Advertisement
Browse Subject Areas
?

Click through the PLOS taxonomy to find articles in your field.

For more information about PLOS Subject Areas, click here.

  • Loading metrics

Operant Conditioning in Honey Bees (Apis mellifera L.): The Cap Pushing Response

Operant Conditioning in Honey Bees (Apis mellifera L.): The Cap Pushing Response

  • Charles I. Abramson, 
  • Christopher W. Dinges, 
  • Harrington Wells
PLOS
x

Abstract

The honey bee has been an important model organism for studying learning and memory. More recently, the honey bee has become a valuable model to understand perception and cognition. However, the techniques used to explore psychological phenomena in honey bees have been limited to only a few primary methodologies such as the proboscis extension reflex, sting extension reflex, and free flying target discrimination-tasks. Methods to explore operant conditioning in bees and other invertebrates are not as varied as with vertebrates. This may be due to the availability of a suitable response requirement. In this manuscript we offer a new method to explore operant conditioning in honey bees: the cap pushing response (CPR). We used the CPR to test for difference in learning curves between novel auto-shaping and more traditional explicit-shaping. The CPR protocol requires bees to exhibit a novel behavior by pushing a cap to uncover a food source. Using the CPR protocol we tested the effects of both explicit-shaping and auto-shaping techniques on operant conditioning. The goodness of fit and lack of fit of these data to the Rescorla-Wagner learning-curve model, widely used in classical conditioning studies, was tested. The model fit well to both control and explicit-shaping results, but only for a limited number of trials. Learning ceased rather than continuing to asymptotically approach the physiological most accurate possible. Rate of learning differed between shaped and control bee treatments. Learning rate was about 3 times faster for shaped bees, but for all measures of proficiency control and shaped bees reached the same level. Auto-shaped bees showed one-trial learning rather than the asymptotic approach to a maximal efficiency. However, in terms of return-time, the auto-shaped bees’ learning did not carry over to the covered-well test treatments.

Introduction

Operant conditioning as originally envisioned by B. F. Skinner is characterized by goal-directed motor manipulation of the environment [1]. This manipulation was achieved by substituting an arbitrary response such as a lever press for the locomotive response associated with such commonly used apparatus as the runway, running wheel, and maze. The role of species typical behavior was minimize by using an arbitrary response in the expectation that it would help stimulate a “functional analysis” of behavior. This arbitrary response was created by a process known as shaping (or “response differentiation by successive approximations”) [2]. Shaping consists of the use of reward and non-reward to reinforce increasingly accurate responses by a subject, leading to a final, arbitrary, response the organism would not otherwise produce. The creation of this arbitrary behavior by the use of shaping has become one of the defining characteristics of operant behavior and, in our view, one of the most important behavioral principles in psychology. As one of the most widely employed treatment techniques used in applied behavior analysis, shaping is most commonly used in applied treatments for individuals with autism spectrum disorder [3], and to aid stroke patients in overcoming learned non-use of limbs [46]. Despite its popularity in applied fields and vertebrate research, shaping is rarely investigated in invertebrates, an area potentially ripe for investigation of the neuronal basis of learning and memory. The present study uses a newly developed shaping protocol to examine operant conditioning in an invertebrate, the honey bee.

Honey bees represent an important insect model in the study of learning and memory, and two techniques have dominated this research: PER and SER [7, 8]. The proboscis extension reflex (PER) pairs a scent with an unconditioned stimulus to elicit an appetitive associative learning response. In contrast, the sting extension reflex (SER) pairs a scent with an electric shock to illicit an aversive associative learning response.

The PER technique in immobilized honey bees produces a true associative learning situation [9]. Although the first published use of a proboscis extension reflex technique in honey bees was 70 years ago [10], only after refinements making experimental results highly repeatable and applicable to psychology learning models [1114] has the protocol become an attractive mean for studying appetitive classical conditioning roles in a wide variety of learning and memory scenarios [7]. PER methodology has elucidated many cognitive similarities between honey bees and vertebrates [15, 16], and is proving to be an important tool for discovering the underlying cellular and molecular processes involved in classical conditioning [1722]. PER experiments show that honey bees may accomplish the same task as vertebrates but cognitively do so in a different manner. A case in point is the PER demonstrated inability of honey bees to use the removal of an odor stimulus as a conditioning cue [23].

In contrast to the PER modality, the sting extension reflex is an aversive, classical conditioning experimental design that also utilizes a harnessed bee [24]. Like the PER technique, the SER protocol has also been an important tool in studying the molecular basis of classical conditioning. It is clear that aversive long-term memory involves protein synthesis [25], and that the ecdysone/dopamine signaling pathway is involved in aversive classical conditioning [26]. The SER methodology actually has its roots in electro-shock aversive learning of free flying honey bees [27, 28], and has also been successfully used in shuttle box experiments [2831]. The shuttle box design allows a broader range of learning models to be tested than simple SER, and in fact showed that the SER technique (harnessed bee) actually presented an aversive rather than positive ‘attack’ conditioning situation [29]. The shuttle box experimental design also opens a new leaning domain for study: operant conditioning.

Operant conditioning centers on learning from the consequences of behavioral choices [32], and is less well studied in invertebrates than is classical conditioning when compared to the vertebrate counterparts [33, 34]. Thus far, accounts of operant conditioning in invertebrates are limited to behaviors such as discrimination in Y- or T-mazes [35, 36], place preference in shuttle boxes [31, 37], and response rate change operant chambers [34, 3840]. Specific genetic mutants in Drosophila affect classical conditioning, and it is interesting that some of those decrease classical learning mutants have no effect on operant learning [41]. Thus, even on the most basic level there are fundamental differences between classical and operant learning.

Like classical conditioning, operant conditioning has been a dominant element in the analysis of learning across the animal kingdom [42], not only for the insight provided in basic animal cognition but also as a means to explore how environmental factors impacts behavioral responses. Illustrative examples of the latter include alcohol in primates [43], pesticide ingestion by insects [9, 44], and the broad field of ‘cognitive ecology’ [45]. Various forms of operant conditioning provide the determinants for advanced cognitive processes [46, 47] and capacities for conceptual learning even in organisms with neuronal systems as relatively simple as the honey bee [48]. Drosophila, Aplysia, and Lymnaea have been important invertebrate models that have provided insights into the cellular and molecular basis of operant learning beyond what can be gleaned from their vertebrate counterparts [33], and Apis mellifera is now proving as valuable due to its rich behavioral repertoire associated with foraging and social interactions [46, 49, 50].

In this paper we present a new modality, the cap pushing response (CPR), for exploring operant behavior that lends itself to studies involving behavioral shaping of honey bees. The CPR protocol requires bees to both exhibit a truly novel behavior and in doing so operate ‘manipulata’ in a process that mirrors the lever-pushing protocols used by Skinner with vertebrates such as rats and pigeons [32]. Using the CPR protocol we test for the effect of explicit- versus auto-shaping on learning curves.

Materials and Methods

Experimental Design

Experiments were conducted spring of 2015 at the Department of Entomology of the Pontificia Universidad Católica de Valparaíso, in Quillota Chile. Honey bees (Apis mellifera L.) housed in a laboratory hive at the Pontificia Universidad Católica de Valparaíso were trained to a feeding platform 18 m from the hive following the methods of Abramson (1990) [51]. The training platform contained a 59 mm diameter gray plastic disk with a 6 ml drop of 50% (w/v) sucrose solution in the center as reward. Free-flying forager honey bees making repeated trips to the foraging platform were used in the experiments. One bee was tested at a time, with N = 55 bees in total used in experiments. All other bees were captured and removed from the location.

The experiment was initiated when the test subject returned to the hive. The gray plastic disk was removed and replaced by an experimental target that consisted of a clear 88 mm diameter plastic disk. The disk had a red 50 mm diameter circle painted in the center of the disk on the underside. In the center was a feeding well 10 mm in diameter and 6 mm deep capable of holding 20 ml of 50% sucrose solution (enough for three visits from one bee). A hollow plastic cap, measuring 12 mm in diameter, 10 mm in height, and weighing 0.12g, was used to cover the feeding well in some trials of the experiment. The cap was hollow, open only on the bottom. To control for odor cues, feeding disk and cap were changed between and within experimental trials. The feeding disks and caps were washed, triple rinsed with distilled water, and air dried between uses.

Three different experiments were performed: 1) Control, 2) Explicit-Shaping, and 3) Auto-Shaping. Each bee experienced only one of the three experimental protocols (N = 20 bees Control, N = 20 bees Explicit-Shaping, N = 15 bees Auto-Shaping). The first two experiments were run initially, and the results of those experiments led to the Auto-Shaping experiment. To control for calendar variables, bees from the Control and Explicit-Shaping groups were run intermixed, but just one bee at a time. Each run of an experiment consisted of giving a bee 20 sequential, uninterrupted trials, where each trial represented a return trip from the hive. Recorded on each trial in both experiments were: 1) return time, 2) number of cover pushes per-trial, and 3) latency to push cover from time of landing on the disk.

The Auto-Shaping experiment was subsequently performed. Each run of an experiment consisted of giving a bee 20 sequential, uninterrupted trials, where each trial represented a return trip from the hive. Recorded on each trial in both experiments were: 1) return time, 2) number of cover pushes per-trial, and 3) latency to push cover from time of landing on the disk.

Control group

Each bee was given 20 trials. The first 5 trials (baseline) utilized a disk with an uncovered feeding well. Beginning on trial 6, bees encountered the feeding well fully covered by the cap on each return trip (each trial). If a Control group bee failed to push the cap to gain access to the reward within 10 minutes, the experiment was terminated for that bee since our previous work shows that bees will abandon a situation if they are not rewarded within 10 minutes [28, 40]. If the Control group bee pushed the target, the experiment continued for an additional 15 trials for a total of 20 trials.

Explicit-Shaping group

Each bee was given 20 trials. Like the Control group, the first 5 trials (baseline) utilized a disk with an uncovered feeding well. During trials 6–10 bees experienced a behavioral shaping regiment to push the cap. On trial 6, half of the feeding well was covered by the cap (5mm). On trials 7, 8, 9 and 10 the cap progressively covered 1 mm more of the feeding well. The purpose of the shaping phase was to give the bee explicit experience in pushing the cover. Trials 11–20 matched the Control group where the cap completely covered the feeding well, and to gain access to the sucrose reward the bee had to push the cap to uncover the reward.

Auto-Shaping Experiment

Each bee was given 20 trials. Like the Control and Explicit-Shaping groups, the first 5 trials (baseline) utilized a disk with an uncovered feeding well. During trials 6–10 bees experienced a situation made to elicit behavioral auto-shaping to push the cap. During these 5 trials bees were given the cap inverted, covering all but 0.5mm of the feeding well. This allowed bees to access the reward via proboscis extension without moving the cap. Trials 11–20 matched that of the Control and Explicit-Shaping groups where the cap covered the feeding well completely; to gain access to the sucrose reward the bee had to push the cap to uncover the feeding well.

Statistical Analysis of Data

Data were fitted to the Rescorla-Wagner model of learning curves. The model predicts that the rate of learning is proportional to the difference between the current ability to solve the problem (measure of ability) and the physiological limit possible. The model is semi-log in nature and can be expressed in the form ln(cn − c) = −an + ln(c0 − c) where cn is the trial accuracy, c is the physiologically possible best accuracy, co is the accuracy on the initial trial, a is the learning-rate parameter, and n is the trial number [52]. The model fit is thus a line with y = ln(cn − c) and x = n where cn is the current measure of learning and c is a constant. We used 3 separate measure of learning, and analyze each separately. Thus, cn is ‘return-time’, ‘latency-time’, or ‘pushes’ depending upon the analysis.

Each treatment of each experiment experienced by a cohort of bees was fit to the model and statistically analyzed via regression analysis using the SAS program JMP [53, 54]. For Return-Times we used c = 180 sec which was 9 sec faster than observed for any bee in any treatment for Control, Explicit-Shaping, or Auto-Shaping group subjects. For Latency-Times we used c = 1 sec which was 1.2 sec faster than observed for any bee in any treatment for Control, Explicit-Shaping, or Auto-Shaping group subjects. When dealing with the number of cap pushes, the results were integer values greater than or equal to 1 and many bees were able to move the cover in a single push. Thus, we used c = 0.9 to resolve this issue (i.e. ln(0)) and were able to analyze the entire data set.

Results

Our results, presented in detail in the following sections, demonstrate that honey bees are capable of developing the novel CPR tactic to access a concealed food source, both with and without explicit shaping. Experience is critical to rapid mastery of this strategy. The results of the Control and Explicit-Shaping experiment led to the Auto-Shaping experiment presented below.

A. Control and Explicit-Shaping Experiments

Bees in the Control group were given two treatments: ‘baseline’ where the feeding well was uncovered (trials 1 through 5), and ‘covered’ where the feeding well was covered by a moveable cap (trials 6 through 20). Bees in the Explicit-Shaping group were given 3 treatments: ‘baseline’ where the feeding well was uncovered (trials 1 through 5), ‘shaping’ where the feeding well was progressively covered by a movable cap over multiple trial (trials 6 through 10), and ‘covered’ where the feeding well was completely covered by a moveable cap (trials 11 through 20).

Return times.

The Control group consisted of 20 bees. Five of those 20 bees failed to return on trial 6, abandoning the foraging site upon experiencing the cap-covered well. The remaining 15 bees completed all 20 trials. The regression coefficient was not significant in the baseline treatment for either the 5 bees that abandoned the site (ANOVA: F1,23 = 0.0315, P = 0.8606) or the 15 bees that completed both the baseline and covered treatments, which together encompassed 20 trials (ANOVA: F1,73 = 0.0845, P = 0.7721). Further, the mean return time did not differ significantly between the 5 bees that abandoned the site and the 15 bees that complete all treatments during the baseline period (t test: T98 = 0.01498. P = 0.9881). Combining the baseline data from these two sets of bees (5 bees that abandoned the site and 15 bees that completed all treatments), the average return-time fitting the mean over time was 240.7 sec (Fig 1).

thumbnail
Fig 1. Average Return-Time δ by Trial for Control and Explicit-Shaping Groups.

Return-Time δ = (cncc∞) is the difference between the current-trial (cn) and physiologically shortest possible return time (cc∞ = 180 sec in this study). Average return-time δ with standard error bars is presented. Each square, ■, is the mean of a trial for the Control group. Each diamond, ◊, is the mean of a trial for the Explicit-Shaping Experiment group. Vertical dashed lines mark treatment boundaries. Trials 1–5 are the baseline treatment for both Control and Explicit-Shaping bees. Trials 6–10 are the shaping treatment for the Explicit Shaped group. Trials 11–20 are the covered treatment for the Explicit-Shaping bees, while trials 6–20 are the covered treatment for the Control bees. Control bees only experienced two treatments (baseline and covered) while the Explicit-Shaping group of bees experienced three treatments (baseline, shaping, and covered). Regression lines shown are the least-square fit for the Rescorla-Wagner learning model, (cn − c) = −a n + ln(c0 − c).

https://doi.org/10.1371/journal.pone.0162347.g001

Bees in the Control group that completed the 20 trials initially experienced difficulty reaching the sucrose when the well was covered, but progressively became more efficient at doing so. The regression coefficient was significant (ANOVA: F1,223 = 282.9589, P<0.0001), and learning appeared to occur at the rate predicted by the model across the covered trials, trials 6 through 20 (Fig 1). Return time diminished by half in 7.3 trials (model half-life). The Lack-of-Fit test for the model was not significant (ANOVA: F13,210 = 0.7683, P = 0.6931).

Bees in the Explicit-Shaping group all completed the 20 trials (20 bees). This represented a significant difference from the Control group (X12 = 5.7143, P = 0.0168). However, like the Control group bees, bees in the Explicit-Shaping group did not have a significant regression coefficient for the baseline trials (ANOVA: F1,198 = 0.5632, P = 0.4548). Bee average return-time fitting the mean over time was 248.1 sec (Fig 1).

Like the Control group, bees in the Explicit-Shaping group initially experience difficulty reaching the reward when the cap was present in the shaping treatment even though it did not completely cover the well, but the bees rapidly mastered the task as the well become progressively more obscure. The regression coefficient was significant (ANOVA: F1,98 = 17.3725, P<0.0001), and the Lack-of-Fit test for the model was not significant (ANOVA: F3,95 = 1.1631, P = 0.3280). The regression-coefficient of the Explicit-Shaping group over the shaping trials was greater (negative) than the Control group regression-coefficient over the covered 15 trials where learning occurred (F test: F1,321 = 22.8587, P<0.0001). Learning for Explicit-Shaping group bees appears complete by the end trial in the shaping treatment (Fig 1), and the regression coefficient was not significant in the covered treatment for these bees (ANOVA: F8,190 = 1.8836, P = 0.1715).

Latency to hit cover.

The Control group consisted of 15 returning bees, which experienced 15 consecutive trials (trials 6–20, since trials 1–5 were the baseline treatment). The regressions was significant (P<0.0001), but also was the Lack-of-Fit (P<0.0001). The areas where the data did not fit the Rescorla-Wagner model can be seen (Fig 2) to be the initial trial (trial 6) and the last five trials (trials 16–20). For trials 7 through 15 the regression coefficient was significant (ANOVA: F1,133 = 449.2533, P<0.0001), and the Lack-of-Fit test for the model was not significant (ANOVA: F7,126 = 1.2427, P = 0.2845). Latency-time diminished by half in 1.7 trials (model half-life). Learning stopped during the last five trials (trials 16–20); the regression coefficient was not significant (ANOVA: F1,73 = 1.4429, P = 0.2335). Bee average latency-time fitting the mean over time was 3.03 sec (Fig 2).

thumbnail
Fig 2. Average Latency-Time δ by Trial for Control and Explicit-Shaping Groups in the Covered Treatment.

Latency-Time δ = (cncc∞) is the difference between the current-trial (cn) and physiologically shortest possible latency time (cc∞ = 1 sec in this study). Average latency-time δ with standard error bars is presented. Each square, ■, is the mean of a trial for the Control group. Each diamond, ◊, is the mean of a trial for the Explicit-Shaping Experiment group. Vertical dashed lines mark treatment boundaries. Trials 6–10 are the shaping treatment for the Explicit Shaped group. Trials 11–20 are the covered treatment for the Explicit-Shaping bees, while trials 6–20 are the covered treatment for the Control bees. Trials 1–5 (not shown) are the baseline treatment for both Control and Explicit-Shaping bees. Since there was no cap covering the feeding well, there was no latency-time between landing and pushing the cap. Regression lines shown are the least-square fit for the Rescorla-Wagner learning model, (cn − c) = −a n + ln(c0 − c).

https://doi.org/10.1371/journal.pone.0162347.g002

The Explicit-Shaping group (N = 20) experienced the covered treatment in trials 11–20, since the baseline treatment was trials 1–5 and the shaping treatment was trials 6–10. Like the Control group, the regression was significant (P<0.0045), but also was the Lack-of-Fit to the Rescorla-Wagner model (P<0.0001). Learning had stopped after the first 3 trials (Fig 2). The regression coefficient was significant for trials 11 through 13 (ANOVA: F1,58 = 23.4244, P<0.0001), and the Lack-of-Fit test for the model was not significant (ANOVA: F1,57 = 0.0003, P = 0.9857). As expected, learning to solve the problem had occurred in the shaping treatment, with the mean latency time in the first trial (trial 11) being only 3.60±0.16 sec (mean ± se) compared to 39.18±4.25 sec in the first trial (trial 6) of the Control group. Surprisingly, a slight increase in latency-time occurred in the Explicit-Shaping group over trials 14–20 (ANOVA: F1,138 = 5.3522, P = 0.0169), with a non-significant Lack-of-Fit (ANOVA: F5,133 = 0.3666, P = 0.8707).

Shaping led to a significant difference in task performance by trial 11 when compared to the Control group. On trial 11 the mean latency time differed between the two groups (t test: T33 = 4.9668, P<0.0001).

Number of cover pushes.

The regession coefficient was significant for the Control group bees (ANOVA: F1,223 = 406.3820, P<0.0001) and the Lack-of-Fit was not significant (ANOVA: F13,210 = 1.2371, P = 0.2547) for trials 6 through 20. Thus, the Rescorla-Wagner model explained the learning curve for the Control group well over the entire covered set of trials (Fig 3). The number of pushes needed to reveal the feeding well diminished by half in 2.2 trials (model half-life).

thumbnail
Fig 3. Average Pushes δ by Trial for Control and Explicit-Shaping Groups in the Covered Treatment.

Pushes δ = (cncc∞) is the difference between the current-trial (cn) and physiologically least possible number of cap pushes (cc∞ = 0.9 used in this study—see Materials and Methods). Average cap Pushes δ with standard error bars is presented. Each square, ■, is the mean of a trial for the Control group. Each diamond, ◊, is the mean of a trial for the Explicit-Shaping Experiment group. Vertical dashed lines mark treatment boundaries. Trials 6–10 are the shaping treatment for the Explicit Shaped group. Trials 11–20 are the covered treatment for the Explicit-Shaping bees, while trials 6–20 are the covered treatment for the Control bees. Trials 1–5 (not shown) are the baseline treatment for both Control and Explicit-Shaping bees. Since there was no cap covering the feeding well, there was no cap pushes to gain access to the feeding well. Regression lines shown are the least-square fit for the Rescorla-Wagner learning model, ln(cn − c) = −a n + ln(c0 − c).

https://doi.org/10.1371/journal.pone.0162347.g003

In contrast, the regression was significant (ANOVA: F1,198 = 113.3754, P<0.0001) but the Lack-of-Fit was also highly significant (ANOVA: F8,190 = 22.9589, P<0.0001) for the Explicit-Shaping bees over the entire set of covered trials (trials 11–20). Bees became very proficient at moving the cover by trial 14. The regression was not significant when performed on the Explicit-Shaping group trials 14–20 (ANOVA: F1,138 = 0.7596, P = 0.3850). Bee average Number of Pushes fitting the mean over time was 1.0053 (Fig 3). For trials 11 to 13 where learning occurred in the Explicit-Shaping group, the regression was significant (ANOVA: F1,58 = 52.2006, P<0.0001) and the Lack-of-Fit was not significant (ANOVA: F1,57 = 1.6314, P = 0.2067).

The shaping treatment of the Explicit-Shaping group did not have a significant effect on learning when compared to the Control group which experienced the covered treatment over the same set of trials. On trial 11 the mean number of pushes did not differ between the two groups (t test: T33 = 1.2565, P = 0.2177), which was in contrast to the observation of Latency times. However, shaping did lead to an accelerated reduction in the number of pushes to move the cover off of the feeding well in the Explicit-Shaping group once the covered treatment began. The regression slopes were significantly different (ANOVA: F1,284 = 12.9603, P = 0.0004) with the Explicit-Shaping group having the greater negative slope.

B. Auto-Shaping Experiment

Bees in the Auto-Shaping experiment were given 3 treatments: ‘baseline’ where the feeding well was uncovered (trials 1 through 5), ‘auto-shaping’ where the feeding well was covered by an inverted cap that allowed bees to access the reward via proboscis extension without moving the cap (trials 6 through 10), and ‘covered’ where the feeding well was completely covered by a moveable cap (trials 11 through 20).

Return times.

Like the Explicit-Shaping group, bees in the Auto-Shaping experiment all completed the 20 trials (15 bees). Further, the regression coefficient for the baseline trials was not significant (ANOVA: F1,73 = 3.2370, P = 0.070). Bee average return-time fitting the mean over time was 273.9 sec (Fig 4).

thumbnail
Fig 4. Average Return-Time δ by Trial for the Auto-Shaping Group in Shaping and Covered Treatments.

Return-Time δ = (cncc∞) is the difference between the current-trial (cn) and physiologically shortest possible return time (cc∞ = 180 sec in this study). Average return-time δ with standard error bars is presented. Each diamond, ♦, is the mean of a trial for the Auto-Shaping Experiment group. Vertical dashed lines mark treatment boundaries. Trials 1–5 are the baseline treatment, trials 6–10 are the auto-shaping treatment, and trials 11–20 are the covered treatment for the Auto-Shaping bees. Regression lines shown are the least-square fit for the Rescorla-Wagner learning model, (cn − c) = −a n + ln(c0 − c). Arrow pointing down represents one-trial leaning event, which does not fit the Rescorla-Wagner model.

https://doi.org/10.1371/journal.pone.0162347.g004

The learning curve for the Auto-Shaping bees did not fit the Rescorla-Wagner model during the shaping treatment. Interestingly, the regression coefficient was significant (ANOVA: F1,73 = 92.7751, P<0.0001), as was the Lack-of-Fit test for the model (ANOVA: F3,73 = 6.9976, P = 0.0003). This is due to a one-trial learning event occurring midway through the shaping trials, rather than the Rescorla-Wagner asymptotic approach to the physiological best possible over the set of trials (Fig 4). Thus, return-time response of the Auto-Shaped group and Explicit-Shaped group were fundamentally different during the shaping treatment (trials 6–10). Explicit-Shaped bees became progressively more efficient over the series of trials in a Rescorla-Wagner learning curve manner, while the Auto-Shaped group reached its greatest proficiency in just one trial. Further, while learning from the explicit-shaping carried over to the covered treatment, the auto-shaping learning did not.

Unlike the Explicit-Shaping group, learning of the Auto-Shaped group in the shaping treatment did not carry over well to the covered treatment in the Auto-Shaping experiment (Fig 4). The regression coefficient for the final treatment (cover, trials 11–20) was significant (ANOVA: F1,148 = 31.0671, P = 0.0001) as was the Lack-of-Fit (ANOVA: F8,140 = 4.1883, P = 0.0002). The reason for the model lack of fit is that learning was complete by trial 13: bees had reached their best physiological performance (Fig 4). The regression for trials 11–13 was significant (ANOVA: F1,42 = 35.3336, P = 0.0001) with the Lack-of-Fit not significant (ANOVA: F1,42 = 0.0895, P = 0.7663). Further, the regression for trials 14–20 was not significant (ANOVA: F1,103 = 0.4299, P = 0.5135). The mean return time for trials 14–20 was 254.7 sec.

Latency to hit cover.

The results for Latency time have many similarities to the Return time. The learning curve for the Auto-Shaping bees did not fit the Rescorla-Wagner model. The regression coefficient was significant (ANOVA: F1,148 = 48.9478, P<0.0001), but so was the Lack-of-Fit test for the model (ANOVA: F8,140 = 35.7525, P<0.0001). Considering only trials 12–20, the regression was not significant (ANOVA: F1,133 = 3.2200, P = 0.0750). The mean latency time for trials 13–20 was 3.16 sec. Like the return times, latency time learning may best be described as a one-trial learning event (Fig 5) rather than the Rescorla-Wagner asymptotic approach to the physiological best possible over the set of trials. Notice that the latency times start at trial 11 at the same approximate value as the Control group on trial 11. However, the leaning model is completely different from trials 12–30. Like return-time, latency-time response of the Auto-Shaped group and Explicit-Shaped group were fundamentally different during the shaping treatment (trials 6–10). Explicit-Shaped bees became progressively more efficient over the series of trials in a Rescorla-Wagner learning curve manner, while the Auto-Shaped group reached its greatest proficiency in just one trial.

thumbnail
Fig 5. Average Latency-Time δ by Trial for the Auto-Shaping Group in Shaping and Covered Treatments.

Latency-Time δ = (cncc∞) is the difference between the current-trial (cn) and physiologically shortest possible latency time (cc∞ = 1 sec in this study). Average latency-time δ with standard error bars is presented. Each diamond, ♦, is the mean of a trial for the Auto-Shaping Experiment group. Vertical dashed lines mark treatment boundaries. Trials 6–10 are the auto-shaping treatment (no points because bees did not have to push the cap to access reward), and trials 11–20 are the covered treatment for the Auto-Shaping bees. Trials 1–5 (not shown) are the baseline treatment. Since there was no cap covering the feeding well, there was no latency-time between landing and pushing the cap. Regression lines shown are the least-square fit for the Rescorla-Wagner learning model, ln(cn − c) = −a n + ln(c0 − c). Arrow pointing down represents one-trial leaning event, which does not fit the Rescorla-Wagner model.

https://doi.org/10.1371/journal.pone.0162347.g005

Number of cover pushes.

The results for Cover-Pushed are like that for Latency time. The learning curve for the Auto-Shaping bees did not fit the Rescorla-Wagner model. The regression coefficient was significant (ANOVA: F1,223 = 23.1550, P<0.0001), but so was the Lack-of-Fit test for the model (ANOVA: F13,210 = 6.6573, P<0.0001). Considering only trials 12–20, the regression was not significant (ANOVA: F1,208 = 2.0701, P = 0.1517), and neither was the Lack-of-Fit (ANOVA: F12,196 = 0.9097, P = 0.5383). The mean Cover-Pushes for trials 12–20 was 1.029 (SE 0.011). Like Latency times, Cover-Pushes learning may best be described as a one-trial learning event (Fig 6) rather than the Rescorla-Wagner asymptotic approach to the physiological best possible over the set of trials.

thumbnail
Fig 6. Average Pushes δ by Trial for the Auto-Shaping Group in the Covered Treatment.

Pushes δ = (cncc∞) is the difference between the current-trial (cn) and physiologically least possible number of cap pushes (cc∞ = 0.9 used in this study—see Materials and Methods). Average cap Pushes δ with standard error bars is presented. Each diamond, ♦, is the mean of a trial for the Auto-Shaping group. Vertical dashed lines mark treatment boundaries. Trials 6–10 are the auto-shaping treatment while trials 11–20 are the covered treatment. Trials 1–5 (not shown) are the baseline treatment. Since there was no cap covering the feeding well, there was no cap pushes to gain access to the feeding well during the baseline treatment. Regression lines shown are the least-square fit for the Rescorla-Wagner learning model, ln(cn − c) = −a n + ln(c0 − c). Arrow pointing down represents one-trial leaning event, which does not fit the Rescorla-Wagner model.

https://doi.org/10.1371/journal.pone.0162347.g006

Discussion

Existing literature focuses primarily on visual or odor discrimination; here we have provided the protocol foundation to investigate motor tasks in free flying foragers without the use of prohibitively expensive or complex apparatus. This protocol allows for a more varied and flexible response from the animal without sacrificing measurement accuracy. Further, the unambiguous cap pushing response can be readily applied to additional insect species, and allows for comparative analysis through a simple foraging task with universal applications.

Honey bees developed the novel CPR tactic to access a concealed food source. Experience is shown to more rapid mastery of this strategy. Control bees did not experience the shaping treatment, and 25% were never able to solve the problem of accessing the covered nectar well. Nevertheless, a majority (75%) of Control group bees (no shaping treatment) accessed the food by happenstance. These bees were eventually able to consistently and rapidly access the sucrose reward after repeated contact with the cap, but acquired proficiency much more slowly than bees explicitly shaped to do so (Explicit-Shaping group). These results were seen in all three measures of learning: return time, latency time, and cap pushes. The Rescorla-Wagner learning-curve model [52] fit the observations well for all three measures in both shaped and control bees. However, learning stopped after a set of trials rather than continuing endlessly as an asymptotic approach to the physiological leaning limit as predicted by the model.

The results of the Control and Explicit-Shaping experiments led to the Auto-Shaping experiment. In this experiment, the honey bees performed a self-shaping task without explicit shaping by the experimenter. Specifically, the well was covered but the inverted cap allowed a bee’s proboscis access to the reward through small cracks anywhere around the cap. Thus, access to the reward technically did not require pushing the cap but produced a situation where it was likely to occur. The auto-shaping trials resulted in a one-trial learning situation for return time rather than a Rescorla-Wagner learning curve. However, the auto-shaping learning did not carryover when the covered trials began, where a Rescorla-Wagner curve [52] was observed. Latency time was also a one-trial learning situation during auto-shaping trials, but interestingly carried over to the covered trials. The same result occurred for number of Cover-Pushes.

In this study, bees were trained to access a concealed food source by pushing a plastic cap. Bees successfully acquired this novel cap pushing response following either successive approximations (Explicit-Shaping group), or self-shaping through a task of intermediate difficulty (Auto-Shaping group). Replications of this experiment performed with heavier 3D printed caps (S1 Text) resulted in identical results for both Explicit-Shaping (S1 Video) and Auto-Shaping groups (S2 Video), but bees were unable to access the food source by happenstance in Control groups (S3 Video). Pilot manipulations have revealed interesting errors and patterns (S4, S5 and S6 Videos), these and the results of this study suggest this method to be a valuable alternative to existing protocol exploring foraging and operant behaviors in the honey bee.

These results suggest the strategies to access concealed food sources exist naturally, but experience can hasten the bee’s mastery of such strategies. Not only must bees solve complex discrimination problems, they must also solve mechanically challenging tasks in their daily routine [55, 56]. Indeed, honey bees have been observed exploiting flowers with unsuitable morphologies for pollinator resource collection such as plants with anemophilous characteristics [57]. Honey bees have also been observed manipulating papilionate flowers, such as Robinia pseudoacacia. These flowers have pollen release mechanisms for which the honey bee is often too physically weak to activate [58]. However, honey bees appear to learn to favor R. pseudoacacia flowers that are easier to trip and in many cases were able to trip the pollen release mechanism while accessing the nectar or finding a suitable foothold [57, 58]. Learning to manipulate flowers to access a nectar and pollen reward may be the natural foundations for motor-task operant behavior in honey bees.

We have previously argued [32, 59] that any behavior sensitive to response-reinforcer contingencies should not be automatically assumed to be an example of operant behavior. When the term operant behavior is applied to a particular invertebrate, the invertebrate should not only be able to manipulate an object but show that they know how to use it. The vertebrate literature is full of demonstrations in which an organism can be taught to press a lever at a particular speed, force, or directions. As far as we know there are no such demonstrations in the invertebrate literature. We believe the Cap Pushing Response method to be a prime candidate for continued exploration of operant behaviors in honey bees, and provides a potentially valuable comparative method for the functional analysis of behavior.

This method may also be of use for other psychological investigations in honey bees. The utilization of strategies to gain access to a concealed food source suggests honey bees are capable of utilizing knowledge of representation in regard to working memory tasks [60] which provides supporting evidence for ‘aboutness’ or ‘intentionality,’ a form of mental representation [61]. Representation in arthropods is an important point to consider in the discussion of consciousness [62]. However, when considering consciousness from different zoological levels, interpretations of behavior must consider the natural history of the animal [6365].

Conclusions

While honey bees are established as useful models to study intermediate levels of cognitive complexity and associated neural substrates [46, 66], our results provide further reason to investigate cognitive skills likened to advanced vertebrates. Relatively complicated foraging tasks where not all foragers solve the problem, or do so in different ways, are useful for examining the role of neurotransmitters in invertebrate decision processes [49]. What seemed like simple chance event mediated individual decision process development now appears to be influenced by slight differences in insect neurotransmitters [49]. Indeed, these techniques are advantageously suited for apicultural and ecological relevant studies such as investigating the effects of pesticides and pollutants on the foraging capabilities of honey bees [67]. Further, the learning-curve model fit demonstrated in this study provides a powerful tool for examining subtle differences in both the rate and degree of learning. Here we saw a pronounced difference in learning curves between auto- and explicit-shaped bees. Due to the ease of access to procedural materials, the cap pushing response provides a powerful tool for basic and applied research related to insect operant behavior and cognition.

Supporting Information

S1 Text. Methods for 3D Printed Materials.

This document contains modified methodology for experiment procedure using 3D printed caps and feeding wells.

https://doi.org/10.1371/journal.pone.0162347.s001

(DOCX)

S1 Video. Explicit-Shaping.

This bee is recruited to the food well and allowed to return five times before the cap is added (only the fifth visit shown in video). She is then given five trials of explicit-shaping through successive approximations. Following shaping, the well is completely covered and the bee pushes the cap to access the well. She becomes increasingly more proficient each visit. This video is cut, time-lapsed for length, and muted for file size. This file is for example purposes and should not be used for analysis. Different and complete videos are available upon request.

https://doi.org/10.1371/journal.pone.0162347.s002

(MP4)

S2 Video. Auto-Shaping.

This bee is recruited to the food well and allowed to return five times before the cap is added (not shown in video). She is then given five trials of auto-shaping with a task of intermediate difficulty. Following auto-shaping, the well is completely covered and the bee pushes the cap to access the well. She becomes increasingly more proficient each visit. This video is cut, time-lapsed for length, and muted for file size. This file is for example purposes and should not be used for analysis. Different and complete videos are available upon request.

https://doi.org/10.1371/journal.pone.0162347.s003

(MP4)

S3 Video. Control Bees Example.

Time-lapsed video of bees recruited to the food well but do not have experience pushing the cap.

https://doi.org/10.1371/journal.pone.0162347.s004

(MP4)

S4 Video. Error Example 1.

Following five trials of successful cap pushing, a discrimination task was performed. The previously trained cap was moved to the side and a novel cap was used to cover the sucrose feeding well. The trained cap was counterbalanced between cross patterned and solid caps, and the novel cap presented for the discrimination task was different in shape. Upon first return, 11 of 12 bees pushed the cap previously trained, ignoring the cap that actually covered the sucrose well. If the bee pushed the cap off the plate, it was quickly returned. Bees would continue to push the incorrect cap when available (see S5 Video: Error Example 2).

https://doi.org/10.1371/journal.pone.0162347.s005

(MP4)

S5 Video. Error Example 2.

Bees trained with one shape of cap (in this case cross patterned) repeatedly make the same error and push the previously trained cap. This is generalized to similar shaped caps as can be seen in the video.

https://doi.org/10.1371/journal.pone.0162347.s006

(MP4)

S6 Video. Error Example 3.

This bee missed the well when pushing the cap and continued to push the cap for some time before reorienting to the sucrose well.

https://doi.org/10.1371/journal.pone.0162347.s007

(MP4)

Acknowledgments

This research was supported in part by grants from the Chilean grant agency Conicyt—PAI, the Pontificia Universidad Católica de Valparaíso Chile, and National Science Foundation grants DBI 1560389, OISE 1545803, and GRF 1144467. We would like to thank Ms. Solange Elizabeth Huaiquil for her assistance in Chile.

Author Contributions

  1. Conceptualization: CIA CWD HW.
  2. Data curation: CIA HW.
  3. Formal analysis: HW.
  4. Funding acquisition: CIA CWD HW.
  5. Investigation: CIA CWD.
  6. Methodology: CIA CWD HW.
  7. Project administration: CIA.
  8. Resources: CIA HW.
  9. Supervision: CIA.
  10. Validation: CIA CWD.
  11. Visualization: CIA CWD HW.
  12. Writing – original draft: CIA CWD HW.
  13. Writing – review & editing: CIA CWD HW.

References

  1. 1. Lee VL. Beyond behaviorism. Hillsdale: Lawrence Erlbaum Associates; 2011.
  2. 2. Skinner BF. Science and human behavior. Oxford: Macmillan; 1953.
  3. 3. National Research Council. Educating children with autism. Committee on educational interventions for children with autism. In: Lord C, McGee JP, editors. Division of Behavioral and Social Sciences and Education. Washington DC: National Academy Press; 2001.
  4. 4. Taub E, Crago JE, Burgio LD, Groomes TE, Cook EW, DeLuca SC, et al. An operant approach to rehabilitation medicine: Overcoming learned nonuse by shaping. J Exp Anal Behav. 1994;61: 281–293. pmid:8169577
  5. 5. Ballester BR, Maier M, San Segundo R, Castañeda V, Duff A, Verschure PFMJ. Reinforcement-Induced Movement Therapy: A novel approach for overcoming learned non-use in chronic stroke patients Reinforcement-Induced Movement Therapy: A novel approach for overcoming learned non-use in chronic stroke patients. In: International Conference on Virtual Rehabilitation (ICVR) Proceedings. Valencia; 2015.
  6. 6. Ballester BR, Maier M, SS Mozo RM, Castañeda V, Duff A, MJ Verschure PF. Counteracting learned non-use in chronic stroke patients with reinforcement-induced movement therapy. J Neuroeng Rehabil. 2016;13(74).
  7. 7. Giurfa M, Sandoz JC. Invertebrate learning and memory: Fifty years of olfactory conditioning of the proboscis extension response in honeybees. Learn Mem. 2012;19: 54–66. pmid:22251890
  8. 8. Tedjakumala SR, Giurfa M. Rules and mechanisms of punishment learning in honey bees: the aversive conditioning of the sting extension response. J Exp Biol. 2013;216: 2985–2997. pmid:23885086
  9. 9. Abramson CI, Sokolowski MBC, Wells H. Issues in the study of proboscis conditioning. In: Columbus S, editor. Social Insects: Structure, Function, and Behavir. Hauppaug, NY: Nova Science Publishers; 2011. pp. 25–49.
  10. 10. Frings H. The loci of olfactory end-organs in the honey-bee, Apis mellifera Linn. J Exp Zool. 1944;88: 65–93.
  11. 11. Takeda K. Classical conditioned response in the honeybee. J Insect Physiol. 1961;6: 168–179.
  12. 12. Vareschi E. Duftunterscheidung bei den honigbiene: Einzellzel-ableitungen und verhaltensreaktionen [Odor discrimination by the honeybee: single cell recording and behavior reaction]. Z Vgl Physiol. 1971;75: 143–173.
  13. 13. Bitterman ME, Menzel R, Fietz A, Schäfer S. Classical conditioning of proboscis extension in honeybees (Apis mellifera). J Comp Psychol. 1983;97: 107–119. pmid:6872507
  14. 14. Abramson CI, Boyd BJ. An automated apparatus for conditioning proboscis extension in honey bees (Apis mellifera L.). J Entomol Sci. 2001;36: 78–92.
  15. 15. Deisig N, Lachnit H, Giurfa M, Hellstern F. Configural olfactory learning in honeybees: Negative and positive patterning discrimination. Learn Mem. 2001;8: 70–78. pmid:11274252
  16. 16. Scheiner R, Abramson CI, Brodschneider R, Crailsheim K, Farina WM, Fuchs S, et al. Standard methods for behavioral studies of Apis mellifera. In: Dietemann V, Ellis JD, Neumann P, editors. The COLOSS BEEBOOK: standard methodologies for Apis mellifera research. J Api Res. 2013;52(4).
  17. 17. Joerges J, Küttner A, Galizia CG, Menzel R. Representation of odours and odour mixtures visualized in the honeybee brain. Nature. 1997;387: 285–288.
  18. 18. Szyszka P, Ditzen M, Galkin A, Galizia CG, Menzel R. Sparsening and temporal sharpening of olfactory representations in the honeybee mushroom bodies. J Neurophysiol. 2005;94: 3303–3313. pmid:16014792
  19. 19. Schwarzel M, Muller U. Dynamic memory networks: Dissecting molecular mechanisms underlying associative memory in the temporal domain. Cell Mol Life Sci. 2006;63: 989–998. pmid:16596333
  20. 20. Okada R, Rybak J, Manz G, Menzel R. Learning-related plasticity in PE1 and other mushroom body-extrinsic neurons in the honeybee brain. J Neurosci. 2007;27: 11736–11747. pmid:17959815
  21. 21. Rath L, Galizia CG, Szyszka P. Multiple memory traces after associative learning in the honey bee antennal lobe. Eur J Neurosci. 2011;34: 352–360. pmid:21692886
  22. 22. Strube-Bloss MF, Nawrot MP, Menzel R. Mushroom body output neurons encode odor-reward associations. J Neurosci. 2011;31: 3129–3140. pmid:21414933
  23. 23. Abramson CI, Nolf SL, Mixson TA, Wells H. Can honey bees learn the removal of a stimulus as a conditioning cue? Ethology. 2010;116: 843–854.
  24. 24. Vergoz V, Roussel E, Sandoz JC, Giurfa M. Aversive learning in honeybees revealed by the olfactory conditioning of the sting extension reflex. PLoS ONE. 2007;2: e288. pmid:17372627
  25. 25. Giurfa M, Fabre E, Flaven-Pouchon J, Groll H, Oberwallner B, Vergoz V, et al. Olfactory conditioning of the sting extension reflex in honeybees: memory dependence on trial number, interstimulus interval, intertrial interval, and protein synthesis. Learn Mem. 2009;16: 761–765. pmid:19933880
  26. 26. Geddes L, McQuillan HJ, Aiken A, Vergoz V, Mercer A. Steroid hormone (20-hydroxyecdysone) modulates the acquisition of aversive olfactory memories in pollen forager honeybees. Learn Mem. 2013;20(8): 399–409. pmid:23858051
  27. 27. Núñez JA, Denti A. Respuesta de abejas recolectoras a un estímulo nociceptivo. Acta Physiol Lat Am. 1970;20: 140–146.
  28. 28. Abramson CI. Aversive conditioning in honeybees (Apis mellifera). J Comp Psychol. 1986;100: 108–116.
  29. 29. Carcaud J, Roussel E, Giurfa M, Sandoz JC. Odour aversion after olfactory conditioning of the sting extension reflex in honeybees. J Exp Biol. 2009;212: 620–626. pmid:19218512
  30. 30. Agarwal M, Giannoni Guzman M, Morales-Matos C, Del Valle Díaz RA, Abramson CI, Giray T. Dopamine and octopamine influence avoidance learning of honey bees in a place preference assay. PLoS ONE. 2011;6(9): e25371. pmid:21980435
  31. 31. Dinges CW, Avalos A, Abramson CI, Craig DPA, Austin ZM, Varnon CA, et al. Aversive conditioning in honey bees (Apis mellifera anatolica): A comparison of drones and workers. J Exp Biol. 2013;216: 4124–4134. pmid:24133154
  32. 32. Abramson CI. A primer of invertebrate learning: The behavioral perspective. Washington DC: American Psychological Association; 1994.
  33. 33. Brembs B. Operant conditioning in invertebrates. Curr Opin Neurobiol. 2003;13: 710–717. pmid:14662373
  34. 34. Sokolowski MBC, Desmi G, Abramson CI. A paradigm for operant conditioning in blow flies (Phormia terrae novae Robineau-Desvoidy, 1830). J Exp Anal Behav. 2010;93: 81–89. pmid:20676269
  35. 35. Giurfa M, Zhang S, Jenett A, Menzel R, Srinivasan MV. The concepts of ʻsamenessʼ and ʻdifferenceʼ in an insect. Nature. 2001;410: 930–933. pmid:11309617
  36. 36. Zhang S, Si A, Pahl M. Visually guided decision making in foraging honeybees. Front Neurosci. 2012;6: 88. pmid:22719721
  37. 37. Wustmann G, Rein K, Wolf R, Heisenberg M. A new paradigm for operant conditioning of Drosophila melanogaster. J Comp Physiol A. 1996;179: 429–436. pmid:8785009
  38. 38. Kisch J, Erber J. Operant conditioning of antennal movements in the honey bee. Behav Brain Res. 1999;99: 93–102. pmid:10512576
  39. 39. Sokolowski MBC, Abramson CI. From foraging to operant conditioning: A new computer-controlled Skinner box to study free-flying nectar gathering behavior in bees. J Neurosci Methods. 2010;188: 235–242. pmid:20171985
  40. 40. Craig DPA, Varnon CA, Sokolowski MBC, Wells H, Abramson CI. An assessment of fixed interval timing in free-flying honey bees (Apis mellifera ligustica): An analysis of individual performance. PLoS ONE. 2014;9(7): e101262. pmid:24983960
  41. 41. Putz G, Bertolocci F, Raabe T, Zar T, Heienberg M. The S6KII (rsk) gene of Drosophila melanogaster differentially effects an operant and a classical learning task. J Neurosci. 2004;24: 9745–9751. pmid:15525759
  42. 42. Jenkins HM. Animal learning and behavior theory. In: Hearst E, editor. The first century of experimental psychology. Hillsdale: Earlbaum; 1979.
  43. 43. Field M, Schoenmakers T, Wiers RW. Cognitive processes in alcohol binges: a review and research agenda. Curr Drug Abuse Rev. 2008;1: 263e279.
  44. 44. Abramson CI, Squire J, Sheridan A, Mulder PG Jr. The effect of insecticides considered harmless to honey bees (Apis mellifera L.): proboscis conditioning studies using the insect growth regulators Confirm® (Tebufenozide) and Dimilin® (Diflubenzuron). Environ Entomol. 2004;33: 378e388.
  45. 45. Dukas R, Ratcliffe JM. Cognitive Ecology II. Chicago: University of Chicago Press; 2009.
  46. 46. Menzel R, Giurfa M. Cognitive architecture of a mini-brain: the honeybee. Trends Cogn Sci. 2001;5: 62–71. pmid:11166636
  47. 47. Menzel R. The honeybee as a model for understanding the basis of cognition. Nat Rev Neurosci. 2012;13: 758–768. pmid:23080415
  48. 48. Avarguès-Weber A, Giurfa M. Conceptual learning by miniature brains. Proc R Soc B. 2013;280: 20131907. pmid:24107530
  49. 49. Giray T, Abramson CI, Chicas-Mosier A, Brewster T, Hayes C, Rivera-Vega K, et al. Effect of octopamine manipulation on honeybee decision making: Reward and cost differences associated with foraging. Anim Behav. 2015;100: 144–150.
  50. 50. Seeley TD. The wisdom of the hive: The social physiology of honey bee colonies. Cambridge: Harvard University Press; 1995.
  51. 51. Abramson CI. Invertebrate Learning: A Laboratory Manual and Source Book. Washington DC: American Psychological Association; 1990.
  52. 52. Stepanov II, Abramson CI. The application of the first order system transfer function for fitting the 3-arm radial maze learning curve. J Math Psychol. 2008;52: 311–321.
  53. 53. Sokal RR, Rohlf FJ. Biometry, 3rd. New York: W. H. Freeman; 1995.
  54. 54. Sall F, Lehman A. JMP in SPSS. Belmont: Ducksberry Press; 1996.
  55. 55. Raine NE, Chittka L. Pollen foraging: learning a complex motor skill by bumblebees (Bombus terrestris). Naturwissenschaften. 2007;94: 459–464. pmid:17149583
  56. 56. Cakmak I, Sanderson C, Blocker TD, Pham LL, Checotah S, Norman AA, et al. Different solutions by bees to a foraging problem. Anim Behav. 2009;77: 1273–1280.
  57. 57. Giovanetti M, Aronne G. Honey bee interest in flowers with anemophilous characteristics: first notes on handling time and routine on Fraxinus ornus and Castanea sativa. Bull Insectology. 2011;64: 77–82.
  58. 58. Córdoba SA, Cocucci AA. Flower power: its association with bee power and floral functional morphology in papilionate legumes. Ann Bot. 2011;108: 919–931. pmid:21821623
  59. 59. Abramson CI. Where have I heard it all before: Some neglected issues of invertebrate learning. In: Greenberg G, Tobach E, editors. Comparative Psychology of Invertebrates: The Field and Laboratory Study of Insect Behavior. New York: Garland Publishing; 1997. pp. 55–78.
  60. 60. Baddeley A. Working memory: theories, models, and controversies. Annu Rev Psychol. 2012;63: 1–29. pmid:21961947
  61. 61. Dennett DC, Haugeland JC. Intentionality. In: Gregory RL, editor. The Oxford Companion to the Mind. Oxford: Oxford University Press; 1987. pp. 383–386.
  62. 62. Barron AB, Klein C. What insects can tell us about the origins of consciousness. Proc Natl Acad Sci USA. 2016;113(18): 4900–4908. pmid:27091981
  63. 63. Nagel T. What is it like to be a bat? Phil Review. 1974;83(4): 435–450.
  64. 64. Etienne A. The meaning of object permanence at different zoological levels. Hum Dev. 1984;27: 309–320.
  65. 65. Gómez J. Species comparative studies and cognitive development. Trends Cogn Sci. 2005;9: 119–125.
  66. 66. Menzel R, Müller U. Learning and memory in honeybees: from behavior to neural substrates. Annu Rev Neurosci. 1996;19: 379–404. pmid:8833448
  67. 67. Abramson CI, Giray T, Mixson TA, Nolf SL, Wells H, Kence A, et al. Proboscis conditioning experiments with honey bees (Apis mellifera caucasica) with butyric acid and DEET mixture as conditioned and unconditioned stimuli. Insect Sci. 2010;10: 1–17.