Our study illustrates the robustness of bumblebee learning. We challenged bees with a free-flight conditioning task. This was learned quickly and presenting bees with conflicting information seen on arrival and departure from the feeder had minimal impact on either the rate of learning (Fig. 2a), or on the specificity of what had been learned. Learning of a horizontal CS + was more affected by conflicting information than a vertical CS-.
Our assay had features of trace conditioning, since our stimuli could not been seen by bees when they were feeding, hence the CS did not overlap with attaining the US. Trace conditioning is considered a cognitive form of learning and is even considered by some as evidential of conscious processing (Clark and Squire, 1998; Lovibond and Shanks, 2002, Birch et al. 2020, Droege et al. 2021), but it is a robust finding in insects (Menzel 2001, Dylla et al. 2013, Perisse et al. 2011, Klappenbach et al. 2021, Paoli et al. 2023). Lehrer was the first to show bees learn information seen on both arrival and departure from a sucrose solution feeder (1991, 1993). She found that if bees were presented with different stimuli on arrival and departure from a feeder their learning rate slowed. While there was evidence bees could learn a stimulus seen on departing a feeder, they showed a prioritisation of the stimulus seen before feeding (Bitterman and Couvillon 1991, Lehrer 1993). Our work differs from Lehrer (1993) in that while she presented bees with two different stimuli on arrival and departure (essentially two CS+), we used a discriminant learning paradigm and presented some bees with a conflict situation (CS + seen on arrival, CS- seen on departure). In this case, we saw no reduction in learning rate when compared to learning a consistent CS + for a vertical CS + stimulus, and only a minor reduction in learning rate for learning a horizontal CS+. Similarly, in generalisation tests the conflicting information had minimal impact. It is clear, therefore, that, when presented with both the CS + flipped to the CS- on departure from a sucrose feeder bumblebees did not generalise between the two stimuli, nor was there interference between the two stimuli. Bees in the switching groups appear to prioritise the relevant CS + information and entirely disregard the conflicting CS- information, but we may not need to invoke cognitive concepts such as “prioritisation” to explain our findings.
The most plausible anatomical locus for the associative learning phenomena studied here are the mushroom bodies (Barth et al. 1997, Li L. et al. 2017). The Kenyon cells of the mushroom bodies receive processed sensory input, and output from premotor regions (Mobbs 1982; Fahrbach 2006). There is experience-dependent neuroplasticity at both the input and output of the Kenyon cells that is sensitive to neurochemicals released in response to appetitive or aversive reinforcers (Barnstedt et al. 2016). It is theoretically possible for the mushroom body to support trace conditioning (Menzel 2001, Menzel & Giurfa 2001). Certainly, an enduring “trace” of neural activation can be held by the mushroom body structure for a short period of time. The Kenyon cells have a prolonged accommodation property (Strausfeld et al. 2009), and in Drosophila, recurrent connections have been detected between Kenyon cells (Dylla et al., 2013, Lyutova et al. 2019, Chandra et al. 2010, Aso et al. 2014, Bennett et al. 2021). These could, in theory, support a reverberation of neural activity in the Kenyon cell populations. Either or both mechanisms could maintain a trace of neural activity that persists beyond the presentation of a stimulus. This could support elementary forms of trace conditioning.
Learning stimuli on departure from the feeder most likely also involves the mushroom bodies. In classic associative learning theory, a CS that comes after the US is typically not learned since it is not predictive of the occurrence of the US. And yet, bees demonstrate a specific behaviour – the turn back and look – at a feeder on departure and learn features of a feeder during this behaviour. This form of learning could either be a form of secondary reinforcement or latent learning (Menzel 2001). Secondary reinforcement would assume that the feeder station and/or feeder location has become a reinforcer following pairing with food reward, in which case the feeder could now act as a conditioned reinforcer for any view directed at the feeder. Latent learning is simply learning with no explicit reinforcer and is presumed to be important for many forms of spatial learning. Both secondary reinforcement and latent learning are believed to involve the mushroom bodies in conjunction with the spatial systems of the lateral accessory lobes (Wystrach et al. 2023).
If mushroom bodies are involved in learning the stimuli seen both before and after feeding, how is it that learning performance is largely unchanged even if this information conflicts? In terms of the robustness of bees to learning conflicted information, here we should consider the mechanisms of decision making in bees as well as the learning mechanisms. Ultimately, the outcome of learning is to influence a decision of whether a bee should land at a feeder marked by a horizontal or vertical stimulus. The mushroom body alone is not a decision maker (Galizia 2014, Bhazenov et al. 2013, Huerta et al. 2004, 2009). It can perhaps best be thought of a as a classifier – learning to associate presented stimuli with different outcomes which are conveyed by mushroom body output neurons to premotor regions (Galizia 2014, Maboudi et al. 2023). The punished stimuli were consistent in all groups therefore the rate of learning to avoid the CS- would be the same in all groups. In both the switching and constant groups, the CS + was seen on approach to the feeder, therefore in all groups the CS + was reinforced for approach behaviour only, whereas the CS- stimulus would be reinforced for avoidance of punished stimuli in all groups and departure from the CS + in the switching groups. If we consider the mushroom body as classifying stimuli by behavioural response, this alone is sufficient to resolve any conflicting information associated with a feeder. In our paradigm, the CS + was only associated with approach responses, regardless of training groups.
In this experiment, learning of a horizontal CS + was more disrupted by the switching manipulation than learning of a vertical CS+. Why this might be is not clear, but there are other reports of insects responding differently to vertical and horizontal stimuli or learning them at different rates (Srinivasan et al. 1999, Wang, Tie et al. 2014, Wolf et al. 2015). It is possible these are processed differently by the visual system or have different innate responses.
In summary, our study demonstrates remarkable speed and proficiency for bumble bees learning a trace conditioning paradigm. Their learning was rapid, specific and largely unaffected if the CS + feeder was linked with conflicting information. Our study speaks to the remarkable efficacy of the bee brain for learning food related stimuli.