ANIMAL BEHAVIOUR, 2002, 64, 175184 doi:10.1006/anbe.2002.3055, available online at [Link]

com on

Nonelemental visual learning in honeybees

*Laboratoire de Cognition Animale, Universite Paul Sabatier Neurobiology, Institute of Biology, Free University of Berlin, Germany Department of Psychology, Philipps-University of Marburg (Received 22 January 2001; initial acceptance 10 April 2001; final acceptance 19 January 2002; MS. number: 6816R)

Free-flying honeybees, Apis mellifera, learn visual stimuli in the appetitive context of food search. Visual compound stimuli are relevant in nautre as bees learn flower images that consist of many visual elements. We studied whether elemental associations between each visual element and the reinforcement (elemental approach) are enough to explain the solving of visual discrimination problems that raise ambiguity at the elemental level. We asked whether bees could solve three different visual discriminations: (1) positive patterning (A , B , AB+); (2) negative patterning (A+, B+, AB ); and (3) biconditional discrimination (AB+, CD+, AC , BD ). In experiments 1 and 2 bees had to discriminate a yellowviolet chequerboard from the yellow or the violet squares alone. In experiment 3, four different gratings combining one colour (yellow or violet) with one orientation (vertical or horizontal) had to be discriminated. In all three problems binary compounds were trained in such a way that each element appeared equally often as rewarded and nonrewarded. Bees could solve the three discrimination problems. They always chose the reinforced stimulus despite ambiguity at the level of the elements. For solving positive patterning, elemental processing could be used. For negative patterning and biconditional discrimination, nonelemental processing strategies (unique-cue or configural approach) are necessary to account for these results. Although we cannot decide between a configural and a unique-cue interpretation, we can clearly reject purely elemental processing in these cases.

2002 The Association for the Study of Animal Behaviour. Published by Elsevier Science Ltd. All rights reserved.

The capacity to learn relationships between events in the environment is of central importance for successfully negotiating a complex world. Animals can learn that an originally neutral stimulus (conditioned stimulus, CS) can act as a predictor for a meaningful stimulus (unconditioned stimulus, US). Such an elemental association has been found in a great variety of animals and constitutes the basis of classical conditioning (Pavlov 1927). In the natural world, however, stimuli rarely occur in isolation. They usually form compounds that enter into an association with a US. Although it is clear that animals can process and learn such compounds, the nature of the associations enabling such learning is still debated. Two main approaches account for compound processing and learning: (1) an elemental approach (e.g. Rescorla & Wagner 1972) assuming that a compound AB is represented as two elements A and B, each of which becomes connected to the representation of the US (i.e. the whole equals the sum of its parts) and (2) nonelemental
Correspondence: M. Giurfa, Laboratoire de Cognition Animale, Universite Paul Sabatier, 118 Route de Narbonne, 31062 Toulouse, France (email: giurfa@[Link]). H. Lachnit is at the Department of Psychology, Philipps-University of Marburg, Gutenbergstr. 18, D-35032 Marburg, Germany.
00033472/02/$35.00/0

approaches assuming that the representation of a compound AB differs from the simple sum of the individual representations of the elements A and B (i.e. the whole is different from the sum of its parts). Among the latter, two alternatives can be cited: (1) the unique-cue approach (Rescorla 1972, 1973; Whitlow & Wagner 1972) and (2) the configural approach (Pearce 1987, 1994; Rudy & Sutherland 1992, 1995). The unique-cue approach (Rescorla 1972, 1973; Whitlow & Wagner 1972) assumes that a compound consists of its elements plus an additional configural stimulus that is unique to the compound, but that can otherwise be dealt with as an additional element (i.e. the whole is the sum of the elements plus the unique cue). This unique cue codes the common presentation of a certain configuration of stimuli. The response to a compound may thus be analysed as the sum of the associative strengths of the elements plus that of the unique cue. Although elemental associations are also invoked in the unique-cue approach, it constitutes a nonelemental approach as problem solving cannot be explained purely on the basis of the physical presence of the elements of a compound. The configural approach (Pearce 1987, 1994) assumes that the elements of a compound collectively enter into a

175 2002 The Association for the Study of Animal Behaviour. Published by Elsevier Science Ltd. All rights reserved.

176

ANIMAL BEHAVIOUR, 64, 2

single association that constitutes an entity different from the elemental representations (i.e. the whole is different from the sum of its parts). According to this theory, generalization of associative strength between the elements and the compound occurs depending on their similarity. An extreme configural theory adheres to the basic assumption of the configural theory but pushes the reasoning to the extreme of postulating that a compound is coded as a new stimulus that is totally unrelated to its elements (for a comparison of the extreme configural and configural point of view see Williams & Braker 1999). Studying the animals performance in patterning experiments may help in deciding between the elemental and the nonelemental options. The decisive aspect of a patterning experiment depends on whether an element appears alone, or together with another element as a compound. In a negative patterning discrimination, for instance, two single stimuli are reinforced (A+, B+), while the compound is not (AB ). Animals should learn to respond to the elements but not to the compound. Mastering this task can be explained only if nonelemental associations are invoked, that is, if the compound is a new entity that is different from its elements (configural approach) or if a unique cue is present in addition to the elements (unique-cue approach). If only the associative strengths of the elements were summed during compound presentation, the animals would never learn to suppress their reaction to the compound and would always respond to it. In a positive patterning discrimination, on the other hand, the compound is reinforced (AB+) while the single elements are not (A , B ). In this case, discrimination could be solved through elemental associations because the associative strength of the elements could be subthreshold for the response but once added upon compound presentation they might result in a suprathreshold associative strength. Such a summation would yield higher associative strength and, therefore, higher responsiveness to the compound. Therefore results of positive patterning are not necessarily conclusive in analyses of nonelemental processing and learning. In contrast, biconditional discrimination allows one to determine whether animals can process and learn compounds in nonelemental terms. In such a paradigm (Saavedra 1975), four elements (A, B, C, D) are arranged to form four binary compounds (AB, CD, AC, BD), two of which are followed by reinforcement (AB+, CD+) and two are not (AC , BD ). In this situation, it is impossible to assign one element to reinforcement or to nonreinforcement because every element is followed equally often by both outcomes. The compounds, however, are unambiguous, and correct predictions can be made if they, rather than their elements, become associated with the respective outcome. The honeybee, Apis mellifera, is a valuable model for studying the problem of elemental versus nonelemental associations. Honeybees learn a great variety of sensory cues associated with their nest and their food sources, flowers (Menzel et al. 1993; Menzel & Mu ller 1996; Menzel & Giurfa 2001). The ecological conditions under which learning and memory formation take place are well defined (Menzel 1985, 1999) and, therefore, the issue of

nonelemental learning has been addressed with bees as experimental subjects. Classical olfactory conditioning of the proboscis extension reflex (Takeda 1961; Bitterman et al. 1983) has been used to study whether bees learn olfactory compounds in nonelemental terms. When the antennae of a hungry, harnessed honeybee are touched with sucrose solution (US), the insect reflexively extends its proboscis. If the presentation of an originally neutral olfactory stimulus (CS) to the antennae is paired with sucrose solution (US), an association is formed that enables the odour to release the proboscis extension reflex in a subsequent test. Nonelemental learning of olfactory compounds has been shown using this paradigm (Hellstern et al. 1995; Chandra & Smith 1998; Mu ller et al. 2000; Deisig et al. 2001). Conditioning of free-flying honeybees can also be used to study the problem of elemental versus nonelemental associations. In this case, bees are trained to fly to a visual target (see Giurfa et al. 1997 for review) to receive a reward of sucrose solution on it. Couvillon & Bitterman (1988) trained bees to solve a negative and a positive patterning discrimination problem using stimuli of two different modalities, a green ring (colour) and a peppermint scent (odour) delivered through small orifices on the cover of a petri dish. Their results support the interpretation that odourcolour compounds were processed in a configural fashion. We studied the problem of elemental versus nonelemental associations in free-flying bees using only visual stimuli. We used coloured patterns because colour vision in the honeybee is well studied (see review in Menzel & Backhaus 1991). Colour coding in the bee brain is also relatively well known (see Menzel & Backhaus 1991 for review) and several models for honeybee colour vision have been established (Backhaus 1991; Vorobyev & Brandt 1997). This allows us to quantify precisely the psychophysical properties of the elemental and compound stimuli (Giurfa et al. 1997). We studied whether bees process and learn visual compounds in nonelemental terms. We asked whether bees could solve three different visual discriminations: (1) positive patterning; (2) negative patterning; and (3) biconditional discrimination. In all three problems binary compounds were trained in such a way that each element appeared equally often as rewarded and nonrewarded. At least in cases 2 and 3, problem solving, if any, thus required nonelemental processing and learning. GENERAL METHODS We trained free-flying honeybees, A. m. mellifera carnica, to enter an experimental set-up to collect a food reward (50% sucrose solution). Foragers individually marked with coloured spots on the thorax and/or abdomen were rewarded on achromatic discs and progressively brought to the set-up. No chromatic stimuli were used in pretraining. We did the negative patterning experiment (NP) during summer 1999 and the positive patterning experiment (PP) and the biconditional discrimination experiment (BiDi) during summer 2000. Thus, a direct comparison

SCHUBERT ET AL.: VISUAL LEARNING IN BEES

177

(a)

Entrance chamber

(c) Decision chamber Back wall (20 20 cm) UV-transparent ceiling

Entrance (sliding door) Entrance hole

20 cm 20 cm Entrance chamber 20 cm Decision chamber 20 cm Decision chamber Decision line Back wall (20 20 cm) Entrance chamber 20 cm

(b)

Entrance (sliding door) Entrance hole

20 cm

20 cm

Dividing wall

30 cm

Figure 1. The experimental set-up (see text for details). In the decision chamber, bees faced one of two situations during training: (a) a single stimulus was offered and rewarded or (b) two stimuli were offered and only one was rewarded. (c) Front view of the set-up in a dual-choice situation.

between NP and PP is not possible. However, withinexperiment analyses are possible and allow us to determine whether bees process and learn colour compounds in a nonelemental way.

Apparatus
The apparatus (Fig. 1) consisted of two connected wooden chambers: an entrance and a decision chamber. A sliding door allowed the bees to enter the entrance chamber, one at a time. The entrance and the decision chamber were separated by a wall with a circular orifice, 5 cm in diameter, in the middle, through which the bee could enter the decision chamber. The whole apparatus was covered with an ultraviolet-transparent ceiling. In the decision chamber bees faced one of two situations during training. (1) The decision chamber presented a unique, rewarded stimulus in the middle of the back wall (20 20 cm). Bees had to fly to the target and land on it to get the reward. (2) The decision chamber presented two different stimuli simultaneously. In this case, its distal part was divided by an upright wall (8 20 cm), perpendicular to the back wall. Bees had thus to decide between two subcompartments (each 10 20 cm), each with a stimulus presented in the centre of its back wall. One was rewarded and the other was not. In both situations, the reward was delivered in a plastic micropipette, 0.3 cm in diameter, in the centre of the stimulus. Such a pipette could not be resolved by the bees eyes when entering the decision chamber. The nonreinforced stimulus also presented an empty plastic micropipette. During tests, bees had to choose between two nonrewarded stimuli. The decision chamber was slightly modified with respect to the dual-choice situation used

during training. We replaced the original dividing wall (8 cm long) by a shorter one (4 cm). This was done to facilitate the switch between the left and right subcompartments. In all cases the experimenters position was behind the maze to avoid interfering with the arrival and choices of the bees. Once in the maze and before making their choices, the bees could not see the experimenter. In addition, to avoid a bias in the bees choices determined by seeing the experimenters position before entering the maze, the experimenter always stood aligned with the central hole entrance of the maze (see, for example, the dashed line in Fig. 1a).

Stimuli
We used colours cut from HKS-1N and HKS-36N papers (Hostmann-Steinberg, Celle; K+E Druckfarben, Stuttgart; H. Schminke & Co., Erkrath, Germany) that were yellow and violet for humans, respectively. All stimuli were presented on a neutral, grey background provided by a HKS-92 N paper. Grey was never specifically associated with the presence or absence of reinforcement. Figure 2 shows the spectral reflection curves of stimuli and background (Fig. 2a) and their loci in the colour space (colour opponent coding space) of the honeybee (Backhaus 1991; Fig. 2b). This two-dimensional diagram represents for every colour stimulus the excitations of the two kinds of colour opponent cells, A and B, responsible for the coding of colour in the honeybee brain. The space allows one to read the perceptual difference between colours according to the City-Block metric (Backhaus 1991). Yellow and violet could be perfectly differentiated by the bees as shown by their separated loci in the colour space (Fig. 2b). Furthermore, they presented a similar

178

ANIMAL BEHAVIOUR, 64, 2

(b) 1.0 (a) 0.8 Spectral reflectance 0.6 0.4 300 0.2 0.0 300 Yellow Violet Grey 400

*
Violet

*
Yellow *

Grey 500

400 500 600 Wavelength (nm)

700

Figure 2. Spectral properties of the coloured stimuli (yellow and violet) and background (grey). (a) Spectral reflection curves of yellow (HKS-1N paper), violet (HKS-36N) and grey (HKS-92). (b) Loci of the elementary colours and background in the colour opponent coding diagram. The diagram represents the responses of two types of colour opponent coding cells, A and B. The origin represents the grey background. The closed line gives the loci of spectral colours in 10-nm steps, and the mixtures of 300 and 550 nm marked in 10% steps.

perceptual salience as both had a similar amount of chromatic contrast to the grey background (ca. 3 COC units).

Procedure
We began each experiment by pretraining a group of individually marked foragers to enter the maze to collect sucrose solution. During training and tests only one bee was present at a time in the apparatus as the other pretrained bees were captured in a flight cage before the experiment started. We excluded recruited bees by closing the sliding entrance door. At the entrance chamber, the bee could not see the stimuli that were presented on the back walls of the subsequent decision chamber. Once the bee entered the decision chamber by flying through its entrance hole, she could find either a single, rewarded stimulus (NP and PP) or two different stimuli, one rewarded and the other not (NP, PP and BiDi). In the first case, the bee simply flew along the nondivided chamber and landed on the stimulus to get the reward. In the second case, she had to decide between the stimuli; if she chose the subcompartment with the reinforced stimulus, she was rewarded with sucrose solution; if she chose the subcompartment with the nonreinforced stimulus, she was immediately tossed away from the maze by hand before she reached it. The bee then had to re-enter through the entrance hole and make a new decision until the correct stimulus was chosen. In that way, wrong decisions were penalized and choice performance improved. Under such conditions (differential conditioning procedure) bees trained in a maze similar to the one used here learn the properties of both the reinforced and the nonreinforced stimulus (Giurfa et al. 1999). This is an important point as in all experiments reported here bees were expected to learn about both stimuli. When two stimuli were presented during training, the sides of the reinforced and the non-

reinforced stimulus were randomized to ensure that bees did not associate the reward with a particular position. In all cases, training and test stimuli were regularly replaced by fresh ones to avoid influences of olfactory cues. To avoid the bees learning visual stimuli as retinotopical fixed images, which implies memorizing a pattern on a pixel basis, a capacity repeatedly suggested in insects (Wehner 1972; Gould 1985; Dill et al. 1993; Giurfa et al. 1995a), we constantly varied, between trials, the spatial position of each element in the visual field of the bee. In addition, when presenting a compound (e.g. a chequerboard with two colours in NP and PP), we constantly rotated it on its central point, between trials, to present the coloured areas at different retinal positions. Chequerboards (in NP and PP) and gratings (in BiDi) were also constantly replaced by similar stimuli with a reversed arrangement of colours. In all experiments training followed a fixed number of trials (60 for NP and PP and 80 for BiDi). Previous experiments showed that this number of trials is enough to improve performance during acquisition. In all cases training took at least 1.5 (NP and PP) or 2 days (BiDi) per bee. After completing the training sequence, bees were tested in a dual-choice situation in which no reward was provided. In such tests, we presented fresh stimuli on exchangeable walls for 2 min to the bee under study. To control for position tendencies, we did each test twice, with the sides of the stimuli exchanged from one test to the other. To avoid extinction resulting from the absence of reward, we introduced a refreshment training period between tests. For this we presented each training combination once.

Response Measured and Statistics

SCHUBERT ET AL.: VISUAL LEARNING IN BEES

179

Relative choice frequencies for each test stimulus were calculated from the absolute number of decisions registered. We analysed preference for one stimulus within a test by means of a replicated goodness of fit G test (Sokal & Rohlf 1995), after verifying that heterogeneity (evaluated through a G test for heterogeneity) did not make a significant contribution to the data as a whole (Sokal & Rohlf 1995, pp. 722724). The responses to the elemental colours within an experiment were compared with a 2 2 G test (Sokal & Rohlf 1995). Absolute choice frequencies were used to compute G statistics. EXPERIMENT 1: POSITIVE PATTERNING In PP two single stimuli were nonreinforced (A , B ) whilst the compound of the same stimuli was reinforced (AB+). We asked whether bees learn to respond to a yellowviolet chequerboard and not to its elements, that is, a yellow chequerboard or a violet chequerboard (see Fig. 3). The latter resulted from presenting the arrangement of yellow or violet squares on the grey background.

high-quality laser printer on white copying paper of constant quality. It had the same diameter and square size as the chromatic chequerboards. The precise location of the reward was in the middle of each stimulus, between two adjacent yellow-square corners in the case of the elemental-yellow chequerboard, between two adjacent violet-square corners in the case of the elemental-violet chequerboard, and in the middle of the four central squares of the yellowviolet compound chequerboard.

Procedure
Each bee was trained with the yellowviolet chequerboard alone and reinforced (AB+), or on separate trials with a nonreinforced yellow (A ) or violet (B ) chequerboard presented against the reinforced black-andwhite chequerboard (XY+). During training, all stimuli (A , B and AB+) were presented 20 times in a pseudorandom sequence such that no stimulus was presented consecutively more than twice. The total number of training trials was thus 60. After completing the training trials, each bee was presented with the nonrewarded, dual-choice tests A versus AB and B versus AB in random succession. Between tests, a refreshment training of three visits (one with each patterning situation: A versus XY+; B versus XY+; AB+) was done to keep the bees foraging motivation high.

Methods
The schedule of patterning experiments had to be adapted to the particular situation of working with freeflying honeybees. The experimenter has no control over the bees decision to return to the experimental set-up and bees are extremely sensitive to decreases in reward quantity or quality (Nu n ez 1970, 1982) such that they may decide not to return to the set-up when confronted with negative experiences. It is therefore impossible to offer exclusively a pure nonreinforced situation, that is, offering a nonreinforced element alone in PP (A or B ) because the bee could interrupt foraging and thus the whole experiment. Therefore, we adapted our training schedule in such a way that whenever a nonreinforced element was presented, a reinforced alternative stimulus was presented simultaneously. The alternative stimulus, a black-and-white chequerboard (henceforth XY), was chosen to avoid any kind of generalization between it and the elemental colours or the coloured compound used. This stimulus was never presented during the tests. Owing to its lack of chromaticity and absence during the tests, its role was irrelevant for the outcome of the experiment. It simply ensured we could offer a reward if the alternative to it was nonrewarded. Thus, the training schedule of PP was A versus XY+, B versus XY+ and AB+.

Results
We recorded 746 choices of eight bees during the 2-min extinction tests. Figure 3a shows the bees performance in the tests confronting A versus AB and B versus AB. In both cases, bees significantly preferred the previously reinforced yellowviolet compound to the single element, either yellow or violet. The preference for the compound was ca. 60% in both cases (yellowviolet versus yellow: 60.72 versus 39.28%; G8 =31.20, P< 0.001; yellowviolet versus violet: 61.13 versus 38.87%; G8 =51.89, P< 0.001). There were no differences in the preference for yellow or violet between the two tests (2 2 G1 =0.04, NS). Thus, after being trained in a modified PP schedule (A versus XY+, B versus XY+, AB+), bees significantly preferred the compound chequerboard to the elemental chequerboards. EXPERIMENT 2: NEGATIVE PATTERNING In NP two single stimuli were reinforced (A+, B+) whilst the compound of the same stimuli was nonreinforced (AB ). We asked whether bees learn to respond to the yellow and violet chequerboards but not to the yellowviolet chequerboard.

Stimuli
Colour stimuli were made from coloured squares, 2 2 cm, and covered a circular area, 10 cm in diameter (Fig. 3). Each square was perfectly resolvable for a bee flying within the decision chamber of the maze. Thus, no perceptual colour mixture occurred. When the bee had to decide between two test options and therefore enter one of the two subcompartments of the chamber, the coloured squares appeared chromatic to it. The blackand-white chequerboard (XY) was produced with a

Methods
The schedule of NP was inverted with respect to that of PP. Thus, the training schedule was A+, B+, AB versus XY+. The colour stimuli were the same as those used in the PP experiment (Fig. 3).

180

ANIMAL BEHAVIOUR, 64, 2

Figure 4. Choice performance in the biconditional discrimination experiment. N =15 bees in tests CD+ versus AD (882 choices), AB+ versus CB (781 choices) and CD+ versus CB (711 choices); in the remaining test AB+ versus AD N =14 bees (756 choices). The stimuli used are shown at the right. *P< 0.001.

Figure 3. Choice performance in the A versus AB and B versus AB tests. A: yellow; B: violet; AB: yellowviolet chequerboard. The stimuli used are shown below. *P< 0.001. (a) Performance after positive patterning training. N =8 bees; 746 choices. (b) Performance after negative patterning training. N =8 bees, 473 choices.

yellow or violet elemental chequerboard to the yellow violet compound. The preference was higher for violet (violet versus yellowviolet: 90.69 versus 9.31%; G8 =154.34, P< 0.0001) than for yellow (yellow versus yellowviolet: 69.44 versus 30.56%; G8 =58.38, P< 0.001). The difference between the response elicited by the violet chequerboard and that elicited by the yellow chequerboard was significant (yellow versus violet: 2 2 G1 =33.25, P< 0.0001). Despite this asymmetry, bees in this experiment clearly preferred the elemental stimuli and not the compound, conversely to what happened in the PP experiment, where they preferred the compound to the elemental stimuli. Thus, after being trained in a modified NP schedule (A+, B+, AB versus XY+), bees significantly preferred the elemental chequerboards to the compound chequerboard. EXPERIMENT 3: BICONDITIONAL DISCRIMINATION BiDi involved, not two single elements as in the patterning experiments, but four different elements. These four elements (A, B, C, D) were arranged to form four binary compounds (AB, CD, AC, BD), two of which were reinforced (AB+, CD+) while the others were not (AC , BD ) (Saavedra 1975). In the case of free-flying bees it is possible to design a biconditional discrimination experiment using colour (yellow and violet; Y and V, respectively) and orientation (horizontal and vertical; h and v, respectively). Honeybees easily learn both kinds of discrimination separately (for colour discrimination see Menzel & Backhaus 1991; for orientation discrimination see van Hateren et al. 1990). In this experiment, one group of bees had to choose yellow if it was presented in a horizontal grating (Yh+) but not in an identical vertical grating (Yv ) and violet if it was presented in a vertical grating (Vv+) but not in an identical horizontal grating (Vh ), while in a second group of bees the contingencies were reversed (Yh , Yv+, Vv , Vh+). We thus asked whether bees could

Procedure
Each bee was trained with the nonreinforced yellow violet chequerboard (AB ) presented against the reinforced black-and-white chequerboard (XY+) or on separate trials with the yellow (A+) and violet (B+) chequerboards presented alone and reinforced. During training, all stimuli (A+, B+ and AB-) were presented 20 times in a pseudorandom sequence such that no stimulus was presented consecutively more than twice. The total number of training trials was thus 60. After completing the training trials, each bee was presented with the nonrewarded, dual-choice tests A versus AB and B versus AB in random succession. Between tests, a refreshment training of three visits (one with each patterning situation: A+, B+, AB versus XY+) was done to keep the bees foraging motivation high.

Results
We recorded 473 choices of eight bees during the tests. Figure 3b shows the performance of bees in the tests confronting A versus AB and B versus AB. In both cases, bees significantly preferred the previously reinforced

SCHUBERT ET AL.: VISUAL LEARNING IN BEES

181

learn to solve such discriminations, a fact that would be indicative of nonelemental processing and learning of visual information.

Methods
When free-flying bees are trained to a saliently coloured stimulus with a predominant orientation in an absolute conditioning procedure, they generalize their choice to different orientations of the same stimulus (Giurfa et al. 1995a). Thus, a priori, bees trained with a yellow vertical grating should generalize their choice to the same yellow grating placed horizontally. However, our conditioning procedure generates an ambiguity at the level of the associative strength of each separate element, such that bees should not generalize from one situation to the other if they want to solve the BiDi problem.

stimuli from one subcompartment to the opposite one from one test to the next. Between tests, a refreshment training of four visits (one with each training situation: Yh versus Vh, Yh versus Yv, Vv versus Vh and Vv versus Yv) was done to keep the bees foraging motivation high. We report results only from bees that completed at least three of the four tests.

Results
We trained seven bees with the problem Yv+, Vh+, Yh , Vv (AB+, CD+, AD , BC ) and eight bees with the problem Yv , Vh , Yh+, Vv+ (AD , BC , AB+, CD+). We compared the performance of these two groups in the four tests AB+ versus AD , CD+ versus AD , AB+ versus CB and CD+ versus CB (ANOVA for repeated measurements). Since no significant difference was found in their performance (F1,12 =3.78, NS), we pooled the data of the two groups. Figure 4 shows the test performance of the bees trained in the BiDi problem. Bees preferred the correct stimulus (AB+ and CD+) in all four tests (AB+ versus AD : 70.90 versus 29.10%; G14 =195.72, P< 0.0001; CD+ versus AD : 62.23 versus 37.77%; G15 =153.51, P< 0.0001; AB+ versus CB : 65.65 versus 34.35%; G15 =147.48, P< 0.0001; CD+ versus CB : 59.92 versus 40.08%; G15 =91.86, P< 0.0001). Thus, bees chose significantly more often the compounds that were followed by reinforcement than those that were not followed by reinforcement. These results show that honeybees were capable of solving a BiDi in which colour and orientation were used as elemental stimuli. GENERAL DISCUSSION Our study shows that bees can solve discrimination problems that involve ambiguity at the level of elemental associative strength and that require, therefore, nonelemental processing of stimulus compounds. In the appetitive context of food search, free-flying honeybees could master a positive and a negative patterning discrimination and a biconditional discrimination. After being trained in a modified PP schedule (A versus XY+, B versus XY+, AB+), they significantly preferred the compound to the elements, whilst they preferred the elements to the compound after being trained in a modified NP schedule (A+, B+, AB versus XY+). They also preferred the reinforced compounds to those that were nonreinforced in the BiDi task (AB+, CD+, AC , BD ) despite ambiguity at the elemental level. The solving of the PP task could also be explained in elemental terms: the associative strengths of the yellow and violet chequerboards could both be subthreshold for the response but once added together to make the compound presentation they could result in a suprathreshold associative strength. Such a summation would yield higher associative strength and, therefore, higher responsiveness to the yellowviolet chequerboard. Therefore results of PP are not necessarily conclusive in analyses of nonelemental processing because it can be considered as a test of elemental learning.

Stimuli
The gratings (Fig. 4) were either vertical (v) or horizontal (h) (0 and 90 with respect to the vertical, respectively). They were defined on the basis of stripes, 2 cm wide, which covered the whole circular area of the stimulus (10 cm in diameter). The stripe period was 2 cm. Such a grating was perfectly resolvable for a bee flying within the decision chamber (Srinivasan & Lehrer 1988). Colours were cut from the same yellow (Y) and violet (V) papers used in the previous experiments (HKS-1N and HKS-36N papers, respectively). They were presented on the same grey background of the previous experiments (HKS-92 N paper). Owing to the resolvable nature of the stripes, no perceptual colour mixture occurred. When the bee had to decide between two test options, the coloured stripes appeared oriented and chromatic to her. Grey was not specifically associated with the presence or absence of reinforcement.

Procedure
During training, two different stimuli were always presented, each in the centre of the back wall of one subcompartment of the decision chamber. If both gratings had the same colour, their orientation was different (Yh versus Yv or Vh versus Vv) and if their orientation was the same, their colour was different (Yh versus Vh or Yv versus Vv). During training, all stimuli combinations were presented 20 times in a pseudorandom sequence such that no stimulus was presented consecutively more than twice in the same subcompartment of the maze. Thus a complete training schedule comprised 80 training trials per bee. We trained two groups of bees in parallel: one was trained with the problem Yv+, Vh+, Yh , Vv (that is, AB+, CD+, AD , BC ) and the other with the reversed problem Yv , Vh , Yh+, Vv+ (that is AD , BC , AB+, CD+). The whole experiment was balanced. The possible stimuli combinations during training were Yh versus Vh, Yh versus Yv, Vv versus Vh and Vv versus Yv. In both experimental groups, bees were tested with the same four stimulus combinations used during training but without reward. We did each test twice, changing the

182

ANIMAL BEHAVIOUR, 64, 2

The solving of the NP task, however, cannot be explained by the elemental summation principle. This principle predicts that the excitatory associations of the elements (A+, B+) will lead to summation when the AB compound is presented. For a pure elemental account, the associative strength of the compound chequerboard should be twice as high as those of the elemental chequerboards, that is, bees should prefer the yellow violet chequerboard to the yellow or the violet chequerboard, which we never observed. The results of the NP experiment could be explained either by the unique-cue or the configural hypothesis. In the former case (uniquecue), the discrimination that had to be learned would be A+, B+, ABC . This means that in AB trials the unique cue C will acquire inhibitory associative strength that will counter the excitatory associative strengths of the elements A and B. In the latter case (configural), the yellowviolet chequerboard would constitute a configuration that would be distinct from the yellow and the violet elemental chequerboards. Further experiments would be necessary to decide between these two options. The same argument applies to the BiDi experiment. Pure elemental associations are not sufficient to explain its results because during training each element was followed as often by reinforcement as by nonreinforcement. Therefore, simple summation of the elemental associative strengths on compound presentation should result in the same level of responding for all four compounds, which was not the case. The observed differences in the response strengths between the reinforced and the nonreinforced compounds can be explained only if one assumes nonelemental processing of the different compounds trained, a strategy that would eliminate the ambiguity of the discrimination problem (Sutherland & Rudy 1989; Pearce 1994; Rudy & Sutherland 1995). As in the case of NP, we cannot decide between the configural and the unique-cue interpretation in accounting for these results. In both NP and BiDi, however, we can clearly reject the purely elemental account. In the patterning experiments, discrimination between the elementary colours yellow and violet and their compound could be easily explained without the necessity for a configural associative theory if adjacent yellow and violet squares in the chequerboard were simply to mix perceptually to form new green squares or stripes. In that way, the yellowviolet chequerboard would not be a compound in which the elementary colours could be identified, but a new, homogeneous green disc, perceptually different from the original elements. If this were the case, the outcome of the experiment would be trivial and no generalization between elementary trained colours and the compound should be expected. However, unit squares in our stimuli were spatially resolvable and thus excluded the possibility of a perceptual mixture between yellow and violet. The chequerboard was indeed perceived as a compound of yellow and violet squares. In addition, the unitary yellow or violet squares were not mixed with the grey background and could be clearly distinguished. Similarly, in the biconditional discrimination experiment, perceptual mixture within gratings was excluded because of their spatial period. If such a

mixture had occurred, bees would never discriminate two gratings with the same colour and different orientations, which they did in all cases. One may argue that in both patterning experiments, bees could solve the task by focusing on the grey squares that were present only in the elementary stimuli (A or B). This argument can, however, be dismissed, as the whole back wall on which the stimuli were presented was the same grey as the stimuli squares and because there was continuity between grey squares and background (see Fig. 3). In fact, all stimuli, elementary and compound ones, were presented on the grey background. Thus, grey could not be useful in this experimental paradigm. The clearest evidence that bees did not use the grey squares as information for their decisions comes from the NP experiment, in which a significant difference between the two elemental stimuli, yellow and violet, was found: the bees preferred violet to yellow squares when presented against the compound. This asymmetry can be explained only if the bees exclusively used the chromatic information, yellow or violet, and ignored the achromatic information provided by the grey squares, which was common to both elementary stimuli. The asymmetry between yellow and violet also suggests that these two colour stimuli did not have the same salience for the bees. Such an effect may be seasonal, as it was not found in the PP experiment done in the previous year. It could result from an innate bias towards violet as found in colour preference experiments with nave bees (Giurfa et al. 1995b) or from the cumulative experience of the bees in the field before they were trained to our set-up. The latter effect was shown to influence olfactory conditioning in honeybees (Gerber et al. 1996). The patterning experiments could, however, be explained in terms of a unique-cue based on the amount of chromatic contrast within each pattern. Chromatic contrast is the perceptual distance between a colour and the background on which it is presented. In the case of the colours we used, the amount of chromatic contrast between yellow and grey and violet and grey was the same. Thus, such a contrast was identical in both the elemental yellow and the elemental violet chequerboard (three COC units in the colour space of the honeybee; see Fig. 2b and Backhaus 1991). The contrast between yellow and violet was much higher (seven COC units). Thus, bees could discriminate the compound yellowviolet chequerboard from the elemental yellow and violet chequerboards on the basis of the amount of chromatic contrast within the pattern. This explanation deserves further study, as the possibility that bees can switch between using different properties of chromatic patterns in visual learning remains an open question. Relying on chromatic contrast as a unique cue does not explain, however, the results of the BiDi experiment as all gratings involved chromatic (yellow or violet) and achromatic grey stripes. As the chromatic contrast between yellow and grey stripes and between violet and grey stripes was the same (three COC units), this cue could not be useful to solve this kind of discrimination. Bees perceive orientation through the achromatic channel of the long-wave receptor type (green receptor)

SCHUBERT ET AL.: VISUAL LEARNING IN BEES

183

alone (Giger & Srinivasan 1996), using as a cue the so-called green contrast (that is the contrast of a target area against its background for the green receptor type alone). If bees reduced the whole BiDi problem to an orientation discrimination problem, they would not be able to solve it: for instance, the tests of the problem Yv+, Vh+, Yh , Vv would be v+ versus h , h+ versus h , v+ versus v and h+ versus v . Thus each orientation appears equally often as rewarded and nonrewarded. Problem solving requires a nonelementary solution. This, however, does not invalidate the unique-cue interpretation in the case of this experiment. We cannot exclude interactions between colour and orientation at a higher level that could give rise to a unique-cue sensation that is different for each grating. Our experiments precluded the formation of a retinotopically fixed template as the spatial position of each element was varied constantly in the visual field of the bee. In addition, when presenting a compound (e.g. a chequerboard with two colours), we constantly rotated it on its central point to present the coloured areas at different retinal positions. In addition, chequerboards and gratings were constantly replaced by similar stimuli with a reversed arrangement of colours. This was important because we wanted bees to learn the elemental information within a stimulus, independently of its spatial coordinates. Although evidence in favour of learning patterns as templates is being revised in the light of different pattern recognition theories in insects (Ernst & Heisenberg 1999; Efler & Ronacher 2000; Stach & Giurfa 2001), some authors maintain that insects memorize patterns in a pixel-based representation that preserves all resolvable spatial details (Wehner 1972; Gould 1985, 1986; Dill et al. 1993; Giurfa et al. 1995a). With such an image in their memory, the insects choice is thought to be determined by the degree of overlap between the memorized image and the observed pattern. In this case, generalization to novel patterns or to single elements forming the patterns is poor. Drosophila flies, for instance, remember the position of stimuli in the visual field and are unable to recognize that two identical patterns displaced by 9 in space are the same (Dill et al. 1993; see Heisenberg et al. 2001 for a review). Our training procedure precluded the building of such a template and therefore made sure that bees took into account the elemental information pertaining to our stimuli. In the appetitive olfactory conditioning of the proboscis extension reflex bees solved PP and NP (Deisig et al. 2001) and BiDi problems (Hellstern et al. 1995; Chandra & Smith 1998). In comparing the difficulty of PP and NP, Deisig et al. (2001) showed that bees solved PP more easily than NP. This difference is consistent with findings in humans in which the skin conductance response was conditioned (Lachnit & Kimmel 1993, 2000). This asymmetry in the difficulty of the task, however, seems to depend on the particular ratio of the number of reinforced and nonreinforced trials for both kinds of patterning discriminations. In our experiments the response differentiation found in NP seemed to be larger than that in PP. However, the two patterning experiments were not done in parallel. Therefore, since

factors such as weather, food source availability, colony conditions and the bees hormonal status were not necessarily constant from one experimental season to the next, it is impossible to compare directly our PP and NP studies. Under certain circumstances, bees trained to a coloured pattern can generalize their choice to elements of the pattern. This indicates that elemental associations between single colour elements and the reinforcement are possible. The kind of training experienced by animals may be critical in implementing either elemental or configural associations. This was shown in olfactory compound learning by spiny lobsters, Panulirus argus, in which absolute or differential conditioning of an olfactory compound resulted in configural or elementary compound processing, respectively (Livermore et al. 1997). In our experiments bees clearly solved the discrimination tasks. We thus conclude that bees built associations other than those between each element and the US. Acknowledgments We thank Randolf Menzel for encouragement, discussions and helpful comments on the manuscript, and Mary Wurm for correcting the English. We also thank Natalie Hempel, Bernhard Komischke, Silvia Lee, Dagmar Malun, Nina Deisig and Silke Stach for valuable discussions. This work was supported by grants DFG Gi 291/3-1, DFG La 564/10-1 and DFG Me 365/23-1 of the German Research Council (Deutsche Forschungsgemeinschaft: DFG) and by a grant ASUPS of the Universite PaulSabatier to M.G. References
Backhaus, W. 1991. Colour opponent coding in the visual system of the honeybee. Vision Research, 31, 13811397. Bitterman, M. E., Menzel, R., Fietz, A. & Scha fer, S. 1983. Classical conditioning of proboscis extension in honeybees (Apis mellifera). Journal of Comparative Psychology, 97, 107119. Chandra, S. & Smith, B. H. 1998. An analysis of synthetic processing of odour mixtures in the honeybee (Apis mellifera). Journal of Comparative Psychology, 201, 31133121. Couvillon, P. A. & Bitterman, M. E. 1988. Compound-component and conditional discrimination of colours and odours by honeybees: further tests of a continuity model. Animal Learning and Behavior, 16, 6774. Deisig, N., Lachnit, H. & Giurfa, M. 2001. Configural olfactory learning in honeybees: negative and positive patterning discriminations. Learning and Memory, 8, 7078. Dill, M., Wolf, R. & Heisenberg, M. 1993. Visual pattern recognition in Drosophila involves retinotopic matching. Nature, 365, 751753. Efler, D. & Ronacher, B. 2000. Evidence against a retinotopictemplate matching in honeybees pattern recognition. Vision Research, 40, 33913403. Ernst, R. & Heisenberg, M. 1999. The memory template in Drosophila pattern vision at the flight simulator. Vision Research, 39, 39203933. Gerber, B., Geberzahn, N., Hellstern, F., Klein, J., Kowalsky, O., Wu stenberg, D. & Menzel, R. 1996. Honey bees transfer

184

ANIMAL BEHAVIOUR, 64, 2

olfactory memories established during flower visits to a proboscis extension paradigm in the laboratory. Animal Behaviour, 52, 10791085. Giger, A. & Srinivasan, M. V. 1996. Pattern recognition in honeybees: chromatic properties of orientation analysis. Journal of Comparative Physiology A, 178, 763769. Giurfa, M., Backhaus, W. & Menzel, R. 1995a. Colour and angular orientation in the discrimination of bilateral symmetric patterns in the honeybee. Naturwissenschaften, 82, 198201. Giurfa, M., Nu n ez, J., Chittka, L. & Menzel, R. 1995b. Colour preferences of flower-nai ve honeybees. Journal of Comparative Physiology A, 177, 247259. Giurfa, M., Vorobyev, M., Brandt, R., Posner, B. & Menzel, R. 1997. Discrimination of coloured stimuli by honeybees: alternative use of achromatic and chromatic signals. Journal of Comparative Physiology A, 180, 235244. Giurfa, M., Hammer, M., Stach, S., Stollhoff, N., Mu ller-Deisig, N. & Mizyrycki, C. 1999. Pattern learning by honeybees: conditioning procedure and recognition strategy. Animal Behaviour, 57, 315324. Gould, J. L. 1985. How bees remember flower shapes. Science, 227, 14921494. Gould, J. L. 1986. Pattern learning by honeybees. Animal Behaviour, 34, 990997. van Hateren, J. H., Srinivasan, M. V. & Wait, P. B. 1990. Pattern recognition in bees: orientation discrimination. Journal of Comparative Physiology A, 167, 649654. Heisenberg, M., Wolf, R. & Brembs, B. 2001. Flexibility in a single behavior variable of Drosophila. Learning and Memory, 8, 110. Hellstern, F., Wu stenberg, D. & Hammer, M. 1995. Contextual learning in honeybees under laboratory conditions. In: Learning and Memory (Ed. by N. Elsner & R. Menzel), page 30. Stuttgart: Thieme. Lachnit, H. & Kimmel, H. D. 1993. Positive and negative patterning in human classical skin conductance response conditioning. Animal Learning and Behavior, 21, 314326. Lachnit, H. & Kimmel, H. D. 2000. Experimental manipulation of a unique cue in Pavlovian SCR conditioning with humans. Biological Psychology, 50, 157171. Livermore, A., Hutson, M., Ngo, V., Hadjisimos, R. & Derby, C. D. 1997. Elemental and configural learning and the perception of odorant mixtures by the spiny lobster (Panulirus argus). Physiology and Behavior, 62, 169174. Menzel, R. 1985. Learning in honey bees in an ecological and behavioural context. In: Experimental Behavioural Ecology (Ed. by B. Ho lldobler & M. Lindauer), pp. 5574. Stuttgart: Gustav Fischer Verlag. Menzel, R. 1999. Memory dynamics in the honeybee. Journal of Comparative Physiology A, 185, 323340. Menzel, R. & Backhaus, W. 1991. Colour vision in insects. In: Vision and Visual Dysfunction. The Perception of Colour (Ed. by P. Gouras), pp. 262288. London: MacMillan. Menzel, R. & Giurfa, M. 2001. Cognitive architecture of a minibrain: the honeybee. Trends in Cognitive Sciences, 5, 6271. Menzel, R. & Mu ller, U. 1996. Learning and memory in honeybees: from behaviour to neural substrates. Annual Review of Neuroscience, 19, 379404. Menzel, R., Greggers, U. & Hammer, M. 1993. Functional organization of appetitive learning and memory in a generalist pollinator, the honey bee. In: Insect Learning: Ecological and

Evolutionary Perspectives (Ed. by D. Papaj & A. C. Lewis), pp. 79125. New York: Chapman & Hall. Mu ller, D., Gerber, B., Hellstern, F., Hammer, M. & Menzel, R. 2000. Sensory preconditioning in honeybees. Journal of Experimental Biology, 203, 13511364. Nu n ez, J. A. 1970. The relationship between sugar flow and foraging and recruiting behaviour of honey bees (Apis mellifera L.). Animal Behaviour, 18, 527538. Nu n ez, J. A. 1982. Honeybee foraging strategies at a food source in relation to its distance from the hive and the rate of sugar flow. Journal of Apicultural Research, 21, 139150. Pavlov, I. P. 1927. Lectures on Conditioned Reflexes. New York: International Publishers. Pearce, J. M. 1987. A model for stimulus generalisation in pavlovian conditioning. Psychological Review, 94, 6173. Pearce, J. M. 1994. Similarity and discrimination: a selective review and a connectionist model. Psychological Review, 101, 587607. Rescorla, R. A. 1972. Configured conditioning in discrete-trial bar pressing. Journal of Comparative and Physiological Psychology, 79, 307317. Rescorla, R. A. 1973. Evidence for unique stimulus account of configural conditioning. Journal of Comparative and Physiological Psychology, 85, 331338. Rescorla, R. A. & Wagner, A. R. 1972. A theory of Pavlovian conditioning: variations in the effectiveness of reinforcement and non-reinforcement. In: Classical Conditioning II: Current Research and Theory (Ed. by A. H. Black & W. F. Prokasy), pp. 6499. New York: Appleton-Century-Crofts. Rudy, J. W. & Sutherland, R. J. 1992. Configural and elemental associations and the memory coherence problem. Journal of Cognitive Neuroscience, 4, 208216. Rudy, J. W. & Sutherland, R. J. 1995. Configural association theory and the hippocampal formation: an appraisal and reconfiguration. Hippocampus, 5, 375389. Saavedra, M. A. 1975. Pavlovian conditioning in the rabbit. Learning and Motivation, 6, 314326. Sokal, R. R. & Rohlf, F. J. 1995. Biometry. New York: W. H. Freeman. Srinivasan, M. V. & Lehrer, M. 1988. Spatial acuity of honeybee vision and its spectral properties. Journal of Comparative Physiology A, 162, 159172. Stach, S. & Giurfa, M. 2001. How honeybees generalize visual patterns to their mirror image and leftright transformation. Animal Behaviour, 62, 981991. Sutherland, R. J. & Rudy, J. W. 1989. Configural association theory: the role of the hippocampal formation in learning, memory, and amnesia. Psychobiology, 17, 129144. Takeda, K. 1961. Classical conditioned response in the honeybee. Journal of Insect Physiology, 6, 168179. Vorobyev, M. & Brandt, R. 1997. How do insect pollinators discriminate colours. Israel Journal of Plant Sciences, 45, 103113. Wehner, R. 1972. Dorsoventral asymmetry in the visual field of the bee, Apis mellifica. Journal of Comparative Physiology, 77, 256277. Whitlow, J. W. & Wagner, A. R. 1972. Negative patterning in classical conditioning: summation of response tendencies to isolable and configural components. Psychonomic Science, 27, 299301. Williams, D. A. & Braker, D. S. 1999. Influence of past experience on the coding of compound stimuli. Journal of Experimental Psychology: Animal Behavior Processes, 25, 461474.