarins could generalize to novel Xs located in the 283 center of the arrays. 284 Subjects responding above chance to Test 1 and Test 2 moved on to Training B, in which we 285 presented a total of 36 sessions to each subject, using the same stimuli presented in Test 1. 286 Test 3 was identical to Test 1, except that new Xs were located at the edges of the arrays, so 287 that stimuli followed patterns similar to XA(X) n BX or XB(X) n AX. Up to this stage, A and B 288 tokens always occupied the edges of the sequence. Two alternative hypotheses could account 289 for tamarins' success until Test 2. Tamarins could have learned, instead of the relative 290 position of A and B tokens, their absolute position with respect to the edges. In that case, they 291 would not have been able to generalize to arrays not containing A or B tokens at the edges, as 292 presented in Test 3. As an alternative hypothesis, if tamarins had learned that A must be on 293 the left with respect to B, their performance should not have been affected by the insertion of 294 novel tokens at the edges. Subjects responding above chance in Test 3 proceeded to Training 295 C (with the same stimuli used in training B) and then to Test 4. 296 Test 4 was identical to Test 1, except that at least one X token was located between As or Bs 297 and the central Xs, so that stimuli followed patterns of the form XA(X) n XB or BX(X) n AX. If 298 tamarins learned that the position of A and B with respect to X tokens were irrelevant, or that 299 the symmetry of A and B with respect to the center of the array were irrelevant, than their 300 performance should not be affected by the insertion of novel Xs adjacent to the central tokens. These results license the conclusion that RK and RB (but not the other two subjects, EM and 328 SH) were able to use the experience gained during Training A to successfully distinguish 329 between novel consistent and inconsistent stimuli. Thus, at least two cotton-top tamarins 330 could learn a positional rule as A(X) n B and generalize this regularity to novel arrangements. 331 The positive performance of RK and RB is noteworthy considering that, during the training, 332 subjects had previous experience with only a small set of stimuli (n=13 token combinations), 333 which then increased to a set of 24 novel exemplars presented in Test 1, and that before Test 334 14 338 339

Training B and Test 2: Novel Xs in the center 344
Before moving to Test 2, subjects that succeeded in Test 1 were trained with 36 sessions 345 identical to those presented in Test 1 (Training B). The success rate over the 36 sessions for 346 each subject was 328/432 correct choices, 76% for RK, and 315/420 correct choices, 75% for 347

RB. 348
In Test 2, we investigated whether RK and RB were able to generalize to new X tokens 349

Test 3 and Test 4: Novel Xs on the edges and between As, Bs and Xs 361
In Test 3 (Figure 3 In Test 4 (Figure 3, green line)

Analyses of responses to inconsistent stimuli 394
We ran further analyses to investigate the individual strategies used by the subjects that 395 succeeded in Test 1 and went through the other test stages, by looking at the pattern of 396 responses to arrangements which were inconsistent with the A(X) n B grammar in the first eight 397 sessions of each test stage. 398 To investigate the presence of difficulties or enhanced performance in the presence of 399 specific tokens we analyzed the responses to each A (Figure 4) and B token ( Figure 5) 400 presented in ungrammatical stimuli. As showed in Figure 4, both RB and RK performed 401 significantly worse with inconsistent arrangements that contained A 3 (RB: chi-training, we had not used the A 3 token in any unrewarded arrangements (see Table 1) to test 404 for generalization to items presented in novel positions. Hence the specific failure with A 3 405 exhibited by both monkeys suggests an incomplete generalization of the positional A(X) n B 406 rule that can be due to the selective inexperience with A 3 as unrewarded token. 407 We did not find learning difficulties with any other token, but the performance of each 408 monkey was enhanced in the presence of some tokens, although these effects were not 409 consistent across subjects. These are the results with A tokens (Figure 5)  human readers, exhibit the transposed-letter effect (see Grainger, 2008), so that their 490 performance was lowered by letter transpositions. In this work, though, it was not clear 491 whether monkeys relied on an absolute or on a relative position strategy. A significant 492 decrease in performance with transpositions mediated by relative-position encoding has been 493 noticed in human literates (Dunabeitia et al., 2014;Grainger, 2008) and in pigeons trained to 494 orthographical discriminations (Scarf et al., 2016). We ran several tests to clarify whether 495 tamarins had encoded the absolute position or the relative position of As and Bs (Versace, 496 2008). In our experiments, the absolute position of a token within an array, such as "A 1 must 497 be located as first token on the left part of the array", is defined independently from the 498 identity of other tokens. On the contrary, its relative position, such as "A 1 must be located to 499 the left of B 1 , B 2 , B 3 or B 4 ", depends on the specific identity and position of other tokens. 500 These two alternative strategies of encoding can be probed changing the absolute position of 501 As and Bs with respect to the edges and the center of the array by inserting novel Xs in 502 different positions within the visual arrays. If during the training tamarins had encoded the 503 absolute and not the relative position of As and Bs, they were expected to fail when the 504 absolute position of As and Bs was changed. On the contrary, tamarins' performance was not 505 disrupted when novel Xs were added in the center of the arrays and when the absolute 506 position of As and Bs was changed with respect to the edges and the center of the array. We 507 can hence conclude that tamarins did not rely on the mere absolute position of As and Bs. The 508 fact that tamarins' performance was not compromised by transpositions suggests that they 509 might have used a visual similarity strategy to solve the task. This conclusion is supported by 510 the fact that both successful subjects had a significantly lower performance with 511 ungrammatical stimuli that contained a token -A 3 -never presented in unrewarded stimuli 512 during the training. 513 Overall, we documented a preferential encoding of the relative position. Preferential 514 encoding of relative vs. absolute position has not been documented only in primates and 515 pigeons. Relative rather than absolute encoding in spatial positions had been previously 516 shown in chicks of the domestic fowl during foraging (Vallortigara & Zanforlin, 1986). In a 517 series of experiments, chicks were trained to discriminate between two boxes according to 518 either their relative position to each other or their absolute position (the position with respect 519 to the geometry of the cage or other features of the environment). When the boxes were 520 located close to each other, learning on the basis of the relative position was faster than 521 learning on the basis of absolute position. The advantage of relative position was reduced located tokens show a preference for relative encoding of visual stimuli presented in 524 simultaneous configurations. It seems that encoding of relative rather than absolute position 525 is the default strategy observed across different species and taxa, irrespectively of the possess 526 of language. Further studies should investigate this phenomenon and clarify its spread, 527 neurobiological and evolutionary basis.