The impact of learning opportunities on the development of learning and decision-making: an experiment with passerine birds

1. Introduction

The developmental environment can shape how animals behave. Early-life experiences can impact and even constrain morphology and behaviours displayed in adulthood. Such was the case with jumping spider (Marpissa muscosa) siblings that were raised with either physical enrichment (i.e. objects added), social enrichment (i.e. 15 individuals living together) or no enrichment [1]. The authors found that the developmental environments in which the individuals were reared caused personality differences, with higher exploratory behaviours observed in the enriched treatments (see also [2–4]).

The developmental environment can also affect cognitive abilities. Environments can vary in their degree of certainty, where stable environments have greater certainty of resource availability and fluctuating environments have lower certainty [5]. When environmental certainty is high, individuals do not need to pay the time and energy cost associated with acquiring new information and rely more heavily on innate behaviours or prior experience. Comparatively, fluctuating resources require individuals to physically sample, gather information about their environment, store information and adjust their behaviours accordingly by relying on additional cues to optimize their decisions [6]. Juvenile African cichlid fishes (e.g. Simochromis pleurospilus) that experienced environmental fluctuation in the form of changes in feeding rations were found to outperform individuals that experienced stable feeding rations when tested in an associative learning task during adulthood [7]. Presumably, cichlids that acquired more information about their environment during development were better able to adjust to environmental changes encountered later in the learning test than those that had not experienced environmental changes in their lifetime. When individuals find themselves in uncertain environments with high variability of resource availability, they can learn to use cues in order to make adaptive choices and cope with the changing environment. Through learning, an individual can increase its resource intake and energetic gain through experience [8]. When the environment is uncertain, animals have been known to learn to use reliable cues in order to optimize their foraging decisions [9]. Thus, models such as the components of change model (i.e. the ‘flag’ model [10]) have suggested that experiencing environmental changes causes animals to gather more information about predictive cues in their environment which could lead to more optimal behavioural patterns [11–13]. Still, it is not yet well understood how experience can cause differences in behaviour and cognition based on the informational properties (e.g. presence of cues predicting resources and their reliability) of the developmental environment.

In theory, the more information individuals have about their environment, the better they can adjust to it. However, acquiring information for accurate decisions can be time-consuming and can affect the overall speed of choice. A speed–accuracy trade-off (SAT) has been suggested where physical and cognitive constraints hinder individuals from reaching high choice accuracy at high speeds [14]. SATs have been studied in order to better understand the mechanisms of decision-making and their relationship to the current environment, and are found across taxa including humans and invertebrates [15–17]. When exposed to a colour discrimination task, zebrafishes (Danio rerio) were found to have an increased average decision time as their choice accuracy increased, thus supporting the proposed SAT in cognitive performance [18]. Still, the effects of the developmental environment and experience on the individual variation seen in SATs are not yet well understood.

We aim to address how early experience affects learning performance and speed of choice by manipulating the informational properties (i.e. presence and number of reliable cues) of the developmental environment of juvenile captive zebra finches (Taenopygia guttata). Zebra finches are altricial passerine birds, as opposed to precocial species, with a long developmental period which draws a parallel with humans [19,20]. Like humans, zebra finches have been found to exhibit sensitive periods during juvenile development that can affect behaviour later in life (song learning: [21–23]). We exposed juvenile zebra finches to one of three experimental conditions; i.e. an environment where (i) colour cues reliably predicted the presence of food (associative learning), (ii) a combination of two-colour cues reliably predicted the presence of food (conditional learning), or (iii) colour cues were non-informative (control). We then measured learning performance in two discrimination tasks and predicted that birds with more experience with information acquisition during development would have increased choice accuracy (i.e. conditional learning treatment > associative learning treatment > control). Furthermore, we hypothesized that decision-making would also be influenced by experimental manipulation of information acquisition during development; we predicted that birds will choose faster when they have had experience with more information acquisition (i.e. conditional learning treatment > associative learning treatment > control). Finally, we tested for a SAT during learning using individual variation in speed and accuracy of choice, where we predicted that speed of choice should decrease as choice accuracy increases.

2. Material and Methods

(b) Experience

Individuals from each of the two groups were randomly assorted into three experimental conditions while counterbalancing for sex. The three experimental treatments offered different information sources, i.e. an environment where (i) colour cues reliably predicted the presence of food (associative learning or 1-cue learning), (ii) a combination of two colour cues reliably predicted the presence of food (conditional learning or 2-cues learning), or (iii) colour cues were non-informative (control). The three experimental conditions were achieved by using a silo bird feeder that could be matching or mismatched, with four possible feeder combinations: top green–bottom green (matching), top white–bottom green (mismatched), top white–bottom white (matching), top green–bottom white (mismatched). For the control treatment, all four feeder combinations had equal chance of being rewarding and thus had no informative cues. Individuals in the associative learning treatment only had to pay attention to the bottom colour of the feeder (i.e. white or green) as this was the only informative cue. Only matching feeders (i.e. top green–bottom green and top white–bottom white) were rewarding for the conditional learning treatment, as the combination of both cues would be the reliable predictor of the correct feeder (figure 1).

• Download figure
• Open in new tab
• Download PowerPoint

Birds were first exposed to both matching feeders (green–green and white–white) containing ab libitum seed mix simultaneously for a habituation period of 2 days in order to reduce neophobia (fear of novelty) responses. During experiments, two feeders with seed mix were presented simultaneously as a binary choice, each located close to a perch at the same height in the cage. The first four trials alternated between rewarded sides (i.e. R, L, R, L or L, R, L, R) in order for individuals to experience both sides as rewarding at least twice and prevent side biases. For the remainder of trials, the rewarded side was chosen randomly while ensuring that both sides were rewarded equally often. We also ensured that birds in the control and associative learning treatments would have an equal chance of matched and mismatched feeders being rewarding. Because matching feeders would only be rewarding in the conditional learning treatment, we counterbalanced for colour, making sure that matching green and white feeders would be equally rewarding. Birds were exposed to 20 trials per day for a total of 500 trials, where trial is defined as a presentation of the binary feeder choice, whether or not a choice was made by the bird. Each bird was given a total of 3 min to choose (i.e. look at the contents within the feeder approached) before the trial was terminated and the feeders were removed. If a bird chose the rewarded feeder, it was given 10 s to feed before the feeders were removed and we continued trials for the next bird. If a bird chose the unrewarded feeder, both feeders were removed immediately and the bird was not allowed to feed until the next trial, i.e. after all other birds had completed this trial. Individuals had to choose a feeder for a minimum of 400 trials to be included in the sample; all birds reached at least 400 choices (range: 400–500). Feeder chosen and latency to choose were recorded. Trials took three to four months to complete.

(c) Learning

Zebra finches were subjected to the first learning task within two months after the experience phase finished; the second learning task took place within two months after the first task and three to four months after the experience phase had ended. Sample size was reduced to 22 (N = 22) for the first learning task as four birds had by then died. For both learning tasks, individuals were exposed to 50 binary choices between a rewarding and non-rewarding feeder. Once again, the first four trials alternated between sides, individuals experiencing each side as rewarding at least twice; rewarded side was then chosen randomly and counterbalanced for the remainder of trials. Birds would then be exposed to 10 trials per day for a total of 50 trials. Individuals were given a total of 5 min to choose before a trial was terminated, and trials proceeded as in the experience phase, both feeders being removed either upon the bird reaching into the unrewarded feeder or after the bird had fed for 10 s in the rewarded feeder. Two more birds died before the second learning task and sample size was further reduced to 20 (N = 20). Individuals had to make a minimum of 40 choices of a feeder to be included in the sample; all birds reached at least 40 choices for both tasks. We recorded the feeder chosen and the latency to choose.

The first learning task was a discrimination task composed of a binary choice between two feeders: a rewarding purple bowl with vertical black electrical tape and an unrewarding purple bowl with a horizontal black electrical tape. The second learning task was based on three different shades of blue (e.g. dark blue, medium blue, light blue). Similar to the first task, birds were exposed to a discrimination task composed of a binary choice between two feeders: a rewarding bowl with a brown sticker and an unrewarding bowl with no sticker (figure 2b). We ensured that all three shades were equally rewarding throughout the 50 choices.

• Download figure
• Open in new tab
• Download PowerPoint

3. Results

(b) Learning

We found a strong and significant difference in mean accuracy between the two learning tasks (GLMM: estimate = 1.02 ± 0.12, χ² = 74.81, p < 0.001). For this reason, we decided to separate our initial model into two separate models, one for each learning task (figure 3). For task 1, we found a significant interaction between trial number and experimental treatments (GLMM: estimate = 0.03 ± 0.01, χ² = 7.64, p = 0.02) (table 2). Though there was a significant difference between the slope of control treatment and associative learning over trials, there was no difference between the slope of associative and conditional learning treatments over trials (GLMM: estimate = −0.0002 ± 0.01, χ² = 0.98, p = 0.99). A steeper learning slope for the control treatment in comparison with the associative learning and conditional learning seems to result from lower choice accuracy by control birds in initial trials (figure 3). To further investigate differences between experimental treatments in their performance, we ran a model without the interaction: task 1 showed significantly higher choice accuracy in the associative treatment when compared with the control treatment (GLMM: estimate = 0.58 ± 0.29, χ² = 4.08, p = 0.048). We found no significant effect of the conditional treatment when compared with the control (GLMM: estimate = 0.58 ± 0.29, χ² = 4.08, p = 0.46) and associative treatment (GLMM: estimate = −0.37 ± 0.27, χ² = 4.08, p = 0.18). There was no significant effect of sex or group (table 2). For task 2, we found no significant trial by treatment interaction (GLMM: estimate = −0.003 ± 0.02, χ² = 0.26, p = 0.88). After eliminating the interaction, we found that trial was positive and significant; treatment, sex, task and group were not significant (table 3). Results show that, contrary to our initial predictions, experience with more information did not always lead to higher learning performance, as individuals that experienced 2-cue learning (i.e. conditional learning treatment) did not perform better in comparison with the control treatment. However, higher accuracy in task 1 in the associative learning treatment compared with control birds does support our prediction.

• Download figure
• Open in new tab
• Download PowerPoint

• Download figure
• Open in new tab
• Download PowerPoint

Table 2. Learning phase. GLMM (N = 22 individuals) used to analyse learning task 1 using binary choice (incorrect = 0, correct = 1) as the response variable. Reference category is ‘Control’ for ‘treatment’, and ‘Females’ for ‘sex’.

Collapse

	estimate ± s.e.	χ2	p-value
intercept	1.46 ± 0.42	12.32	0.04
trial	0.04 ± 0.01	21.68	<0.001
treatment	0.55 ± 0.30	3.66	0.16
sexMales	−0.13 ± 0.24	0.30	0.59
group	−0.20 ± 0.23	0.77	0.38
treatment × trialControl versus Associative	−0.03 ± 0.01	7.64	0.02
treatment × trialControl versus Conditional	−0.03 ± 0.01	7.64	0.01
treatment × trialAssociative versus Conditional	−0.0002 ± 0.01	0.98	0.99

Table 3. Learning phase. GLMM (N = 20 individuals) used to analyse learning task 2 using binary choice (incorrect = 0, correct = 1) as the response variable. Reference category is ‘Control’ for ‘treatment’, and ‘Females’ for ‘sex’.

Collapse

	estimate ± s.e.	χ2	p-value
intercept	3.09 ± 0.78	17.09	<0.001
trial	0.05 ± 0.01	42.32	<0.001
treatmentControl versus Associative	−0.14 ± 0.53	0.34	0.79
treatmentControl versus Conditional	0.15 ± 0.5	0.34	0.77
treatmentAssociative versus Conditional	0.29 ± 0.5	0.34	0.79
sexMales	−0.54 ± 0.42	1.62	0.20
group	−0.50 ± 0.41	1.43	0.23

4. Discussion

The goal of this research was to understand how the informational properties of the developmental environment could affect decision-making and learning performance. Our experimental treatments proved to be effective as individuals that experienced predictive cues (i.e. associative and conditional learning treatments) had higher choice accuracy than control birds in the experience phase, i.e. birds in these treatments were indeed learning to use the predictive cues during the initial phase of the study. Then, we found that early experience with information caused differences in learning rate and accuracy, as well as decision-making speed, when tested several weeks later.

During training, when the environment provided information (associative and conditional learning treatments), choice accuracy was higher when compared with individuals reared in an environment with no information, supporting the predicted effect of our experimental treatments. Still, there was a significant difference in choice accuracy contingent on the amount of information provided. Juvenile birds that developed with only one source of information (i.e. associative learning treatment) had higher choice accuracy than individuals with two sources of information (conditional learning treatment). This result suggests that too much information could constrain choice accuracy as the task of paying attention to two different sources could be more difficult than paying attention to just one. Indeed, when given the choice between paying attention to one primary source of information or a second conditional signal, captive blue jays (Cyanocitta cristata) preferentially chose to pay attention to one information source unless the certainty of the primary source of information decreased [26]. There seems to be a cost associated with acquiring too much information which could in turn limit choice accuracy and learning [27].

In order to test our prediction that experience with an increasing number of information sources would lead to higher learning performance (i.e. conditional treatment > associative treatment > control), we ran two different learning tasks several weeks after the experience phase. Our first learning task resulted in significant differences in learning rates between treatments as shown by a significant treatment by trial number interaction—although caution is required with this result as we could not fit random slopes for individual over trials owing to convergence issues [28,29]. Individuals that had no experience with predictive cues had a faster learning rate, which we argue could be due to having no experience with making associations between a stimulus and a food reward; individuals that lack information are at disadvantage and will try to compensate by learning quickly after some inaccurate initial trials. Since the slope difference seems to be dependent on the high intercept for our learning treatments, we looked at overall choice accuracy in a model without the interaction and found a significant effect for our associative learning treatment when compared with control birds. Our results for learning task 1 suggest that experience with simple predictive cues (i.e. learning opportunities) allows individuals to learn to form associations, which can then be transferred to optimize learning later in life. Overall, our results show that (when there is no interaction) individuals with experience with the formation of associations have a 0.64 higher estimated probability (range 0.57–0.71) of choosing correctly than an individual that has had no experience with the formation of associations. However, we cannot reject an alternative explanation based on the development of a ‘win–stay/lose–shift’ learning rule [30,31]. For instance, birds in the learning treatments could have learned that some visual elements of the environment (feeder parts) are associated with 100% payoffs and others 0%, and thus expressed more accurate choices in the subsequent learning task where a win–stay/lose–shift rule could also be used. Future experiments could use multiple learning tasks, including some that cannot be solved with the same rule as that encountered during the experience phase, to assess the generality of the effect of these developmental experiences on learning ability.

Although it is commonly predicted that greater experience with environmental information can lead to greater learning performance [32], and that providing opportunities for learning contributes to increases in general intelligence in young mice [33] and humans [34], we did not find the predicted difference in either learning slope or learning accuracy between the conditional and associative learning treatments. In fact, our results indicate that less information (i.e. associative treatment) leads to increased choice accuracy when compared with the control treatment. One possible explanation for this is the similarity of the associative treatment experience phase (1-cue learning) and the learning tasks (1-cue learning). Another explanation could be that too much information may actually constrain learning, as individuals could try to form associations that do not exist.

Our results for the second task show that although learning took place, there was no difference between treatments in learning rates or choice accuracy. The lack of significant difference between treatments here may be due to a high choice accuracy being reached much faster than in task 1, thereby leaving less variation in learning performance to be explained by predictor variables. Furthermore, our statistical analyses could have benefitted from an increased statistical power and our conclusions may be limited by our smaller sample size.

Our results show that zebra finches in the learning treatments decided on a feeder faster in comparison with birds in the control treatment. Shorter latencies in learning treatments support our predictions that birds that have more experience with information gathering should be faster decision makers, although this result was apparent only in one of the two learning tasks. A possible mechanism for the shorter latency in the learning treatments may involve the detection and perception of information. Theoretically, once individuals have experienced informative cues, they should be able to detect and perceive information quicker, which could lead to faster choices and, potentially, more optimal behaviours [35].

Our results also suggest that accumulating experience in a given feeding situation causes individuals to choose faster, as latency to express a choice decreased over successive trials in the experience phase. It is assumed that animals can have a belief or an estimate about the state of their current environment or the value of a given resource [36]. The degree of belief can fluctuate between uncertain, where an individual is unsure of the estimate and its choice, and certain, where the individual will be more confident about its estimate and its choice [37]. It is possible that the decrease in choice latency over experience trials observed in our experiment indicates a higher certainty in the birds' belief as they gradually gain knowledge on the task presented to them. Little is known about the link between the speed of choice and the certainty of belief; we suggest this would be a topic worthy of investigation. However, an alternative explanation for the decreased choice latencies with trial number in the experience phase could be related to a reduction in neophobia (fear of novelty) towards the new feeders or experimental set-up over successive trials, given that this reduction in latency over trials was not found in the subsequent learning tasks.

Learning can imply costs and constraints [38,39]. One of the implied constraints of learning is a cognitive and/or physiological trade-off between speed and accuracy where individuals can choose fast with minimal amount of information, or slowly with more information, which should result in higher accuracy [40]. We predicted that the latency to choose should increase with choice accuracy in the current trial, because birds would then take more time to choose accurately. We found no support for a SAT in our results; however, it is possible that these results are due to a lack of statistical power. This relationship should be examined with a larger sample size and by partitioning between- and within-individual variance [39]. Similarly, work on jumping spiders (Portia labiate) found no support for a SAT on prey choice, and suggested that faster decision makers may also gather information faster [41]. In our experiment, individuals that developed in one of the environments that provided information (associative learning) were found to be more accurate than controls in learning task 1, and although birds from the learning treatments were not more accurate in learning task 2 they did choose faster. Together, these results seem to indicate that these individuals could express higher accuracy or speed, but maybe not both at the same time, providing indirect evidence for a SAT.

Our experiment provides empirical evidence for theoretical models that predict that experience with information early in life should lead to distinct phenotypic variation [42]. We conclude that experience with predictive cues can lead to individual differences in decision-making and learning performance, which are important components in how animals respond to their environment. By exposing young birds to either predictive or non-predictive cues and testing them on learning months after, we provide a rare longitudinal and experimental examination of the effect of exposure to learning opportunities on the development of cognitive traits. Our work suggests that there are benefits to having experience with making associations and the optimality of learning performance. These results emphasize the fact that learning abilities are not fixed, but that learning is a developmental process that can be shaped by information gathering opportunities in the environment. Though work still needs to be done on the evolution of learning at an empirical level, we suggest that the informational properties of the developmental environment can work as a factor that could promote phenotypic variation that may have long-term adaptive consequences [43].

Ethics

Individuals were deprived of food for 0.5 h before the lights turned off for the day and the birds rested for the night. Birds were further deprived for 1 h the next morning and before beginning test trials. They had ad libitum access to water at all times. Tests were conducted in the birds' home cages and dyads were separated by a white divider to avoid handling and associated stress before conducting the assays. Work was conducted under University of Ottawa Animal Care protocol 1758.

Data accessibility

Our data are available in the Dryad Digital Repository: https://doi.org/10.5061/dryad.9w0vt4bbn [44].

Authors' contributions

I.R.-F. designed the study, collected data, ran statistical analyses and drafted the manuscript. J.M.-F. conceived and helped design the study, participated in the statistical analyses and helped draft the manuscript. Both authors gave final approval for publication.

Competing interests

We declare we have no competing interests.

Funding

This work and I.R.-F. were supported by a National Science and Engineering Research Council of Canada (NSERC) Discovery grant to J.M.-F. (435596-2013), as well as a Human Frontiers Science Program grant to J.M.-F. (RGP0006/2015).