Anne Reboul’s contribution to this conference has open the debate on causality in terms of the distinction between associative and explanatory behavior. It is argued that most animals (the example of Anne’s dog is illustrative) are left with the association, “naked as if it were” while Humans capable of causal learning build explanations that bring them back from the effects to the causes, as a way to grasp the causal relationship between events. Retrospection would be therefore a critical aspect in causal knowledge (see also John Watson’s comment on Anne Reboul’s article). Moreover, the distinction between extracting the relationship between observable and unobservable entities is also invoked as an essential point to distinguish between associative and causal learning. It is argued that associative learning allows extracting predictive relationships between perceptible entities in the world while causal learning allows going beyond the observable and learning about stimuli that are not present.

I will concentrate here on experiments on non-elemental forms of learning in honeybees. My objective will be to provide experimental evidence allowing to discuss a statement that was at the origin of this conference, namely that causality is at the base of the acquisition and use of categories and concepts (see conference primer). I will focus on chosen examples of categorization and rule learning in honeybees. I maintain that different forms of cognitive processing underlie these two forms of learning despite experimental commonalities. While an elemental associative account can explain categorization performances, rule learning requires a different explanatory basis. I argue that prospective / predictive analysis applies to categorization but that retrospection may be involved in rule learning. As honeybees can solve both kinds of problems, demarcating between species on the basis of these capacities seems inappropriate.

Causality as the cognitive basis for the acquisition of categories?

Categorization refers to the classification of perceptual input into defined functional groups (Harnard 1987). It can be defined as the ability to group distinguishable objects or events on the basis of a common attribute or set of attributes, and therefore to respond similarly to them (Troje et al 1999; Delius et al 2000; Zentall et al 2002). Categorization deals, therefore, with the extraction of these defining attributes from objects of the animal’s environment. Our use of the term categorization will be restricted to those cases in which animals transfer their choice to novel stimuli that they have never met before on the basis of common features shared with known stimuli.

The question of whether an insect brain can categorize visual objects in its environment has been recently answered affirmatively (Giurfa et al. 1996). Although this capacity can appear as surprising in the case of insects, considering that categorization obeys simple associative rules should allow demystifying this performance. We will restrict our examples to the visual modality as this is the modality in which categorization studies have been performed in the honeybee.

The honeybee constitutes a good model for addressing the question of visual categorization due to its remarkable learning and memory capabilities for visual stimuli (Menzel and Giurfa 2001; Giurfa 2003). Bees can be easily trained to fly towards a visual target on which a reward of sucrose solution is delivered by the experimenter. The associations built in this context link visual stimuli and reward, but also the response of the animal (e.g. landing) and reward, i.e. bees learn that a given visual cue (e.g. a color) will be associated with a reward of sucrose solution and that they have to land on it to get the reward. Using this basic design in which procedural modifications can be introduced, several studies have shown recently the ability of visual categorization in honeybees trained to discriminate different patterns and shapes. As mentioned above, such a demonstration requires that bees are able to transfer appropriate responding to novel stimuli belonging to the trained category.

Such a transfer has been demonstrated for a variety of visual features. For instance, van Hateren et al (1990) trained bees to discriminate two given gratings presented vertically and differently oriented (e.g. 45° vs. 135°) by rewarding one of these gratings with sucrose solution and the other not. Each bee was trained with a changing succession of pairs of different gratings, one of which was always rewarded and the other not (Fig. 1). Despite the difference in pattern quality, all the rewarded patterns had the same edge orientation and all the non rewarded patterns had also a common orientation, perpendicular to the rewarded one. Under these circumstances, the bees had to extract and learn the orientation that was common to all rewarded patterns to solve the task. This was the only cue predicting reward delivery. In the tests, bees were presented with novel patterns, which they were never exposed to before, which were all non-rewarded, but which exhibited the same stripe orientations as the rewarding and non-rewarding patterns employed during the training. In such transfer tests, bees chose the appropriate orientation despite the novelty of the structural details of the stimuli. Thus, bees could categorize visual stimuli on the basis of their global orientation. This conclusion led to a model of orientation detection in the honeybee, based on the existence of three types of orientation detectors, with a defined preferred orientations and tuning (Srinivasan et al 1994), comparable to those available in the mammalian visual cortex (Hubel and Wiesel 1962). Such detectors were found later by means of electrophysiological recordings in the visual areas of the bee brain (Yang and Maddess 1997).

Figure 1:Categorization of edge orientation by honeybees. (a) Training stimuli (P1 to P10) used in van Hateren et al’s experiments (1990). Pairs of stimuli were presented in a random succession to the bees. Within each pair, one was oriented at 45° and the other at 135°. In this case, gratings oriented at 45° were rewarded with sucrose solution while those at 135° were non rewarded. b) Tests performed with stimulus pairs not used during the training. In each case, there was a significant preference for the pattern presenting the orientation rewarded during the training. Bars indicate the proportion of choices for each stimulus. Bees transferred their choice from the known to the novel patterns and classified them according to their orientation (from van Hateren et al 1990).

Thus, honeybees show positive transfer of appropriate responding from a trained to a novel set of stimuli, and their performances are consistent with the definition of categorization. Visual stimulus categorization is not, therefore, a prerogative of certain vertebrates. However, is it so surprising and do causality judgments (sensu Reboul) underlie this ability, as stated at the origin of this conference? I don’t think so. In my opinion, categorization does not reflect retrospective analysis of events but results from simple associative learning. To explain this view, the neural mechanisms underlying categorization could be considered, in particular with respect to the organization of the bee brain.

If we admit that visual stimuli are categorized on the basis of specific features such as orientation or symmetry, the neural implementation of category recognition could be relatively simple. The feature(s) allowing stimulus classification would activate specific neuronal detectors in the optic lobes, the visual areas of the bee brain. Examples of such feature detectors are the orientation detectors whose tuning and orientation have been already characterized by means of electrophysiological recordings in the honeybee optic lobes (Yang and Maddess 1997; see above). Thus responding to different gratings having a common orientation of, say, 45°, is simple as all these gratings will elicit the same neural activation in the same set of orientation detectors despite their different structural quality. In the case of category acquisition, the activation of an additional neural element is needed. Such element would be necessary and sufficient to represent the reward (sucrose solution) and should contact and modulate the activity of the visual feature detectors in order to assign value to appropriate firing. This kind of neuron has been found in the honeybee brain as related to the olfactory circuit. VUMmx1 is a neuron present in the honeybee brain that receives its name from its localization (the name is the abbreviation of “ventral unpaired median neuron of the maxillary neuromere 1”). The dendrites of VUMmx1 arborize symmetrically in the brain and converge with the olfactory pathway at different sites (Hammer 1993). The essential property of VUMmx1 is that it responds to sucrose solution delivered both at the antennae and the proboscis of the bee with long lasting spike activity (Hammer 1993). Furthermore, the activity of this neuron constitutes the neuronal representation of reward in the case of olfactory learning as shown by the fact that bees can learn an olfactory stimulus which was paired with an artificial depolarization of VUMmx1 instead of sucrose reward (Hammer 1993). Other VUM neurons whose function is still unknown are present in the bee brain. It could be conceived that one of them (or more than one) contact the visual circuit to function as reinforcement in associative visual learning. Category learning could be thus reduced to the progressive reinforcement (through Hebbian rules, for instance) of an associative neural circuit relating visual-coding and reinforcement-coding neurons, similar to that underlying simple associative (e.g. Pavlovian) conditioning.

Thus, caution is needed before relating categorization to causality, i.e. to explanatory, retrospective behavior. One of the original statements of this conference, namely that “causality is the cognitive basis for the acquisition and the use of categories and concepts” may not hold, as indicated by experiments on honeybee visual categorization. Categorization, even if viewed as a higher-order cognitive performance, may simply rely on elemental links between conditioned and unconditioned stimuli. It may thus be based on prospective/predictive analysis and not on retrospection.

Causality as the cognitive basis for rule learning?

Like categorization, rule learning also presupposes positive transfer of an appropriate response from a known set to a novel set of stimuli. Despite this common experimental basis, I maintain that these processes do not rely on common mechanisms. In rule learning, the animal bases its choice, not on the perceptual similarity between the novel and the known stimuli which may not share, contrarily to categorization problems, any common feature, but on links that transcend the stimuli used to train it. Examples of such rules are “larger than”, or “on top of”, which may apply to stimuli which do not share any common feature but which can nevertheless be classified following the rule. I maintain that simple, elemental associative links cannot account for success in rule learning and that retrospective/diagnosis may be necessary to solve this kind of problem.

An example of rule learning is the learning of the so-called principles of sameness and of difference. These rules are usually uncovered through the delayed matching to sample (DMS) and the delayed non-matching to sample (DNMS) experiments, respectively. In DMS, animals are presented with a sample and then with a set of stimuli, one of which is identical to the sample and which is reinforced. As the sample is being changed regularly, they have to learn the sameness rule ‘choose always what is shown to you (the sample), independently of what is shown to you’. In DNMS, the animal has to learn the opposite, i.e. ‘choose always the opposite to what is shown to you (the sample)’. The interesting point concerning these protocols is that predictive analysis based on stimulus or feature generalization does not necessarily hold as the rule is ideally independent of the physical nature of the stimuli used. To discover the rule, the animal has to operate on the set of examples known such that retrospection and different forms of heuristics can be applied to solve the problem. Neural accounts based on simple associative networks such as that proposed for visual categorization (see above) may not be valid in this case.  Although reinforcement can still be represented by a specific neural pathway or element (such as the VUMx1 neuron or its equivalents; see above), the novel, differing sample (e.g.; a color) will not activate the same network components responding to a previous sample (e.g., an odor). Extracting the rule in a changing learning set means therefore going beyond stimulus modality and performing a form of retrospective or diagnostic analysis of the problem faced.

Is this kind of problem a good candidate for species demarcation? The answer is no. Honeybees foraging in a Y-maze learn to solve both DMS and DNMS rules (Giurfa et al. 2001). Bees were trained in a DMS problem in which they were presented with a changing non-rewarded sample (i.e. one of two different color disks or one of two different black-and-white gratings, vertical or horizontal) at the entrance of a maze. The bees were rewarded only if they chose the stimulus identical to the sample once within the maze. Bees trained with colors and presented in transfer tests with gratings that they have not experienced before solved the problem and chose the grating identical to the sample at the entrance of the maze. Similarly, bees trained with the gratings and tested with colors in transfer tests also solved the problem and chose the novel color corresponding to that of the sample grating at the maze entrance. Transfer was not limited to different kinds of modalities (pattern vs. color) within the visual domain but could also operate between drastically different domains such as olfaction and vision (Giurfa et al. 2001). Furthermore, bees also mastered a DNMS task, thus showing that they also learned a principle of difference between stimuli (Giurfa et al. 2001). In both DMS and DNMS, win-stay/ loose-shift (or win-shift / loose-stay) strategies could not account for the performances of the bees. These results document that bees learn rules relating stimuli in their environment. The capacity of honeybees to solve DMS tasks has been verified in other contexts (see, for instance Zhang et al. 2004, 2005). In particular, introducing longer delays between the offset of the sample and the onset of the comparison stimuli yielded a decay in matching performances and allowed to suggest that honeybees retrospectively code the samples in delayed matching-to-sample task (Zhang et al. 2005).

Retrospective revaluation in honeybees

Associative learning theories account for causal and predictive learning but face the problem of retrospective learning as for such theories, learning can only occur when a stimulus is present. In this sense, the case of backward blocking (Shanks 1985) is interesting as it implies training an animal with a compound stimulus AB reinforced (AB+) in a first phase, and then with A reinforced (A+) in a second phase; if backward blocking occurs, ratings of B in a third test phase are reduced retrospectively by experience with A in the second phase because A alone is enough to predict the outcome of AB. This experiment is therefore interesting because it can be related to the claim made in this conference on the distinction between associative and causal learning. As mentioned in the first paragraph, in Anne Reboul’s contribution it is suggested that associative learning allows extracting predictive relationships between perceptible entities in the world while causal learning allows going beyond the observable and learning about stimuli that are not present. If this is the case, then backward blocking could be an interesting case as animals learn retrospectively about stimulus B.

Backward blocking has been recently studied in honeybees (Blaser et al. 2004). It was shown that responding to B after AB+training was less in animals that also had A+ training than in control animals that were equally often reinforced in the absence of A. Furthermore, responding to B was less after AB+followed by differential training A+C- than after AB+followed by C+A- training. In the first case (AB+, A+C-), retrospective revaluation would decrease the value of B as A is a reliable cause or predictor of the outcome during the compound training. In the second case (AB+, A-C+), retrospective revaluation would have the opposite effect, i.e. it would enhance the value of B as reliable cause or predictor of the outcome during the compound training.

Caution is nevertheless needed when analyzing these data in the light of a possible dichotomy between associative and causal learning. It is worth mentioning that associative accounts have been provided to explain retrospective revaluation. For instance, Van Hamme and Wasserman (1994) suggested that when A and B are paired with the outcome in Phase 1, a within-compound association is formed between them, which then allows the presentation of A in Phase 2 to activate the representation of B. The predictive strength of an expected but absent cue decreases. Therefore, when A alone is followed by the outcome in Phase 2, the associative strength for A increases while the associative strength for the absent cue B simultaneously decreases. In this case, therefore, the proposed distinction between associative learning, which allows extracting predictive relationships between perceptible stimuli in the world, and causal learning, which allows learning about stimuli that are not present, may not be so straightforward.

Conclusions

Although honeybees, as Anne Reboul’s dog, are sometimes left with the associations ‘naked as if they were”, they can also operate on associations between events in their environment in order to extract rules and retrospectively evaluate stimuli and their outcomes. It seems therefore that bees have expectations based on associative, predictive learning but that such learning is not the whole of honeybee cognition. Clearly, research articulated on categorization and rule learning may be useful to distinguish between different levels of complexity of cognitive processing but not to determine what is unique to Humans. This implies, therefore, that either diagnostic / retrospective learning is not unique to Humans, or that such uniqueness resides elsewhere, for instance in the existence of language.

References

Blaser RE, Couvillon PA Bitterman ME (2004) Backward blocking in honeybees. Q J Exp Psychol B 57: 349-60.

Giurfa M, Eichmann B, Menzel R (1996) Symmetry perception in an insect. Nature 382: 458-461

Giurfa M, Zhang S, Jenett A, Menzel R, Srinivasan MV(2001) The concepts of 'sameness' and 'difference' in an insect. Nature 410: 930-933

Giurfa M (2003) Cognitive neuroethology: dissecting non-elemental learning in a honeybee brain. Curr Opin Neurobiol 13: 726-735

Hammer M (1993) An identified neuron mediates the unconditioned stimulus in  associative olfactory learning in honeybees.Nature 366: 59-63

Harnard S (1987) Categorical Perception. The Groundwork of Cognition. CambridgeUniversityPress, Cambridge

Hateren JH v, Srinivasan MV, Wait PB (1990) Pattern recognition in bees: orientation discrimination. J Comp Physiol A 197: 649-654

Hubel DH, WieselTN(1962) Receptive fields, binocular interaction and functional architecture in the cat's visual cortex. J Physiol (London)160: 106-154

Menzel R, Giurfa M (2001) Cognitive architecture of a minibrain: the honeybee. Trends Cognit Sci 5: 62-71

Shanks, D. R. (1985). Forward and backward blocking in human contingency judgement. Quart J Exp Psychol 37B: 1-21.

Srinivasan MV, Zhang SW, Witney K (1994) Visual discrimination of pattern orientation by honeybees: Performance and implications for “cortical” processing. Phil Trans Royal Soc Lond (B) 343: 199-210

Troje F, Huber L, Loidolt M, Aust U, Fieder M (1999) Categorical learning in pigeons: the role of texture and shape in complex static stimuli. VisRes 39: 353-366

Van Hamme LJ, Wasserman EA (1994) Cue competition in causality judgments: The role of nonpresentation of compound stimulus elements. Learn Motivation 25: 127-151

Yang EC, Maddess T (1997) Orientation-sensitive neurons in the brain of the honey bee (Apis mellifera). J Insect Physiol 43: 329-336

Zentall TR, Galizio M, Critchfield TS (2002) Categorization, concept learning and behavior analysis: an introduction. J Exp Anal Behav 78: 237-248

Zhang SW, Srinivasan MV, Zhu H, and Wong J (2004). Grouping of visual objects by honeybees. J Exp Biol  207: 3289-3298.

Zhang S, Bock F, Si A, Tautz J, Srinivasan MV (2005) Visual working memory in decision making by honey bees. PNAS 102:5250-5255.