Fish can infer relations between colour cues in a non-social learning task

Abstract

Transitive inference (TI) describes the ability to infer relationships between stimuli that have never been seen together before. Social cichlids can use TI in a social setting where observers assess dominance status after witnessing contests between different dyads of conspecifics. If cognitive processes are domain-general, animals should use abilities evolved in a social context also in a non-social context. Therefore, if TI is domain-general in fish, social fish should also be able to use TI in non-social tasks. Here we tested whether the cooperatively breeding cichlid Neolamprologus pulcher can infer transitive relationships between artificial stimuli in a non-social context. We used an associative learning paradigm where the fish received a food reward when correctly solving a colour discrimination task. Eleven of 12 subjects chose the predicted outcome for TI in the first test trial and five subjects performed with 100% accuracy in six successive test trials. We found no evidence that the fish solved the TI task by value transfer. Our findings show that fish also use TI in non-social tasks with artificial stimuli, thus generalizing past results reported in a social context and hinting toward a domain-general cognitive mechanism.

1. Introduction

Transitive inference (TI) has been proposed to be an important cognitive ability in social contexts of hierarchy formation and maintenance [1]. TI describes the ability to infer relationships between stimuli that have never been seen together [2,3]. Social species organized in stable groups with rank hierarchies benefit from the ability to infer the social status of others indirectly and are able to use TI for this [4–6]. Assessing and remembering the ranks of conspecifics allows for appropriate behaviour during interactions, and thus is an important aspect of social competence [7].

It is currently debated whether the brain mechanisms underlying cognition are organized in domain-specific modules or are domain-general [1,8–11]. If cognition is domain-general, the context in which an ability is used should not matter. Conversely, if cognition is organized in modules, animals should not be able to use abilities evolved in a social context also in a non-social context, and vice versa [11]. The usage of TI in animals has been shown either in a social (birds [4], fish [5,12]) or in a non-social context (insects [13], birds [4,14], mammals [15–18]). Only in a single bird species, the pinyon jay Gymnorhinus cyanocephalus, was the use of TI demonstrated both in a social [4] and a non-social context [19], suggesting that this ability is domain-general in this species. In Tanganyika cichlids, the use of TI has thus far only been shown in a social context. After witnessing contests between several dyads of conspecifics, experimental individuals were able to infer their dominance status even if they had never observed them interacting directly (Astatotilapia burtoni [5], Julidochromis transcriptus [12]). If we demonstrate that Tanganyika cichlids can use TI in a non-social context, this will support the existence of domain-general cognitive abilities in these fish.

Previous studies suggested two ways to solve a TI task: by logical reasoning [2,20] or by associative learning and value transfer (VT) [1,2,14,21,22]. Logical reasoning has been the main explanation for the human ability to solve TI [2], however, it is difficult to show this in non-human animals. Alternatively, the Value Transfer Theory predicts that animals evaluate a given stimulus depending on the value of a second stimulus presented simultaneously: if a rewarded stimulus A is presented with a second non-rewarded stimulus B, then B will acquire a secondary positive value from A. This means the positive value of A is transferred to some degree to B although B is unrewarded [14,21].

We investigated whether cooperatively breeding cichlids, Neolamprologus pulcher, can infer transitive relationships between non-social stimuli. It is likely that N. pulcher can infer social rank by using TI as shown in other socially living cichlids [9,10]. Neolamprologus pulcher live in groups composed of a dominant breeding pair and subordinate helpers [23], structured by a sized-based linear hierarchy [24]. Finding and maintaining a stable position in the hierarchy is a crucial social skill for N. pulcher, as contests with group members may result in eviction from a group territory imposing a high cost on the evictee [25,26]. In this study, we aimed to test if N. pulcher might use TI in a more general context and can infer the relative value of artificial stimuli in a non-social context. To test this, we trained N. pulcher to solve a colour discrimination task using a well-established conditioning learning paradigm [27–29]. The fish learnt to discriminate four pairs of colours and were trained the five-term series A + B−, B + C−, C + D− and D + E− commonly used in TI experiment (reviewed in [2]; letters are arbitrarily assigned to five stimuli that are rewarded (+) or unrewarded (−)). If N. pulcher can perform TI, we predict that they (1) order the five stimuli in a hierarchical order (e.g. A > B > C > D > E), and consequently, (2) they are able to infer the relationship between B and D, i.e. two colours never seen together before (e.g. to be B > D).

2. Material and methods

(a) . TI training phase

We used 18 N. pulcher, nine males and nine females (see electronic supplementary material [30] for fish maintenance). They were trained in a five-term series, consisting of the presentation of four pairs of colour stimuli [2]. We used two batches of nine fish. For the second batch (combination 2), we reversed the order of colour stimuli as compared to the first batch (combination 1) to control for colour preference (electronic supplementary material [30], figure S3A). We used a similar learning paradigm as [27] (see electronic supplementary material [30]).

(b) . TI test phase

All fish which reached the learning criterion in the last intermixed session (N = 12), were tested on whether they used transitive inference. The test phase was done on 3 consecutive days, with two sessions of three trials each day (in total, 18 trials). The first and third trials of a session consisted of a rewarded trial of a randomly chosen known pair (A + B−, B + C−, C + D− or D + E−) to avoid extinction learning. Only the second trial was a non-rewarded test trial, where we presented colour B against D. Only the non-rewarded trials were used for the analysis of TI. The fish received a 30-min break between the two sessions of a day to prevent carry-over effects of experiences between the two sessions. Fish trained with combination 1 succeeded in the task if they chose B over D, fish trained with combination 2 succeeded when they chose D over B. Due to technical issues, only one fish went through two TI trials and thus we only used the data of the first trial of this fish for analysis.

(c) . Data analysis

All analyses were performed in R (4.0.2) [31]. To analyse the probability of success in the TI test we performed two-tailed binomial tests (i) overall first choices made by fish that participated in the test phase, and (ii) for the individual choices made by every fish over all the six test trials. We compared the choices performed by the fish by fitting three generalized mixed effect models (package ‘lmerTest’, [32]) assuming binomial distribution, two models to analyse the outcomes during the training periods for the two batches and one model on outcome of the TI test. We used observer, sex, type learning session (in models on training periods), rewarded colour and side as fixed factors. Models were simplified by stepwise exclusion of factors, retaining the more parsimonious model if it had a lower AIC. Individual identity was included as a random factor. Post hoc analyses were conducted for pairwise comparison using the Bonferroni method for adjusting p values (package ‘emmeans’, [33]). To test whether the frequency of exposure to colours B and D affected the choice in the TI test, we calculated the ratio between the number of rewards received from B and D (i.e. rewards received from B/rewards received from D). We then correlated the reward ratio to the error percentage during the TI test and during the training phase, respectively, by Spearman correlation tests.

3. Results

(a) . Training phase

The number of correct choices during the training was predicted by the rewarded colour and the side (left or right) of the reward on the well plate (table 1). Fish in batch 1 made more correct choices when the reward was placed on the right side and fish from batch 2 made more correct choices when the reward was placed on the left side. Post hoc analysis showed that fish strongly preferred the colour that was always rewarded in both batches (i.e. red and purple, table 1, figure 1a). In batch 1, there was no preference for colours B and D, i.e. the ones used in the TI test (i.e. black and blue, table 1). In batch 2, fish preferred black (table 1), which during the TI test was the incorrect colour (figure 1a).

Table 1.

GLMM results for the number of correct choices in (a) batch 1 (N = 2477 observation of nine fish) and (b) batch 2 (N = 2637 observation of nine fish) during the training phase and (c) in the transitive inference (TI) test only (N = 68 observations of 12 fish). Type of session: binary variable indicating if it was a single pair or intermixed session. Rewarded colour: rewarded colour in a trial (factor with four levels). Side: binary variable indicating whether the reward was placed on right or left side of the well plate. (d,e) Pairwise comparisons between the different colours in the two batches.

factors	d.f.	LRT	p
a. number of correct choices in batch 1
rewarded colour	3	48.86	<0.0001
type of session	1	2.46	0.12
side	1	7.44	0.006
b. number of correct choices in batch 2
rewarded colour	3	46.66	<0.0001
type of session	1	3.00	0.08
side	1	11.44	0.0007
c. number of correct choices in the TI test
side	1	2.16	0.14
correct colour	1	2.27	0.13
sex	1	2.64	0.07
comparisons	estimates ± SE	Z ratio	p
d. batch 1
red–black	0.87 ± 0.13	6.86	<0.0001
red–blue	0.58 ± 0.15	3.86	0.0007
red–green	0.76 ± 0.13	5.77	<0.0001
black–blue	−0.29 ± 0.14	−2.12	0.20
black–green	−0.11 ± 0.12	−0.94	1.00
blue–green	0.18 ± 0.14	1.30	1.00
e. batch 2
purple–black	−0.54 ± 0.15	−3.66	0.0015
purple–blue	0.35 ± 0.11	3.07	0.01
purple–green	0.27 ± 0.12	2.31	0.12
black–blue	0.89 ± 0.14	6.18	<0.0001
black–green	0.81 ± 0.15	5.52	<0.0001
blue–green	−0.07 ± 0.11	−6.32	1.00

Figure 1.

Success in (a) the transitive inference test from the first trial of all fish (N = 12) and (b) every individual that went through the six TI trials (N = 11; fish K only went through two trials and therefore was excluded here).

(b) . TI test

Out of the 12 tested fish, 11 chose the correct option in the TI test during the first trial (table 2, binomial test, p = 0.006; figure 1a). Of 11 fish taking part in all test trial, five fish performed significantly above chance (6/6 correct choices, table 2), while the other six fish made one between and three mistakes (table 2, figure 1b). The reward ratio B/D (batch 1) or D/B (batch 2) did not predict the error rate in the TI test (Spearman correlation, rho = −0.28, p = 0.41) nor did it predict the error rate in the training phase (Spearman correlation, rho = −0.27, p = 0.22). There was no side or colour effects on the number of correct choices in the TI test (table 2). There was no statistical difference between sexes in TI performance, which might be attributable to a lack of statistical power (table 2).

Table 2.

Results of binomial tests for the TI test trials (N = 12 for first choice, N = 11 for all test trials).

correct choices	N of fish	probability of success	p
first choice correct	11	0.92	0.006
6 out of 6	5	1	0.03
5 out of 6	1	0.83	0.22
4 out of 6	3	0.67	0.69
3 out of 6	2	0.5	1

4. Discussion

Our results demonstrate that N. pulcher use TI when inferring relationships between artificial stimuli in a non-social context. Fish chose the correct colour in the first test trial for TI above chance, and, on the individual level, about half of the fish performed above chance, with 100% accuracy, in the six TI test trials. Our results do not support the use of value transfer as the underlying mechanism of TI in N. pulcher, but our statistical power to draw final conclusions on the mechanism of TI is limited.

While TI in a non-social context was shown in a number of mammals and birds [14,16,34–36], there has been only one such study aiming to test TI in a fish thus far [37]. However, it did not ultimately show the use of TI above chance level, as only four individuals were tested, and not all of these did the correct choice in the first test trial. Moreover, [37] used rewarded test trials, so that the reported indications of the use of TI can alternatively be explained by associative learning during tests. In our study, the TI test trials were not rewarded to exclude associative colour learning and fish performed above chance level. Moreover, our results cannot be explained (i) by simple colour preferences, as the two batches of fish were trained on the opposite order of the five-term colour series or (ii) by different experience with colour pairs, as error rates during the TI test were not influenced by the reward ratio of test colours during the training. We can therefore safely conclude our fish used TI in a non-social task.

In several taxa, highly social species possess the ability to represent transitive relationships among the hierarchy ranks of conspecifics [4,5,12,19]. It has rarely been tested if this ability can be transferred to a non-social context, which would support domain-general organization of cognition [11]. In N. pulcher, there is recent evidence that adaptive social flexibility, i.e. social competence [7], and non-social flexibility are affected similarly by the same physiological manipulation [38]. While previously disputed [39], this provided the first evidence of domain-general cognition in a fish [38]. We propose that our results provide further indication of domain-general cognition in fish spanning the social and non-social domain because TI in a social hierarchy context was shown to exist in the social cichlids A. burtoni and J. transcriptus [5,12], and TI in a non-social context was shown in a social cichlid in this study. Moreover, N. pulcher track the relative ranks of other group members allowing them to reduce their queuing time to territory inheritance when entering group [40], and thus it is likely they also employ TI for inferring ranks. However, to verify domain-general TI in fish, it needs social and non-social TI experiments in the same fish species.

While we demonstrated that fish could make transitive inferences of colour stimuli, the mechanism underlying this ability is yet to be explored. The possibility to use associative learning by value transfer to solve a TI task was first investigated and confirmed in pigeons [14,21,22]. In our study, we performed a separate experiment, in which we trained six fish and tested if they might have used value transfer as a possible mechanism to solve a TI paradigm (see electronic supplementary material [30]). There was no significant evidence that the fish used value transfer in our experiment (see electronic supplementary material [30], figure S4 and Supplementary Results). This may suggest that a more complex cognitive mechanism is at work. Grosenick et al. [5] discussed that value transfer theory alone cannot explain TI in their study species. However, our negative results for value transfer may also be due to low power, as only one of the six fish failed to show a correct first choice. Nevertheless, none of our repeated individual tests provided evidence for an above-chance use of value transfer.

In conclusion, our study adds to existing evidence suggesting that fish can perform complex cognitive tasks such as TI, but which thus far was only shown in a social [5,12,34] context. As brain structures involved in the social decision making network are conserved across vertebrates [41], it should not be surprising that fish often show cognitive abilities comparable to other vertebrates [42]. The underlying mechanisms of these abilities are largely understudied, but may often be simpler than assumed (e.g. value transfer), and may differ among different vertebrate species.

Data accessibility

The data are available from the Dryad Digital Repository: https://doi.org/10.5061/dryad.vhhmgqnx0 [43].

Supplementary material is available online [30].

Authors' contributions

O.L.L.: conceptualization, data curation, formal analysis, investigation, methodology, supervision, writing—original draft, writing—review and editing; A.R.: conceptualization, data curation, formal analysis, investigation, methodology, writing—review and editing; B.T.: conceptualization, funding acquisition, investigation, methodology, project administration, supervision, writing—original draft, writing—review and editing.

All authors gave final approval for publication and agreed to be held accountable for the work performed therein.

Conflict of interest declaration

We declare we have no competing interests.

Funding

This study was financed by the Swiss National Science Foundation (SNSF, Grant 31003A_179208 to B.T.).