Use of temporal and colour cueing in a symbolic delayed matching task by honey bees

The Western honey bee (Apis mellifera) is an established model species for investigating cognitive processes in a miniature insect brain (Srinivasan, 2010). In particular, honey bee learning abilities have been extensively studied using delayed matching-to-sample (DMTS) tasks (Couvillon et al., 2003; Giurfa et al., 2001; Gross et al., 2009). This training procedure is widely used in various model species such as primates (Davenport et al., 1975; Rodriguez et al., 2011), pigeons (Blough, 1959; Dayer et al., 2000) and rats (Hampson et al., 1999; Wallace et al., 1980) to understand how animals make decisions based on information stored in their working memory (Lind et al., 2015). As defined in Giurfa et al. (2001), an animal firstly encounters a stimulus and subsequently after a delay must choose the identically matching or non-matching stimulus (Cooke et al., 2007; Howard et al., 2019a; Thompson and Plowright, 2016). Bees are an especially important animal model as they are currently the only insects that can be trained to perform cognitively difficult tasks such as DMTS (Lind et al., 2015). As such, studying how bees perform in these tasks provides important comparative insights into how miniature brains process different types of information stored in their memory (Srinivasan, 2010). For example, Giurfa et al., (2001) showed that bees could learn to match or non-match visual stimuli such as colours or achromatic spatial gratings, and could also transfer their learning to novel stimuli in the same or a different modality.

Bees are also capable of learning arbitrary relationships between physically dissimilar stimuli in an extension of DMTS tasks (known as symbolic delayed matching to sample). Specifically, bees can learn arbitrary relationships between colours, shapes, patterns and numbers (Cooke et al., 2007; Howard et al., 2019a; Moreno et al., 2012; Zhang et al., 1999). They can also learn arbitrary relationships between stimuli in different modalities, such as visual and tactile (Ravi et al., 2016), or visual and olfactory stimuli (Srinivasan et al., 1998). Relational learning between stimuli of different types and modalities is ecologically meaningful for bees as they can flexibly learn a variety of food-related cues (Moreno et al., 2012), which is very relevant to the natural context of complex foraging environments (Kantsa et al., 2017). However, foraging bees also potentially have access to temporal information during foraging activities (Chittka et al., 1997), but the extent to which honey bees learn relationships involving temporal cues remains unclear.

Animals in general are assumed to use temporal information in a variety of contexts related to their survival (Bateson, 2003; Vasconcelos et al., 2017). Evidence for time sense in the range of seconds to minutes largely comes from interval timing studies (Grondin, 2008; Vasconcelos et al., 2017), where the majority of vertebrates that have been tested show response patterns indicative of possessing a time sense (Craig and Abramson, 2015). In the context of insects, there is only behavioural evidence for interval timing in bumble bees (Bombus impatiens; Boisvert and Sherry, 2006; Boisvert et al., 2007) and Trichogramma wasps (Parent et al., 2016, 2017; Schmidt and Smith, 1987). Although previous studies have not found any interval timing capacity in honey bees (Craig and Abramson, 2015; Craig et al., 2014; Grossmann, 1973), they are still thought to use temporal cues during decision making. For example, it has been suggested that bees are aware of temporal information when processing nectar flow rates (Wainselboim et al., 2002) or when determining distance measurements (Chittka and Tautz, 2003; Esch et al., 2001), in order to maintain an optimal rate of foraging (Skorupski et al., 2006). Furthermore, honey bees are known to communicate the location and quality of a food source using a waggle dance that is temporally structured (Von Frisch, 1967), and the duration of the waggle run is correlated with the perceived distance of a food source (Seeley et al., 2000; Srinivasan et al., 2000). Therefore, honey bees are often assumed to potentially have a time sense at an interval timing resolution, but this time sense has only been tested in a fixed-interval schedule paradigm, where animals make responses based on a learnt time schedule (Grondin, 2008; Vasconcelos et al., 2017).

In this study, we investigated whether free-flying honey bees could use time in a similar way to other cueing stimuli such as colour, in a symbolic DMTS task. If bees can learn arbitrary relationships involving time, the evolution of this capacity may suggest that bees can use temporal information in a similar way during natural foraging activities. We trained individual bees to use illumination colour (colour condition) or duration (temporal condition) as context-setting cues to complete a shape discrimination task. In the colour condition, bees first experienced either a blue or yellow light for 2 s (Fig. 1), and then were trained to select the corresponding achromatic shape stimulus in the subsequent chamber (Fig. 2). In the temporal condition, bees experienced two durations of light (either 1 or 3 s) and again were trained to select the corresponding shape (temporal condition; Fig. 2). Learning performance in the colour condition should provide a baseline for symbolic DMTS for the current experimental context, while performance in the temporal context should reveal whether time can be used as a cueing stimulus for shapes. We used appetite–aversive differential conditioning as this type of conditioning is known to significantly improve the rate of learning in difficult tasks (Avarguès-Weber et al., 2010; Howard et al., 2019b). Individual bees were randomly allocated to one experimental condition for 60 choices, and after completion the bee immediately experienced the other condition for another 60 choices, thus making a total of 120 choices.

Free-flying honey bees, Apis mellifera Linnaeus (n=16), were recruited to the experimental apparatus from a University of Melbourne (Australia) research hive located 3 m away. This apparatus consisted of a presentation tunnel attached to a decision chamber (Fig. 1). Individual bees were marked on the abdomen for identification purposes and were individually trained to land at the entrance of the apparatus, move through the presentation tunnel to enter the decision chamber (20×27×10 cm), and land on plastic poles placed at the back wall of the chamber, where a sucrose reward would be placed (Fig. 1). The presentation tunnel (12.5 cm long, 3 cm diameter) was a transparent acrylic tube within a larger transparent plastic tube, and this larger tube was wrapped in WS2812B LED lights (red LED: 620–630 nm, green LED: 515–525 nm, blue LED: 465–475 nm; Adafruit, New York City, NY, USA). The blue light used in this study was produced by the blue LED, yellow light was produced by mixing of the green and red LEDs, and white was produced by mixing of all three LED types (for detailed specifications, see WS2812B Intelligent Control LED Datasheet, http://www.world-semi.com/Certifications/WS2812B.html). The gap between the two tubes allowed heat from the LED lights to be easily dissipated and we tested the apparatus using a thermocouple to ensure that temperature did not change during an experiment. These two tubes were held in place by a plywood base plate. A plywood box (17.5×13.5×17.5 cm) was used to cover the presentation tunnel in order to shield the apparatus from stray, ambient light and to prevent other bees from becoming attracted to the LED lights. The entrance of the presentation tunnel was gated by an opaque grey plastic slip to control the movements of bees. The exit of the tunnel was gated by a transparent plastic slip coated with tracing paper, as pilot studies indicated that bees had difficultly learning the illumination when the tunnel was completely sealed from external light. Small holes were drilled into the ceiling of the box to allow for heat to escape.

The LED lights were powered by a 5 V DC power source, and connected to an Arduino Uno microprocessor (Arduino AG, Ivrea, Italy) which was used to control the lights. This microprocessor was placed a metre away from the apparatus to ensure there was no electromagnetic interference from the device (Clarke et al., 2013; Greggers et al., 2013). An infrared remote and receiver communicated with the microprocessor to turn on the LED lights during an experiment and to select the required colour. The intensity of lights was not a factor in this study as bees do not process brightness during colour processing tasks (Daumer, 1956; Ng et al., 2018; Reser et al., 2012; Van Der Kooi et al., 2019). Bees viewed the lights at a visual angle greater than 15 deg (Giurfa et al., 1996), therefore driving colour processing mechanisms where brightness is not a confound. The decision chamber was constructed from medium-density fibreboard coated with a water-based varnish, and a fibreglass mesh cover. During the experimental phase, bees were trained towards 7×7 cm laminated cards of different shapes (Fig. 2). There were two sets of shape stimuli and half of the bees were initially trained towards set 1 (circle and square), while the other half were trained to set 2 (star and square pattern). After bees completed their first experimental condition, they were then trained towards the set that they had not yet experienced.

The experimental phase commenced once an individual bee had learnt to consistently enter the decision chamber. Bees were initially allocated into either the colour condition or the temporal condition (Fig. 2). Bees in the colour condition would be trapped upon entering the presentation tunnel, which was subsequently illuminated with either a blue or yellow light for 2 s. Bees were then allowed to enter the decision chamber to make a choice towards one of two shapes presented in the decision chamber, and to land on the corresponding pole (Fig. 1). The colour of the illumination acted as a context-setting cue (Mota et al., 2011), allowing the bee to predict the correct stimulus to choose for a given trial. For example, if a bee was trained to associate blue with the square stimulus, then it would know to land on the pole corresponding to the square card if it experienced blue illumination in the tunnel. Alternatively, if the bee then experienced yellow illumination, the circle card would now be the correct target and the square incorrect. Bees found a drop of 50% sucrose solution as a reward on the correct pole, and a drop of 60 mmol l⁻¹ quinine solution as a punishment on the incorrect pole.

Bees in the temporal condition completed a similar task but instead of changing colour context, they experienced either a short (1 s) duration of white light, or a long (3 s) duration of white light. The LEDs used to produce this white light for a human observer do not have a UV component, and therefore acted as a salient and easy to learn colour cue for bees (Van Der Kooi et al., 2019). In this condition, the duration of the illumination was the context-setting cue for the correct shape. We chose the durations of 1 and 3 s based on limitations of the bee's working memory. Zhang et al. (2005) showed that bees had a reliable working memory capacity up to 5 s, but there was decreasing accuracy following an exponential decay function beyond 3 s. Furthermore, an 8 s delay resulted in chance-level decision making. Therefore, an interval of 1–3 s would probably be optimally retained in the bee's working memory, whilst enabling a 300% stimulus manipulation that should be processed if the bee could perceive time differences within the interval time processing range.

Bees first made a total of 60 choices in their allocated condition, after which the bee experienced the other condition for another 60 choices with a different set of cards (Fig. 2). An equal number of bees were trained towards each of the combinations of illumination type and shapes. The decision chamber, poles and stimulus cards were regularly cleaned with 20% ethanol solution whenever a bee made contact, to exclude any olfactory cues that might be left by the bees.

In an initial analysis, we used a generalised linear mixed model (GLMM) with a logit link function to investigate whether trial number (continuous predictor variable) had a significant effect on the mean proportion of correct choices (dependent variable) in the two experimental conditions. This model also included an interaction term between experimental condition and trial number to test for potential differences in the effect of trial on each experimental condition. The model also included bee ID as a random effect to control for variation within bees. We found a significant interaction between trial number and experimental condition (z=2.30, P=0.02, s.e.m.=0.005; see Table S3) for this model, suggesting an effect of trial condition. We thus followed up this initial analysis by constructing two smaller models, one for each experimental condition. Each one of these models had trial number as a fixed predictor and bee ID number as a random term (Fig. 3). We report the findings of the two reduced models (colour and temporal) in the paper.

We individually trained 16 free-flying honey bees, and each bee experienced both colour and temporal cueing. In the colour condition, we found a significant effect of trial on the mean proportion of correct choices [z=2.6, P<0.01, 95% confidence intervals (CIs) 0.54, 0.63], indicating that choice accuracy increased over time (Fig. 3). These findings show that bees could use blue or yellow illumination colour as a context-setting cue to subsequently complete the shape discrimination task. Thus, this first result demonstrates that bees can learn a symbolic DMTS task under the current experimental conditions. This is consistent with previous work showing that bees can learn relationships between colour and shape information to complete symbolic DMTS tasks (Moreno et al., 2012; Zhang et al., 1999). Bees correctly chose the target 66% (s.e.m.=2.6) of the time in the final 20 choices of the colour condition (see Table S1 for original choice frequencies), indicating that the symbolic DMTS task remained difficult to complete, and is consistent with previous reports for honey bee performance in DMTS experiments (Zhang et al., 2005).

In contrast to the learning performance in the colour condition, honey bees were unable to use the duration of a light stimulus to inform their decisions in the shape discrimination task (Fig. 3; see Table S2 for original choice frequencies). There was no significant effect of trial number on the mean proportion of correct choices (z=−0.64, P=0.52, 95% CIs 0.46, 0.55) in the temporal condition, indicating that there was no increase in accuracy with increased training. This shows that bees did not differentiate a 300% temporal difference between the 1 and 3 s exposures to white light. Our findings therefore suggest that while bees have the cognitive capacity to learn relationships between stimuli such as colours, shapes, patterns and even numbers, they do not appear to use temporal information as a cueing stimulus for shapes. This may imply that bees do not use temporal cues in relation to other signals in natural contexts, and therefore they did not evolve a capacity to associate such signals. This finding was unsuspected as the use of temporal information is classically thought to be ecologically relevant for foraging pollinators (Boisvert et al., 2007; Chittka and Spaethe, 2007). However, it is possible that bees are using proxies for time but cannot measure temporal information directly. For example, Srinivasan et al. (2000) found that honey bees travelling to a feeder within a narrow tunnel performed round dances (signalling a short distance) when there were negligible image motion cues available, but performed waggle dances (signalling a larger distance) when the tunnel was lined with visual texture. Therefore, bees do not measure absolute distance but instead measure optic flow to determine distance travelled to a food source (Esch et al., 2001). This also shows that honey bees do not use time sense in a task where temporal cues are informative, such as measuring the distance of a food source. Evidently, bees can use alternative cues such as optic flow that provide temporal-like information without requiring the capacity to measure absolute time intervals. This is consistent with evidence showing that honey bees have difficulty learning temporal components of signals in other experimental contexts (Craig et al., 2014; Grossmann, 1973; Srinivasan and Lehrer, 1984).

Although honey bees may potentially process temporal differences of a larger magnitude, our current findings suggest that their time sense is limited or only tuned for specific durations or contexts. For example, honey bees communicate using vibration signals that vary in duration (Gil and De Marco, 2010; Ramsey et al., 2018; Seeley et al., 1998). Therefore, honey bees may be more proficient at interval timing in modalities with higher ecological relevance such as when processing mechanosensory signals. As temporal information in visual signals may not be ecologically relevant to bees, bees may consequently ignore or filter out temporal information from their visual system to reduce unnecessary noise in information processing.

Our study is the first to investigate the relational learning capacity of bees involving temporal information. These findings extend the evidence that honey bees appear to have a limited interval timing capacity and cannot use light duration as a cueing stimulus as efficiently as colour signals. Interval timing in insects remains largely understudied and therefore future work on this subject would clarify the ecological benefit of time sense in insects and other invertebrates.

We thank Devi Stuart-Fox for her valuable support and comments for this research. We also thank Hee-won Ham for assisting in the construction of pilot equipment, and Stuart McFarlane for building and programming the LED light tunnel.