Grouping of visual objects by honeybees

Transcription

1 The Jornal of Experimental Biology 27, Pblished by The Company of Biologists 24 doi:.242/jeb Groping of isal objects by honeybees Shaow Zhang, *, Mandyam V. Sriniasan, Hong Zh and Jason Wong 2 Centre for Visal Sciences, Research School of Biological Sciences, Astralian National Uniersity, Canberra, ACT 26, Astralia and 2 School of Moleclar and Microbial Sciences, Uniersity of Sydney, Sydney, Astralia *Athor for correspondence ( swzhang@rsbs.an.ed.a) Accepted 8 Jne 24 Recent work has reealed that monkeys as well as pigeons are able to categorise complex isal objects. We show here that the ability to grop similar, natral, isal images together extends to an inertebrate the honeybee. Bees can be trained to distingish between different types of natrally occrring scenes in a rather general way, and to grop them into for distinct categories: landscapes, plant stems and two different kinds of flowers. They exhibit the same response to noel isal objects that differ greatly in their indiidal, low-leel featres, bt Smmary belong to one of the for categories. We exclde the possibility that they might be sing single, low-leel featres as a ce to categorise these natral isal images and sggest that the categorisation is based on a combination of low-leel featres and configrational ces. Key words: honeybee, learning, memory, groping, categorization, matching-to-sample, ce. Introdction Recent research is reealing that honeybees possess sophisticated sensory and perceptal mechanisms that enable them to naigate to food sorces accrately and efficiently. They are able to se a series of isal images of the enironment acqired en rote to get to their destination (Collett, 996; Collett et al., 993; Jdd and Collett, 998; Wehner et al., 99, 996). Bees are also able to associatiely grop and recall isal stimli along their foraging paths, enabling them to organize and retriee naigational information pertaining to mltiple rotes (Zhang et al., 999). They can learn to se symbolic rles for naigating throgh complex mazes and to apply these rles in flexible ways (Zhang et al., 996, 999, 2). Honeybees are able to form concepts of sameness and difference. They can learn to sole matching-to-sample and non-matching-to-sample discriminations, and transfer the learned rles to noel stimli of the same or a different sensory modality (Girfa et al., 2). Bees can also extract general properties of a stimls, sch as orientation (Wehner, 97; an Hateren et al., 99) or symmetry (Horridge and Zhang, 995; Girfa et al., 996; Horridge, 996), and apply them to distingish between other stimli, which they hae neer preiosly encontered. This demonstrates that honeybees are able to categorise artificial geometrical objects by sing abstract featres. Althogh the bee has only a tiny brain, many of the behaiors referred to aboe reqire considerable perceptal capabilities and central storage and ealation mechanisms associated with experience-dependent adaptations (Sriniasan and Zhang, 998; Menzel and Girfa, 2). The natral enironment in which a bee operates is composed of a ariety of landscapes and a ariety of objects within them, sch as trees, plants and flowers. How do honeybees look at objects and scenes? Can bees grop the different types of objects that they enconter into different categories? Sch a capacity wold facilitate rapid and accrate recognition of important landmarks and targets, and enhance foraging efficiency. Object groping can be thoght of as the ability to link together items that are similar, een thogh they are distingishable from one another. A rose is a rose, regardless of its exact size, color or orientation; it can ths be considered to belong to a grop that is different from the grop of rocks, for example, or trees, or ponds. Recent work has reealed that monkeys and other primates are able to categorise complex isal images, sch as photographs of hman faces, trees and other animals (Daenport and Rogers, 97; Vogels, 999; Freedman et al., 2). Martin-Maliel sggested that the baboon can form amodal abstract concepts of hman and baboon categories (Martin-Maliel and Fagot, 2). Pigeons hae the capacity to grop objects into a nmber of different categories, sch as people, other pigeons, trees, water, landscapes and so on (Mallott and Siddall, 972; Herrnstein, 984; Roitblat, 987; Hber et al., 2). They can een learn to distingish between otline drawings of the leaes of different tree species (Cerella, 979). Bmblebees can learn to associate color with a reward, irrespectie of other isal parameters sch as size or shape (Dkas and Waser, 994). So far, howeer, there hae been no

2 329 S. Zhang and others stdies inestigating the ability of inertebrates to classify complex, natral objects. A Lateral iew Here we explore whether bees can learn to distingish between for different categories of natral isal images that are likely to be releant to their foraging behaior. The categories examined are flowers of two different shapes, plant stems and landscapes. The items belonging to the different categories possess distinct characteristics, and are nambigosly perceied, Entrance C C2 C3 at least by hmans, as belonging to different perceptal classes. On the other hand, items belonging to 8 cm a gien category are perceied, at least by hmans, as being similar, and belonging to the same class. C Patterns A R, 2 R 3, 4 A B Front iew: A A Materials and methods General The experiments were carried ot in the All Weather Bee Flight Facility at the Astralian National Uniersity s Research School of Biological Sciences. The facility consists of a modified glasshose in which the internal temperatre is reglated by a compter to maintain 24±5 C dring the day and 7±3 C at night. A beehie with two entrances, one allowing bees access to the inside of the facility and the other to the otside, was monted on the wall of the facility. Bees foraging indoors obtained sgar water from feeders in the facility. For each experiment, abot 2 bees were marked indiidally at the beginning of the experiment and trained to isit a feeder containing.5 mol l sgar soltion in the maze. The experiment commenced with pre-training for 2 days (training bees to enter the maze and choose among test stimli), followed by training on the actal task. The latter phase inclded training periods, learning tests and transfer tests (Type and Type 2, described below). The entire experiment ran for days. 8 bees remained at the end of the experiment. Grop Grop 2 Grop 3 Grop 4 F f P L F2 f2 P2 L2 F3 f3 P3 L3 F4 f4 Fig.. Illstration of experiments showing (A,B) the mltiple-choice maze apparats and (C) the for grops of stimli that were sed. See text for details. L3 L4 Experimental setp A mltiple-choice maze, located inside the facility, was sed for training and tests. Bees entering the maze encontered a sample stimls at the entrance chamber (C), a connecting chamber (C2) and for additional test stimli in a sbseqent test chamber (C3; Fig. A,B). The bees were trained to fly throgh chambers C, C2 and C3 in sccession. The back wall of the entrance chamber C carried the sample stimls. The bees flew throgh a 3 cm hole in the sample stimls to chamber C2, the back wall of which consisted of a transparent film with a 3 cm diameter apertre in the centre. The transmission of the film is approximately niform in the hman isible spectrm and is redced in the UV. The latter, howeer, is irreleant, as there is relatiely little UV light within the bee flight facility, becase the roof blocks most of it. This apertre restricted the bees speed of flight throgh the apparats, and the transparent film proided the bees with a iew from C2 of the for test stimli, that were monted on a choice board, forming the back wall of the test chamber C3. If the bee chose the correct test stimls in C3, she wold be able to receie a reward of sgar soltion from a feeder that was placed in the reward box, R, behind that stimls, by landing on and crawling throgh a tbe in the centre of the stimls. Visal stimli For grops of complex images (G, G2, G3 and G4), printed on disks of diameter 8 cm sing a color laser printer (Tektronix Phaser 78 Graphics, NWS Corp., NY, USA), were sed in the training and the tests. Each grop consisted of for stimli, each stimls belonging to a different category (Fig. C). The categories were as follows. One category (F) consisted of images of flowers that were star-shaped, bt of different colors. A second category (f) comprised images of

3 Groping of isal objects by honeybees 329 flowers that were nearly circlar in shape, again of different colors. The third category (P) consisted of images of plant stems, of arios shapes. The final category (L) was composed of images of landscapes. Within each category, indiidal images differed in the details of their shape, textre and, sometimes, color. In transfer tests, the sample stimls was always from Grop and the for test stimli on the choice board were from one of the three other grops (G2, G3 or G4). Each test stimls, again, was from a different category, as shown in Fig. C. Training and testing procedres Dring training and learning tests, the sample stimls and the for test stimli on the choice board were from Grop. Each of the for test stimli was from a different category, as shown in Fig. C. Dring training, the sample pattern was changed eery min (after an aerage of two rewarded isits per bee). The relatie positions of the for test stimli were also randomly shffled eery min. This ensred that the bees learned to match the stimli by isal comparison, and not by associating a specific feeder location with each sample. Learning tests began on the third day of training. There were two types of transfer tests. In Type transfer tests, the sample stimls and the for test stimli on the choice board were all from the same grop, bt this grop was different from that sed dring training. In other words, these transfer tests were condcted sing Grop 2 (in some tests) or Grop 3 (in others). In Type 2 transfer tests, the sample stimls was from Grop and the for test stimli on the choice board were from a different grop (Grop 2, Grop 3 or Grop 4). The two types of transfer tests were interleaed. The bees performances were measred in learning tests as well as transfer tests. In each case, performance was ealated by noting which test stimls the bee chose first pon entering the test chamber (by landing on the corresponding entrance tbe). Each transfer test was carried ot only for a brief period ( min, inoling abot two isits per bee). The reward contined to be present dring the tests, in order to minimize extinction and maintain the bees motiation to isit the apparats. The breity of each transfer test ensred that no learning occrred dring the test. Transfer tests were interleaed between segments of contined training that were at least 4 min long, sing Grop stimli. Each transfer test was repeated 4 5 times to gather sfficient data. Controls to check for the se of olfactory ces Controls were carried ot to check whether the bees choice behaior in the learning and transfer tests was inflenced by possible olfactory ces emanating from the feeder, which was placed in a box behind the appropriate test stimls. Two types of controls were sed. In order to minimise the effects of extinction of learning, one type of control was carried ot briefly at the end of the day, whereby bees were continosly tested in the transfer test, which had been carried ot jst before the control check (bt with the feeder remoed). Another type of control was carried ot at the end of the whole series of experiments. In these controls, all for test stimli, as well as the sample, were identical. The stimls sed in these tests was a grey-leel ersion of F (Fig. 5). The feeder was placed behind one of the test stimli (this position was aried randomly from one control test to the next). For control tests of this type were performed, each with the feeder in a different position. Data analysis Each type of transfer test was performed 4 5 times to gather sfficient data. For each test, the choice freqency for each stimls was calclated from the total nmber of first choices for each of the for test stimli. The performance of each bee was ealated separately by pooling all of the choices that it made for a gien test stimls in a gien type of test. The aerage performance for a gien type of test was obtained by aeraging the choice freqencies across bees. The sample size (N) was taken to be the nmber of bees, rather than the nmber of indiidal choices, to ensre that the samples were trly statistically independent. The data was analysed to obtain mean ales of choice freqency, standard deiations (S.D.) and standard errors of means (S.E.M.). Stdent t-tests were sed to determine whether each choice freqency was significantly different from the random choice leel of 25%. The Kolmogoro Smirno two-sample test (Groebner and Shannon, 985) was sed to test for significant differences between histograms obtained in a gien type of test with different samples. For each gien type of experimental test, six sch statistical tests were done (details are gien in the figre legends). Reslts Learning tests Bees were initially trained sing the for stimli labelled as Grop in Fig. C. Each test stimls belonged to a different category. The sample stimls was identical to one of the test stimli, and the bees had to learn to choose the test stimls that matched the sample. This task was, in effect, a matchingto-sample task. The learning tests were commenced on the third day of training (after ca. 8 rewarded isits per bee, on aerage). The bees learned this task well. For each sample, the bees showed a strong and statistically significant preference for the matching test stimls (Fig. 2A). Performance for the learning tests was measred oer a total of 32 isits of 9 bees. The choice freqency in faor of the matching stimls was significantly greater than the random choice leel of 25% (P<.), for each of the for sample stimli. The aerage choice freqency in faor of the matching stimls was 6%. Howeer, the choice freqencies for the other (non-matching) stimli were always either significantly lower than 25%, or not significantly different from it (details in Fig. 2A). Differences between the histograms for different sample stimli were tested

4 3292 S. Zhang and others sing the Kolmogoro Smirno two-sample test. All the histograms are fond to be significantly different from each other (P<.). A Choice freqency.8.6 Sample G B.8.6 Sample G2 C Choice freqency Choice freqency.8.6 Sample G3 Learning tests on G N=9 N=9 N=9 N=9 Type transfer tests on G2 Type transfer tests on G3 N=4 N=4 N=4 N=4 N=2 N=2 N=2 N=2 * * Type transfer tests Type transfer tests are described in Materials and methods. The trained bees were then briefly tested on two noel grops of stimli (Grops 2 and 3), which they had neer preiosly encontered. Each grop consisted of for stimli, one belonging to each category (Fig. C). In each test, the sample was identical to one of the stimli of the grop that was being tested. The bees were immediately able to find the matching stimls in each of the noel grops, withot any training on them. Performance for Type transfer tests sing Grop 2 was measred oer a total 797 isits of 4 bees, and for Grop 3 oer a total 383 isits of 2 bees. The bees showed a strong and statistically significant preference for the matching test stimls: the choice freqency in faor of the matching stimls was significantly greater than the random choice leel of 25% for each of the for samples. The aerage choice freqency for the matching stimls was 62% for Grop 2 and 6% for Grop 3. The differences between the histograms for different sample stimli in Grop 2 as well as Grop 3 were significant (P<.). The reslts of these transfer tests (Fig. 2B,C) show that the bees performed well with each grop. Ths, the bees were able to apply to Grop 2 and Grop 3 the concept of matching that they had acqired whilst being trained on Grop. These reslts confirm and extend earlier work in or and other laboratories, which demonstrated that bees can learn to match colors, stripes or scents and transfer this matching ability to noel stimli (Girfa et al., 2). Type 2 transfer tests Type 2 transfer tests are described in Materials and methods. The experiments described so far demonstrate the ability of bees to learn to match stimli that are identical. Can bees go one step frther, and identify similar stimli as belonging to the same category? To inestigate this, we asked Fig. 2. Reslts of (A) learning tests and (B,C) Type transfer tests. For each grop, the bars show the relatie preferences for the for test stimli when the sample was a star-shaped flower, a circlar flower, a plant stem or a landscape, as shown nderneath the abscissa. In each panel, N denotes the nmber of bees that were tested in each experiment. Asterisks denote statistically significant differences from the random choice leel of 25% (broken horizontal lines). P<.; P<.; *P<.5. Black asterisks refer to leels that are significantly greater than 25%, and red asterisks to leels significantly lower than 25%. The circles denote no significant difference from 25%. Vales are means ± S.E.M. of the data. In each case, the bees are able to learn to choose the test stimls that matches the sample.

5 Groping of isal objects by honeybees 3293 whether the bees, trained as aboe, cold match a sample stimls from one grop, with a stimls of the same category from a different grop. In these tests, the sample was always a stimls from Grop, bt the test stimli were from Grop 2, Grop 3 or Grop 4 (see Fig. C). Performance was measred oer a total of 262 isits of bees for Type 2 transfer tests from Grop to Grop 2, oer a total of 256 isits of bees for the transfer tests from Grop to Grop 3, and oer a total of 28 isits of bees for the transfer tests from Grop to Grop 4. The transfer tests were brief, and were interleaed between segments of the training session. The bees performed ery well in these transfer tests. In each case, the bees showed a clear and significant preference for the test stimls that belonged to the same category as the sample (Fig. 3A,B). Particlarly noteworthy is the transfer test sing Grop 4, in which the test stimli were entirely noel (Grop 4, Fig. C). These stimli had neer been sed in the training phase, or in the learning tests or Type transfer tests. Again, the bees performed ery well at picking the test stimls that was in the same category as the sample (Fig. 3C). We tested for differences between the histograms obtained sing the for different sample stimli for each of the three types of transfer tests, namely, Grop to Grop 2, Grop to Grop 3 and Grop to Grop 4. In each type of transfer test, the for histograms were significantly different from each other (P<.). In a final set of transfer tests, we examined the trained bees ability to match colored stimli with grey-leel ersions of them. Here the sample stimli were from Grop 3, and the test stimli were greyleel ersions of these stimli (Grop 3*, Fig. 4). These transfer tests represent data from a total of 29 isits of bees. The bees performed well at this task, too (Fig. 4). Control tests to check for the se of olfactory ces Fig. 5 shows the reslts of control tests that were carried ot at the end of all transfer tests. Here, all for test stimli and the sample were identical, and the feeder was placed behind one of the test stimli. The data were collected oer a total of 8 isits by 5 bees. The bees chose randomly among all for test stimli, exhibiting no preference for the stimls that was associated with the feeder (P>.5). Ths, the bees choices in or experiments were drien A Choice freqency B Choice freqency C Choice freqency Sample Transfer tests: G to G2 Transfer tests: G to G3 Transfer tests: G to G4 F f P L Fig. 3. Reslts of Type 2 transfer tests examining the G ability of bees, haing encontered a sample stimls from Grop, to choose a test stimls of the same category in (A) Grop 2, (B) Grop 3 and (C) a noel Grop 4. In each case, the bees are able to learn to choose the test stimls that belonged to the same category as the sample. Details as in Fig. 2. N= N= N= N= * * * * * * N= N= N= N= N= N= N= N= * *

6 3294 S. Zhang and others Choice freqency.8.6 Sample F3* f3* P3* L3* F3* f3* P3* L3* F3* f3* P3* L3* F3* f3* P3* L3* N= N= N= N= * F3 f3 P3 L3 Grop 3 Fig. 4. Transfer tests examining the ability of bees, haing encontered a sample stimls from Grop 3, to choose a test stimls of the same category in Grop 3*, which is composed of grey-leel ersions of the stimli in Grop 3. Bees are clearly able to perform this task. Details as in Fig. 2. Choice freqency.5.3. Sample N=5 N=5 N=5 N=5 F (Grey) Fig. 5. Reslts of for control tests to check for the possible se of olfactory ces. Bees trained in the category discrimination experiments of Fig. 2 were presented with tests in which the sample as well as the for test stimli were identical (in this case, they were grey-leel ersions of F). The feeder was placed behind one of the test stimli, as shown by the circlar symbol. The bees chose randomly among the for stimli, indicating that they were not sing olfactory ces to make their discriminations in or experiments. only by the isal ces proided by the patterns, and not by pheromonal ces emanating from the feeder. Analysis of possible biases arising from preiosly rewarded patterns and positions To check whether a bee s choice was inflenced by the identity or the position of the pattern at which it had been preiosly rewarded, we analysed the choice freqencies for the preiosly rewarded pattern and position immediately after the sample stimls was changed, or the test stimli were rearranged. The reslts are smmarized in Tables and 2. Table pertains to the learning tests and Type transfer tests. Table 2 pertains to the Type 2 transfer tests. Statistical analysis shows that the choice freqencies for preiosly rewarded patterns and positions are significantly smaller than random choice leel (25%) or are not significantly different from it (Tables and 2). Ths, there is no preference for the preiosly rewarded pattern or position. For example, if a stimls belonging to the category of starshaped flowers was sed as a rewarding stimls in a preios trail, and a stimls from a different category (circlar flower, landscape or plant) was rewarded in the following experiments, only 5% of the isits wold still be to the star-shaped flower stimls in the sbseqent trails. This leel is significantly lower than 25% (P<.5, Grop, Table ). Similarly, if the correct (rewarded) test stimls was in Position 3 in a preios trail, and in a different position (, 2 or 4) in a sbseqent trail, then only 3% of the isits wold be to the formerly rewarded position (Position 3). This is not significantly different from 25% (P>.5, Grop, Table ). Ths, it is clear that the bees performance in choosing the matching or similar stimls was not significantly affected by preiosly rewarded patterns or their positions. Performance of indiidal bees While space constraints preent s from listing the choice freqencies of each bee in each sitation, data from two indiidal bees (Table 3) demonstrate consistent performance. The reslts from the two bees are statistically indistingishable. Both bees performed well in all of the arios tests.

7 Groping of isal objects by honeybees 3295 Table. Bias analysis for learning tests and Type transfer tests Preiosly rewarded Grop Grop 2 Grop 3 category or position Fr N P Fr N P Fr N P Flower F.5 36 <.5. 4 > <. Flower f 82 > >.5. 2 <. Plant.5 55 < <..2 <. Landscape.5 49 < <. 96 >.5 Position.5 25 < > >.5 Position <.5 7 > >.5 Position > > >.5 Position > > <.5 Bias analysis for learning tests and Type transfer tests. Fr, mean choice freqency; N, total nmber of isits. The P ales are reslts of t- tests for a statistically significant difference from the random-choice leel of 5. Table 2. Bias analysis for Type 2 transfer tests Preiosly rewarded G to G2 G to G3 G to G4 G3 to G3* category or position Fr N P Fr N P Fr N P Fr N P Flower F.3 45 > <..7 6 < <. Flower f.6 6 < > > >.5 Plant.5 38 < > <.. 27 >.5 Landscape.2 45 < < > <. Position.3 49 >.5 56 > > >.5 Position 2 34 > > > >.5 Position > > < >.5 Position > < <.5 53 >.5 Details as in Table. Table 3. Performance of two indiidal bees in the learning and transfer tests Choice freqency for correct test Test type Bee nmber Bee nmber Learning test: G to G 48/83=.58 3/45=.67 Type transfer test: G2 to G2 55/74=.74 45/64=.7 Type transfer test: G3 to G3 38/46=.83 29/43=.67 Type 2 transfer test: G to G2 2/3=.7 27/28=.7 Type 2 transfer test: G to G3 24/42=.57 2/37=.57 Type 2 transfer test: G to G4 26/34=.76 29/34=.85 The nmerator and denominator in each fraction denote the nmbers of correct and total choices, respectiely. Discssion The reslts of the present stdy show clearly that bees can learn a match-to-sample task that inoles choosing between for different complex images. Frthermore, bees can generalize this learned matching concept to () match noel stimli that they hae neer preiosly encontered and (2) match stimli that are not identical, bt similar. Or reslts do not reeal the specific ces by which similarity is jdged. The landscape scenes were all characterized by ble sky in the pper half of the scene and green/brown bsh in the lower half, separated by a horizon. The plant scenes all consisted of a green stem with leaes and, in one case, a red flower at the top. The flower scenes comprised images that were either roghly circlar, or roghly star-shaped. The colors of the flowers were ariable. Gien that the colors of the stimli were not consistent within categories, it is nlikely that the bees were only sing color per se as a ce to distingish between the categories. Rather, form and, possibly, isal textres are likely to be releant ces as well. This is spported by the obseration that category discrimination was not compromised when the colors of the stimli were remoed (Fig. 4). Were the bees distingishing between the patterns by sing their mean lminance as a ce? To inestigate this possibility, we measred the mean lminance of the patterns (Table 4). The measrements show that the star-shaped flowers, circlar flowers and plants display ery similar lminance. Frthermore, the lminance rankings of the patterns in these three categories are different from grop to grop. The landscapes are slightly dimmer. Ths, the bees cold not hae sed mean lminance as the sole ce to distingish between categories. Were the bees distingishing between the patterns in terms of what M. Hertz termed figral intensity (length of contor

8 3296 S. Zhang and others Table 4. Mean lminance of patterns Mean lminance of patterns (lx) Star-type Circlar-type flowers flowers Plants Landscapes Grop Grop Grop Grop Mean lminance of patterns was measred by a TES 33 Digital Lx Meter (TES Electrical Electronic Corp., Taipei, Taiwan). The lminances were measred 2 cm in front of the patterns. per nit area; Hertz, 933)? With the complex images sed in or experiments, it is irtally impossible to measre this parameter, becase many of the patterns contain complex internal textres. In this circmstance, we beliee that the spatial freqency content of the patterns proides a measre that is approximately eqialent to figral intensity. We therefore compted the spatial power spectra of the patterns (Fig. 6). Visal inspection sggests that the spectra do indeed differ between categories. For example, the images belonging Grop Grop 2 Grop 3 Grop 4 F F2 F3 F4 f f2 f3 f4 Fig. 6. Power spectra of patterns in each of the for grops of stimli sed in the experiments. Spatial freqency ranges from 2.22 cycles cm to cycles cm along the and axes, where the nit spatial freqency represents a period eqal to the diameter of the fll image. P P2 P3 P4 to the circlar flower category all possess spectra that are roghly circlar symmetrical, whereas the images in the landscape category all exhibit spectra that contain enhanced power along the x and y axes. To measre these differences qantitatiely, we compared power leels, within a certain spatial freqency band, of patterns belonging to the same category, as well as across categories. The band chosen for analysis was the region spanning.853. cycles cm spatial freqency along the positie and negatie directions of the axis in the spatial freqency domain. The reslts show that, in most cases, the power in this band is similar for patterns that belong to a gien category, bt different for patterns that belong to different categories. Table 5 shows the relatie power leels for patterns in the arios categories. It also shows the reslts of tests for statistically significant difference between the power leels of arios categories. The power leels of the Landscape images are significantly different from those of the star-shaped flowers ( Flower ) and the circlar flowers ( flowers ) (P<.5). The Flower images also differ significantly from the Plant images (P<.5). Other comparisons, howeer, reeal no significant difference. The aboe analysis raises the possibility that spatial freqency content is one ce by which honeybees classify L L2 L3 L4 natral scenes. It is nlikely to be the only ce, howeer. Other ces might be circlar symmetry (for the circlar flowers), anglar periodicity (for the star-shaped flowers), bilateral symmetry (for the plant stems) and the presence of a horizontal, high-contrast edge, the horizon (for the landscape scenes). Frther work is reqired to ealate the possible roles of these other ces. The reslts of the transfer tests with noel stimli (Fig. 3C) show that the bees performed ery well at picking the noel test stimls that was in the same category as the sample. The honeybees exhibit the same response to noel stimli that differ greatly in their indiidal, low-leel featres. That is, bees treat these highly ariable stimli as eqialent. Clearly, the bees performance is not merely de to rote learning of each of the exemplars. When bees display an apparent ability to distingish between different classes of stimli, are they also displaying a tre ability to generalize across stimli that belong to the same category? Or does this apparent generalization come abot simply becase they cannot distingish between stimli that belong to the same category? The latter possibility seems nlikely becase, at least in some of the

9 Groping of isal objects by honeybees 3297 Table 5. Comparison of the spatial power spectra of different categories and statistical test Flowers F Flowers f Plants Landscapes (.5525±.6) (.5835±.2437) (.65±.22) (.652±.267) Flowers F (.5525±.6) t=2.2, f=6, P<.85 t=5.4, f=6, P<.5 t=6.6, f=6, P<.5 Flowers f (.5835±.2437) t=.95, f=6, P<. t=3.69, f=6, P<.5 Plants (.65±.22) t=2.5, f=6, P<.9 For each category, ales are means ± S.D. of the power in the freqency range.825. cycles cm along the positie and negatie directions of the axis in the spatial freqency domain. Also shown are the reslts of Stdent t-tests for statistically significant differences between the power leels for different categories. cases, the stimli belonging to a gien category differed sbstantially in color. It is well known that bees are generally ery good at learning to discriminate between stimli on the basis of color, orientation, shape or other attribtes (Sriniasan, 994; Lehrer et al., 995; Chittka et al., 993; Vorobye and Menzel, 999; Wehner, 98). Neertheless, we inestigated this qestion rigorosly by examining whether bees cold be trained, in a Y-maze, to distingish between two stimli that belonged to the same category. The training and testing procedre sed for these Y-maze experiments was as employed by Sriniasan and Lehrer (988). In for separate training experiments, we examined whether bees cold be trained to distingish between F2 and F3 in Category, between f3 and f4 in Category 2, between P and P2 in Category 3 and between L and L2 in Category 4. We deliberately chose pairs of stimli that were likely to be the most difficlt to discriminate. The reslts showed that honeybees cold learn each of these discrimination tasks. The choice freqency in faor of the positie stimls ranged from 62.9% to 78.4%, depending pon the particlar experiment, bt in each case this freqency was significantly higher than the randomchoice leel of 5% (P<. in three cases, and P<.5 in one case). Ths, while the bees were learning to distingish between for different classes of stimli in or main experiments, they were also exhibiting tre perceptal generalization across stimli that belonged to a gien class. It is clear that the highly ariable complex images in the same category are discriminable by the honeybee. Howeer, they are treated as eqialent when bees are reqired to distingish between images in different categories. So far we hae exclded the possibility that bees might be sing some single low-leel featres as a ce to categorise these complex isal objects. Some other configrational properties cold still be sed in this categorisation. This might inclde circlar symmetry (for the circlar flowers), anglar periodicity (for the star-shaped flowers), bilateral symmetry (for the plant stems) and the presence of a horizontal, highcontrast edge, the horizon (for the landscape scenes). It is likely that this categorisation is based on a combination of low-leel featres as well as some configrational properties. It is also clear that that the behaior of the trained bees does not necessarily represent categorization in the sense of associating a specific concept, meaning or releance, to each of the stimls classes, as is likely to be the case in hmans. With bees, establishing the similarity between images belonging to a gien category cold simply be based on a comparison of the responses eoked by the images across mltiple neral channels representing different featres. Recently, Halford distingished and ranked a series of different leels of cognitie processes that proides a theoretical basis for interpreting findings abot simpler cognitie processes by infants and yonger children as well as hmanlike competencies in animals. (Halford et al., 998; G. S. Halford, S. Phillips and W. H. Wilson, manscript sbmitted). The apparent ability of bees to classify objects may be a process that works at a lower leel than that in hmans. Or findings sggest that the honeybee possesses an ability to grop similar stimli into categories. Frther work is reqired to determine whether the for classes of complex images that we hae sed in or experiments represent for categories of natral scenes that are important to a foraging honeybee, and if so, whether these categories are innately programmed by eoltion, or learned indiidally throgh experience. We thank Rebecca Hillman for assistance with the experiments, Bill Speed and Angela Mitchell for constrcting the maze apparats, and Ang Si and Daid Gez for discssion. This research was spported partly by Grant N from the US Defence Adanced Research Projects Agency to M.V.S. and by Grant DP from the Astralian Research Concil to S.W.Z. References Cerella, J. (979). Visal classes and natral categories in the pigeon. J. Exp. Psychol. Hm. Percept. Perform. 5, Chittka, L., Vorobye, M., Shmida, A. and Menzel, R. (993). Bee color ision the optimal system for discrimination of flower color with three spectral photoreceptor types. In Sensory Systems of Arthropods (ed. K. Wiese), pp Basel, Switzerland: Birkhaeser Verlag Press. Collett, T. S., Fry, S. N. and Wehner, R. (993). Seqence learning by honeybees. J. Comp. Physiol. A 72, Collett, T. S. (996). Insect naigation en rote to the goal-mltiple strategies for the se of landmarks. J. Exp. Biol. 99, Daenport, R. K. and Rogers, C. M. (97). Perception of photographs by apes. Behaior 39, Dkas, R. and Waser, N. M. (994). Categorization of food types enhances foraging performance of bmblebees. Anim. Beha. 48, -6. Freedman, D. J., Riesenhber, M., Poggio, T. and Miller, E. K. (2). Categorical representation of isal stimli in the primate prefrontal cortex. Science 29,

10 3298 S. Zhang and others Girfa, M., Eichmann, B. and Menzel, R. (996). Symmetry perception in an insect. Natre 382, Girfa, M., Zhang, S. W., Jenett, A., Menzel, R. and Sriniasan, M. V. (2). The concepts of sameness and difference in an insect. Natre 4, Groebner, D. F. and Shannon, P. W. (985). Hypothesis testing sing nonparametric tests. Bsiness Statistics A Decision-Making Approach, pp Charles E. Merril Pblishing Company. Halford, G. S., Wilson, W. H. and Phillips, S. (998). Processing capacity defined by relational complexity: Implications for comparatie, deelopmental, and cognitie psychology. Beha. Brain Sci. 2, Herrnstein, R. J. (984). Objects, categories and discrimination stimli. In Animal Cognition (ed. H. L. Roitblat, T. G. Beer, and H. S. Terrace), pp Hillsdale, NJ, USA: Erlbam. Hertz, M. (933). Über figrale Intensitäten nd Qaslitäten in der optischen Wahrnehmng der Biene. Biol Zbl. 53, -4. Horridge, G. A. and Zhang, S. W. (995). Pattern ision in honeybees (Apis mellifera): flower-like patterns with no predominant orientation. J. Insect Physiol. 4, Horridge, G. A. (996). The honeybee (Apis mellifera) detects bilateral symmetry and discriminates its axis. J. Insect Physiol. 42, Hber, L., Troje, N. F., Loidolt, M., Ast, U. and Grass, D. (2). Natral categorization throgh mltiple featre learning in pigeons. Qart. J. Exp. Psychol. 53B, Jdd, S. P. D. and Collett, T. S. (998). Mltiple stored iews and landmark gidance in ants. Natre 392, Lehrer, M., Horridge, G. A., Zhang, S. W. and Gadagkar, R. (995). Shape ision in bees: innate preference for flower-like patterns. Phil. Trans. R. Soc. Lond. B 347, Mallot, R. W. and Siddall, J. W. (972). Acqisition of the people concept in pigeons. Psychol. Rep. 3, 3-3. Martin-Maliel, J. and Fagot, J. (2). Cross-modal integration and conceptal categorization in baboons. Beha. Brain Res. 22, Menzel, R. and Girfa, M. (2). Cognitie architectre of a mini-brain: the honeybee. Trends Cognitie Sci. 5, 6-7. Roitblat, H. L. (987). Langage, concept and intelligence. In Introdction to Comparatie Cognition, pp New York: W. H. Freeman and Company. Sriniasan, M. V. and Lehrer, M. (988). Spatial acity of honeybee ision and its spectral properties. J. Comp. Physiol. A 62, Sriniasan, M. V. (994). Pattern recognition in the honeybee: recent progress. J. Insect Physiol. 4, Sriniasan, M. V. and Zhang, S. W. (998). Probing perception in a miniatre brain: pattern recognition and maze naigation in honeybees. Zoology, an Hateren, J. H., Sriniasan, M. V. and Wait, P. B. (99). Pattern recognition in bees: orientation discrimination. J. Comp. Physiol. A 67, Vogels, R. (999). Categorization of complex isal images by rhess monkeys. Part : behaioral stdy. Er. J. Nerosci., Vorobye, M. and Menzel, R. (999). Flower adertisement for inertebrates: bees. A case stdy. In The Ecology of Vision (ed. S. N. Archer, M. B. A. Djamgoz, E. R. Loew, J. C. Partridge and S. Vallerga), pp Klwer Academic Pblishers. Wehner, R. (97). The generalization of directional isal stimli in the honey bee, Apis mellifera. J. Insect Physiol. 7, Wehner, R. (98). Spatial ision in arthropods. In Vision in Inertebrates, Handbook of Sensory Physiology, Vol. 7/6C (ed. H. Atrm), pp Berlin, Heidelberg: Springer-Verlag. Wehner, R., Bleler, S. and Shah, D. (99). Bees naigate sing ectors and rotes rather than maps. Natrwissenschaften 77, Wehner, R., Michel, B. and Antonsen, V. (996). Visal naigation in insects: copling of egocentric and geocentric information. J. Exp. Biol. 99, Zhang, S. W., Bartsch, K. and Sriniasan, M. V. (996). Maze learning by honeybee. Nerobiol. Learn. Mem. 66, Zhang, S. W., Lehrer, M. and Sriniasan, M. V. (999). Honeybee memory: Naigation by associatie groping and recall of isal stimli. Nerobiol. Learn. Mem. 72, 8-2. Zhang, S. W., Miztani, A. and Sriniasan, M. V. (2). Maze naigation by honeybees: learning path reglarity. Learn. Mem. 7,