- Email: [email protected]
Pattern recognition in insects
Pattern recognition in insects Martin Heisenberg University of WCirzburg,Wiirzburg, Germany After 80 years of research, the field of insect pattern recognition is about to move from the behavioral to the neuronal level. Recent experiments on bees, ants and flies indicate that pattern recognition must be seen as the recruitment of behavioral operations that help the nervous system to solve a task using a small number of potentially simple processing steps. These may now be identified physiologically. Current Opinion in Neurobiology 1995, 5:475-481
Introduction Since the discovery and first characterization of pattern recognition in honeybees by Karl von Frisch and his school in the early 1900s (see [1-4]), this field has progressed slowly. In the post-war era of cybernetics, studying visual course control in insects seemed more appealing, because, with their robust motion-dependent behavioral responses, insects could be conceived as simple input-output systems (see e.g. [5,6]). As a practical matter, patterns did ¢ot drive .behavioral responses as reliably as motion stimuli. On the theoretical side, the nature of the sensory-motor interface in pattern recognition was a vexing problem that defied model building. By the mid-70s, however, analysis of filmed flight manoeuvres of flies [7,8] strongly suggested that visual guidance by motion cues is an optional and not a strictly deterministic affair, and that behavioral programs can be recruited by trial and error [9-12]. The growing interest in insect pattern recognition in the 80s, I believe, had to do with the rediscovery that behavior is a genuinely active process. The importance of behavioral operations in pattern recognition will be the main focus of this review. Landmark orienlation and flower constancy Most of what we know about pattern recognition in insects derives from the analysis of two behaviors, landmark orientation and flower identification. The former behavior refers to insects' ability to use visual landmarks to guide their excursions and to locate their nest or other important sites (see Fig. 1). The latter behavior concerns bees' and other insects' ability to identify flower species by their shapes. In both cases, visual images must be stored and the engrams must be used upon retrieval to trigger a certain behavioral disposition or program. It is likely that distinct motivational states underlie landmark and flower recognition, but it is not clear how much the two
differ. Fast flying insects are evolutionarily much older than flowers; therefore, Landmark recognition should be 'older' than flower recognition. Hence, the latter may be a specialization of the former. Controlled laboratory experiments are often not able to determine to which of the two pattern recognition categories the tested behavior belongs. Nevertheless, the distinction should be kept in mind.
Real image or parametrization? Sensory recognition may be described as the match between an engram of a previous event and the neural correlate of the actual one. In visual pattern recognition, the engram might be conceived as an actual image (i.e. a kind of 'photograph' of the object or scene). For matching, then, the stored image and the actual image would have to be somehow superimposed (see e.g. [13-16]). How are these images stored? Are they 'real' images or parametrizations? Storage of real images may not be very economical. An animal walking or flying around with a comprehensive album of pixel-by-pixel representations of its familiar territory and food sources might be overloaded by irrelevant information. Alternatively, images might be stored as values for a set of parameters, such as size, luminance, contour length, orientation of edges, retinal position of the salient features, etc. [3,17-19]. These parameters could be envisaged as the co-ordinates of a multidimensional perceptual space in which each pattern is defined by a specific point in that multidimensional space [20,21]. A template consisting of a set of parameters might seem less demanding on storage space than a real-image template. However, this may depend on the specific task to be solved. (It may be worth pointing out that any task that can be solved by real-image matching can also be solved by parametrization, because it's always possible to chose a set of parameters that allows for reconstruction of the relevant features of the image. On the other hand, some tasks that can be solved by parametrization, such
Abbreviations Q~size of overlap; R--actual retinal image; sf--similarity function. © Current Biology Ltd ISSN 0959-4388
475
476
Sensory systems
(a)
.2
(b)
Fig. 1. Insects are guided by landmarks
to important locations. (a) In a classic experiment, Tinbergen [68] marked
the entrance of a wasp's nest with a
© 1995 Current Opinion in Neurobiology
as generalizations and invariance formation, can not be solved by matching real images.) Insects often use landmarks to find a previously held orientation or position in space (see Fig. 1). H o w do they do this? One way would be to store the retinal co-ordinates o f some of the features o f the scene that they want to return to as a frame o f reference, while maintaining a specific position or orientation. To re-locate the scene of interest, the insect would have to place itself in the same orientation and position s o that the particular values o f the retinal co-ordinates for the stored features (i.e. the template) match those o f the actual image; this process is called retinotopic matching. The retinal co-ordinates to be matched might be stored as a real image in a retinotopic array o f visual elements or, alternatively, as a set o f parameter values elsewhere in the brain. Below, I will briefly review some o f the experimental evidence showing that insects do indeed store retinal co-ordinates of salient features in a scene
circle made of pine cones. After the wasp had become accustomed to the layout, (b) the cones were moved to a new location while the wasp was out hunting. When the wasp returned, it searched in the center of the pine cone circle for the nest. Modified from [68].
and later match them to the actual image. (See [22] for a review o f the literature up to about 1978.) Insects use retinotopic matching to recognize visual patterns
As early as 1969, Wehner [13] proposed retinotopic matching as the basis o f pattern recognition in insects. Training bees to discriminate between black and white sectors of varying size and orientation on large circular disks, Wehner [14] showed that the relative size o f the overlapping and non-overlapping areas of the training and test patterns was a good predictor o f the degree to which an animal would treat the two patterns as similar. The bees always fixated the center o f the disk before landing so that about the same part o f the retina was stimulated during the training and test trials [15]. Moreover, no interocular transfer or retinal transfer between ventral and dorsal eye regions could be detected in experiments in which the bees' eyes were partially or
Pattern recognition in insects I-leisenberg 477 totally occluded [23]. Gould [24] trained bees on more flower-like patterns, and arrived at similar conclusions. Since the late 70s, retinotopic matching has been postulated to also be the basis for landmark learning in insects (for bees [25-28]; for ants [29]). Typically, individual bees were trained to feed in a place surrounded by several conspicuous landmarks. Once a bee became accustomed to the surroundings, the food was removed and the bee's search pattern was recorded. In some tests, the size or relative position of the landmarks was also changed. The bee usually expected the food to be where the actual image of the landmarks matched most closely the distribution and size of the landmarks seen from the previous feeding position (established during training). Cartwright and Collett [27,28] addressed the difficult problem of how bees make this retinotopic match. They proposed that bees might generate a filtered version of the actual image, containing only the low spatial frequencies, and that the bees would start searching with this modified template. The authors [27,28] designed computer models that were designed to simulate the bees' approach to the site of an optimal match. They programmed the 'computer bees' to minimize the overall deviations between the edges in the template and the actual image. To simplify matters, they had the computer bees always approach the site from the same direction; this behavioral constraint was later found to actually apply to live bees (see below).
A simple similarity function in tethered flies More recently, retinotopic matching has been studied in the fly Drosophila melanogaster under more restrained laboratory conditions, which make it possible to interfere with the normal visual consequences of flight manoeuvres [30]. Flies were attached to a transducers measuring torque and positioned in a flight simulator, in which the yaw torque of the fly determined the angular velocity o f the surrounding panorama. In this artificial situation, patterns could be presented at fixed elevations in the fly's visual field. The flies were trained to keep certain directions of flight with respect to the panorama and to avoid others [30]. They readily learned to discriminate nearly all pairs of different patterns tested, and even discriminated identical patterns if these were presented at slightly different heights. Neither retinal transfer for vertical displacements of the patterns nor interocular transfer was observed [31]. Interestingly, a vertical shift of as little as 9 ° between training and test patterns abolished retrieval of the memory template [32]. Recognition emerged as a simple procedure of measuring the similarity between the memory template and the actual figure. Flies seemed to compare the size of the overlap (Q) between the template and the actual image with the size of the actual retinal image (P,) under conditions of optimal match (similarity function if=Q/P,). No special analysis of shape was observed under these conditions [33].
This leaves little doubt that insects can use retinotopic matching in orientation and, more importantly, it gives the concept of retinotopic matching a concrete, quantitative meaning. Yet, the task studied here is much simpler than that achieved by the computer bees [27]. The latter are guided by the discrepancies between the template and the actual scene in their translatory movements towards the place where the template is optimally matched, and, presumably, where it had been inscribed. In the flight simulator, only horizontal rotatory movement occurs and the fly can find the optimal match by a simple rule: 'revert the direction of turning if the value of the similarity function decreases" One does not have to design a computer fly model to confirm that this algorithm solves the task. Needless to say, the algorithm generated from the flight simulator work only provides for angular orientation. The corresponding behavior in free flight or walking would substitute for, or add to, compass orientation. A conclusive experiment showing that freely moving insects align their direction of locomotion on the basis of landmark memory has been reported by Htlldobler [34], who investigated the effect of the canopy pattern on the homing direction of ants foraging on the forest floor. On their relatively short excursions, these animals apparently refer to the silhouette of the canopy against the sky instead of celestial cues. H/511dobler [34] suggested that the ants leaving the nest take a 'snapshot picture' of their surroundings and use this as a reference vector for the trip. Future experiments will have to show whether in ants the similarity function (if=Q/P,), as derived for tethered flies, is, indeed, used for angular orientation in situations in which the sun compass is not available. The function would be of little use if finding the match required translatory displacement. Therefore, it does not conform with the types of computer models mentioned above [25,27].
Memories are linked to special behavior patterns How can one bridge the gap between behavior in the flight simulator and that of freely moving animals? For nearly a century it has been known that bees and wasps, when leaving a site to which they need to return, perform special flight manoeuvres, called orientation flights, that seem to be crucial for them to find their way back [35]. A bee, for instance, that leaves the hive for the first time, turns around to face the entrance of the hive and circles around it in arcs of increasing size and distance. Only bees that are familiar with the terrain leave the hive in a straight course; however, they resume orientation flights whenever something has changed in the spatial lay out around the hive, or the site of interest (summarized in [22,36]). In the past years, interest in these flight manoeuvres has surged, and remarkable progress has been made in their understanding [37-41,42"',43°']. Collett and Lehrer [41] report that in the orientation flights of the wasp Vespula vulgaris, the end points of successive
478
Sensorysystems arcs are strikingly aligned, suggesting that these might be the moments when images are stored. In agreement with Zeil [38], they observed similarities in the flight patterns of the departures and subsequent arrivals [41]. They speculated that the manoeuvres might serve not only to record the apparent size and angular position of the landmarks (as in a snapshot), but may, by means of motion parallax, also record the distance of nearby landmarks. This has now been elegantly confirmed for ground-nesting bees [42"] and for honeybees [40,43"]. Bees were trained to feed at a fixed distance from a black cylinder [42"']. Some bees were allowed to see the cylinder only during arrival, others only during departure. In the test, the size or distance to the feeder, or both, could be changed. It turned out that on their approach, the bees recorded only the angular size of the cylinder, but in their orientation flights on the first few departures, they also measured its distance. Experienced bees relied predominantly on angular size ([43"']; for wasps, see [38]). In a further set of experiments, Collett and Baron [44"] showed that their bees from Sussex, UK, have a preference for heading south in their final approach towards a feeding site and in the orientation flights leaving it. It would make sense that bees store a coherent (compass-based) memory templates for both arrival and departure. In any case, approaching and leaving a feeding site with a single preferred orientation reduces the number of snapshots required and facilitates retinotopic matching. Apparently, in a familiar surround, piloting - - that is, keeping a certain orientation with respect to a stable reference vector and fixating a target - - brings insects close enough to their goal that retinotopic matching can do the rest. Menzel and co-workers [45] have now demonstrated that bees released at an unfamiliar site choose a direct route towards the hive by interpolating between two learned homing vectors. On the other hand, orientation flights seem to be required to gain familiarity as they are a means of identifying close-by landmarks, which can be used as targets for piloting and for the final search [46]. What exactly bees and wasps learn during these manoeuvres is still not clear. Whether they store a complex distribution and sequence of motion parallax cues in relation to their own movements or whether they take a series of snapshots from different angles remains to be determined. I will briefly return to this problem later. As mentioned above, storage of specific visual memories is linked to certain motor patterns. Even though, under most circumstances, motion parallax cues are just a part of the visual guidance system, during orientation flights, they control the formation of, or become themselves, memory templates. Similarly, retrieval is linked to behavior. Collett et al. [47] have provided suggestive evidence that bees on their foraging routes learn sequences of behavioral operations, each one consisting of a set of detailed instructions about what direction to take and how far to travel. The bee seems to
activate different memory templates in a sequence- and context-dependent manner. It is most intriguing to discover that insects segrnent a scene into landmarks that are close to and landmarks that are far away from a goal. This implies that they restrict certain kinds of processing to particular features in the visual field. As much as they prefer close landmarks in their final search of the goal [42",43"',46], they might use only distant ones as reference for their flight direction. Segmentation of the scene is reminiscent of selective attention in humans [48], and has been observed long ago in hoverflies [49], as well as in the tethered Drosophila at the torque meter ([50]; tk Wolf, M Heisenberg, unpublished data).
Insects can discriminate patterns by certain parameters The notion of retinotopic matching is preceded by 30 years of intense research focused on identifying which parameters insects use to discriminate patterns. Color and luminance contrast were seen as obvious independent parameters contributing to this task. In addition, contour density (figural intensity, high spatial frequencies) and a less well defined parameter, 'figural quality', were also proposed. This early phase of the investigation will not be discussed here (but see [22]). Recently, it has been unambiguously shown that bees can be made to discriminate patterns solely on the orientation of the patterns' contours [51,52]. In these experiments, the bees had to chose between two random gratings oriented at right angles to each other (e.g. horizontal and vertical). Each bee was successively exposed during the training to 10 pairs of gratings differing in the widths of the black and white bars. Bees quickly learned to choose the patterns with the right orientation, even if they were confronted with two patterns they had not encountered before. Hence, they could not have relied on the similarity function (sf=Q/R) for this task. Therefore, insects must possess special feature detectors for the orientation of edges. What these findings do not show is to which degree recognition by contour orientation is independent of the retinal co-ordinates o f the patterns. In other words, is retinotopic matching required even in cases when contour orientation is the only discriminating feature? The same question applies to the discrimination of patterns differing only by color or luminance contrast [21]. Retention of the retinal co-ordinates may be relevant to discrimination tasks, at they ot~en have an orientational component.
Orientation detectors are sensitive to stationary patterns Srinivasan et al. [53] ventured that sensitivity to contour orientation might be due to the intrinsic directionality
Pattern recognition in insects
o f the well known motion detectors [6] stimulated by the bees' own movements relative to the stationary test patterns. As with earlier experiments [51], the authors [53] used 10 different pairs of orthogonally oriented random gratings, but this time they flashed them in a random sequence every 0.5s for only 2 ms on two computer screens so as to eliminate any motion or pseudo-motion stimuli. Still, after some training, the bees were able to consistently choose the orientation paired with the reward. This stimulation procedure probably excludes the above-mentioned directional motion detectors, but the new orientation detectors might still require jittery movement of edges as input. The suggestion that stationary images, not the 'motion signature', of the scenery are stored raises tantalizing questions about the memory template. How can the notion of a snapshot be reconciled with the following three observations: first, that motion parallax seems to be of primary importance during the orientation flights; second, that young worker bees need several minutes to get acquainted with the landmarks around their hive; and, third, that in the flight simulator, the panorama is in permanent motion. How many independent snapshots can an insect store and how can snapshots be elaborated in successive exposures? How are snapshots tied in with the behavioral options they stimulate? How can several slightly different snapshots be combined with the same behavioral option? We will not have to wait long for answers to some of these questions.
A separate processing stage for pattern recognition Like mammals, insects have distinct input channels for motion and color. Motion-dependent behavior is mediated by only one type of receptor, the green receptors [54], and is insensitive to hue contrast [55-57]. Conversely, the input channels to color vision do not mediate motion responses [57]. Having demonstrated that pattern recognition in the bee did not require directional motion [53], Zhang and Srinivasan [58"] discovered that bees still could be primed to use their motion detectors as input channels for pattern recognition. Without prior experience, bees are unable to learn to discriminate textured figures in front of a similarly textured background, although they can detect the edges by motion parallax, as was shown independently [59-61]. If, however, bees are first trained with the same figures in black and white, they subsequently also recognize the textured ones. After this sensitization, they are readily trained to discriminate textured figures in a textured background, even if they had not seen them before. An elegant explanation for this priming effect was proposed by Collett [62"]. If a bee flies straight towards the patterns, these are well camouflaged because motion parallax is small. They would stand out much more clearly if the bee would take a meandering flight trajectory. It should be possible
Heisenberg
to test this explanation by observing bees during their approach. The parallel organization in the insect visual system allowed Zhang et ai. [63"'] to study the relation between the various input channels and pattern recognition. Patterns presented during training as isoluminant color contrast could still be discriminated if in the test they were discernible only by the green receptors (luminance contrast) or by motion parallax. The same general resuk was obtained for all other combinations of input channels during training and testing. Hence, one has to assume that all three input channels - - the color, luminance contrast and movement detection - - feed into an independent station for pattern recognition.
Conclusions and outlook What makes insect pattern recognition so exciting at present is that after eight decades, this field is now approaching the physiological level. Behavioral analysis has become sufficiently fine-grained to permit pertinent conjectures about the underlying neural substrate. We may now speculate where in the insect brain the memory templates for visual pattern recognition are localized. Among the obvious regions to be considered are the optic lobes and mushroom bodies [27]. I n the optic lobes, anatomically obvious candidates for 'storage cells' (e.g. large sets of similar-looking small neurons in each column) have not been detected [64]. O'CarroU [65] has recently described orientation-sensitive neurons in the dragonfly lobula. Such cells, if present in the bee, might be part of the above-mentioned orientation detectors [53]. One would have to search for the templates downstream of these cells [63"']. How about the mushroom bodies? At least for landmark recognition, as assessed in the flight simulator with Drosophila, these are ruled out since they can be genetically and chemically ablated without negative consequences for this behavior [66,67]. In bees, the mushroom bodies, with their massive visual input, might be more important for visual pattern recognition than in flies. Further behavioral analysis is required before a search for the memory templates has a fair chance of success. What this brief account tries to convey to the reader is that visual pattern recognition, which was formerly considered to be a 'cognitive' process of considerable complexity, now appears to be a task-related recruitment of behavioral operations combined with a few potentially simple neuronal processing steps in the visual input. In the brain sciences, behavioral research plays a crucial role. For phylogenetical reasons, higher brain functions can only be adequately understood in the context of a detailed behavioral and micro-behavioral analysis.
Acknowledgements 1 thank M Dill for reading and commenting on the manuscript, and TS Collett, M Lehrer, B Konacher and MV Srinivasan for generously contributing unpublished data.
479
480
Sensory systems
References and recommended reading Papers of particular interest, published within the annual period of review, have been highlighted as: • of special interest •of outstanding interest
Equivalence between size and luminance differences in the context of honeybee vision.] 19.
Horridge GA, Zhang SW: Pattern vision of bees: flower like patterns with no predominant orientation. J Insect Physiol 1995, in press.
20.
Ronacher B: Beitrag einzelner Parameter zum wahmehmungs~ gem~ll~en Unterschied yon zusammengesetzten Reizen hei der Honigblene. Biol Cybem 1979, 32:77-83. [Title translation: Contributions of individual parameters to the perceptual discrimination of modular (composed) stimuli.]
1.
Von Frisch K: Der Farhensinn und Formensinn der Bienen. Zool Jb Abt allg Zool Physiol 1915, 35:1-182. [Title translation: The senses of color and form in the honeybee.]
2.
Hertz M: Die Organisation des optischen feldes hei der Biene I. Z vergl Physiol 1929, 8:693-748. [Title translation: Organization of the optical field in the honeybee.]
21.
Hertz M: Uber figurale Intensit~iten und Qualit~ten in der optischen Wahrnehmung der Biene. Biol Zbl 1933, 53:10-40. [Title translation: On figural intensity and quality in the optical perception of the honeybee.]
Ronacher B: Pattern recognition in honeybees: multidimensional scaling reveals a city-block metric, vision Res 1992, 32:1837-1843.
22.
Wehner R: Spatial vision in arthroflods. In Handbook of Sensory Physiology, vol VII/6C: Vision in Invertebrates. Edited by Autrum H. Berlin: Springer; 1981:287-616.
23.
Wehner R, Flatt h Visual fixation in freely flying bees. Z Naturforsch 1977, 32:469-471.
24.
Gould JL: How do bees remember flower shape? Science 1985, 227:1492-1494.
25.
Anderson AM: Shape perception in the honey bee. Anim Behav 1977, 25:67-79.
26.
Cartwright BA, Collett TS: How honey bees know their distance from a near-by visual mark. J Comp Physiol 1979, 161:335-355.
3.
4.
5.
6.
Zerrahn G: Formdressur und Formunlerscheidung hei der Honigbiene. Z vergl Physiol 1933, 20:117-150. [Title translation: Form conditioning and form discrimination in the honeybee.] Hassenstein B, Reichardt W: Systemtheoretische Analyse der Zeit-, Reihenfolgen- und Vorzeichenauswertung bei der Bewegungsperzeption des Ribselldifers Chlorophanus. Z Natufforsch 1956, 116:513-524. [Title translation: Theoretical systems analysis of the evaluation of time, sequence and sign of motion perception in the beetle Chlorophanos.] EgelhaafM, Borst A: Movement detection in arthropods. In
Reviews of Oculomotor Research, vol 5: Visual Motion and Its Role in the Stabilisation of Gaze. Edited by Miles FA, Wallman J. Amsterdam: Elsevier; 1993:53-77.
27.
Cartwright BA, Collett TS: Landmark learning in bees. ] Comp Physiol 1983, 151:521-543.
28.
7.
Land MF: Head movement of flies during visually guided flight. Nature 1973, 243:299-300.
Cartwright BA, Collett TS: Landmark maps for honeybees. Biol Cybern 1987, 57:85-93.
29.
8.
Collett TS, Land MF: Visual control of flight behaviour in the hoverfly, Syritta pipiens. J Comp Physiol 1975, 99:1-66.
Wehner R, R~ber F: Visual spatial memory in desert ants, Calaglyphis bicolor (Hymenoptera, Formicidae). Experimentia 1979, 35:1569-1571.
9.
Heisenberg M, Wolf R: On the fine structure of yaw torque in visual flight orientation of Drosophila melanoBaster. J Cornp Physiol 1979, 130:113-130.
30.
Wolf R, Heisenberg M: Basic organisation of operant behaviour as revealed in Drosophila flight orientation. J Comp Physiol A 1991, 169:699-705.
10.
Heisenberg M, Wolf R: Vision in Drosophila. Berlin: Springer Verlag; 1984. [Braitenberg V (Ed): Studies of Brain Function, vol 12.]
31.
11.
Heisenberg M, Wolf R: The sensory-motor link in motiondependent flight control of flies. In Reviews of Oculomotor Research, vol 5: visual Motion and Its Role in the Stabilisation of Gaze. Edited by Miles FA, Wallman J. Amsterdam: Elsevier; 1993:265-283.
Dill M: Unterscheidung und Wiedererkennung visuelier Muster bet der Tanfliege Drosophila melanogaster [PhD thesis]. W0rzburg: Universit~it W(Jrzburg; 1995. [Title translation: Discrimination and recognition of visual patterns in the fly Drosophila melanogaster.]
32.
Dill M, Wolf R, Heisenberg M: Visual pattern recognition in Drosophila involves retinotopic matching. Nature 1993, 365:751-753.
33.
Dill M, Heisenberg M: Visual pattern memory without shape recognition. Philos Trans R So(: Lond [Biol] 1995, in press.
34.
HSIIdobler B: Canopy orientation: a new kind of orientation in ants. Science 1980, 210:86-88.
35.
Peckham GW, Peckham EG: On the instincts and habits of solitary wasps. Wisconsin Geol Nat I-list Survey 1898, 2:1-148.
12.
Wolf R, Heisenberg M: Visual orientation in Drnsophi/a is an operant behavior. Nature 1986, 323:154-156.
13.
Wehner R: Der Mecbanismus der optischen Winkelmessong hei der Biene (Apis mellMca). Zool Anz Supp 1969, 33:586-592. [Title translation: Mechanisms of measurement of angles in the honeybee pattern vision.l
14.
Wehner R: Dorsoventral asymmetry in the visual field of the bee, Apis mellifica. J Comp Physiol 1972, 77:256-277.
36.
15.
Wehner R: Pattern modulation and pattern detection in the visual systems of Hymenoptera. In Information Processing in the Visual System of Arthropods. Edited by Wehner R. Berlin: Springer; 1972:183-194.
Von Frisch, K: Tanzsprache und Orientierung der Bienen. Berlin: Springer; 1965. [Title translation: Dance language and orientation of the honeybee.]
37.
Zeil J: Orientation flights of solitary wasps (Cerceri~ Sphecldae; Hymenoptera) I. Description of flight. J Comp Physiol A 1993, 172:189-205.
16.
Cruse H: Versuch einer quantitativen Besc'hreibun8 des Formensehens der Honighiene. Kybernetik 1972, 11:185-200. [Title translation: On the quantitative description of form perception in the honeybee.]
38.
Zeil J: Orientation flights of solitary wasps (Cereeris; Sphecidae; Hymenoptera) II. Similarities helween orientation and return flights and the use of motion parallax. J Comp Physiol A 1993, 172:207-222.
17.
Wehner R: Pattern recognition. In The Compound Eye and vision of Insects. Edited by Horridge GA. Oxford: Clarendon Press; 1975:75-114.
39.
18.
Ronacher B: ,~quivalenz zwischen Grid, n- und Helligkeitsunterschieden im Rahmen der visoellen Wahrnehmung der Honigblene. Biol Cybern 1979, 32:63-75. [Title translation:
Lehrer M: Locomotion does more than bring the bee to new places. In The Behaviour and Physiology of Bees. Edited by Goodman LJ, Fisher RC. Wallingford: C.A.B. International; 1991:185-202.
40.
Lehrer M: Why do bees turn back and look? J Comp Physiol A 1993, 172:549-563.
Pattern recognition in insects Heisenberg 481 41.
Collett TS, Lehrer M: Looking and learning: a spatial pattern in the orientation flight of the wasp Vespula vulgaris. Proc R Soc Lond [Biol] 1993, 252:129-134.
42. •"
Br0nnert U, Kelber A, Zeil j: Ground-nesting bees determine the location of their nests relative to a landmark by other than angular size cues. J Comp Physiol A 1994, 175:363-369. In two earlier papers [37,38], Zeil provided the first thorough account of orientation flights ('turn-back-and-look', [40]). In the tradition of Coilett and Land (see [8]), he presented a detailed frame-by-frame film analysis showing that wasps move laterally in semi-circles of increasing diameter around their nest entrance and the closest landmarks. He also found that a wasp's flight path at the next arrival is sufficiently similar to the previous orientation flight that one could imagine that the animal is storing the 'motion signature' of the relevant objects during that manoeuvre. In this present study [42ee], Zeil and co-workers demonstrate that wild bees estimate the distance of their nest entrance relative to a familiar landmark largely independently of the angular size at which the landmark appears from that position. The result does not directly prove, but supports, the idea that distance might be measured by motion parallax. 43. "
Lehrer M, Collett TS: Approaching and departing bees learn different cues to the distance of a landmark. J Comp Physiol A 1994, 175:171-177. It is during orientation flights that honeybees learn about the distance of landmarks from the goal. During arrival, they memorize only their angular size. This is an excellent demonstration that memories are attached to specific behavioral operations. 44. •
Collett TS, Baron l: Biological compasses and the coordinate frame of landmark memories in honeybees. Nature 1994, 368:137-140. The bees studied by the authors in Sussex, UK consistently approach a feeding site from the north, using the sun compass and the earth's magnetic field. This is a remarkable finding underscoring the general notion that in many situations a behavioral activity dramatically facilitates the computational task of the brain. 45.
56.
Kaiser W: The relationship between movement detection and colour vision in insects. In The Compound Eye and Vision of Insects. Edited by Horridge GA. Oxford: Clarendon Press;
1975:359-377. 57.
58.
Zhang SW, Srinivasan MV: Prior experience enhances paltem discrimination in insect vision. Nature 1994, 368:330-332. his is an especially beautiful example of visual 'priming'. A memory template of a pattern induces the bee to attend to motion parallax cues revealing the same pattern. In other words, the matching memory template pushes the weak motion parallax cues above threshold (see also
151,53]). 59.
Lehrer M, Srinivasan MV, Zhang SW, Horridge GA: Motion cues provide the bee's visual world with a third dimension. Nature 1988, 332:356-357.
60.
Srinivasan MV, Lehrer M, Horridge GA: Visual figure-ground discrimination in the honeybee: the role of motion parallax at boundaries. Proc R Soc Lond [Biol] 1990, 238:331-350.
61.
Lehrer M, Srinivasan MV: Object detection by honeybees: why do they land on edges. J Comp Physiol A 1993, 173:23-32.
Collett TS: Invertebrate vision: bees learn how to look. Curr Biol 1994, 4:717-719. brief review recounts several telling examples in which insects solve certain tasks by activating the specific behavioral operation that provides them with the needed visual input. 62. his
63. ""
Collett TS, Fry SN, Wehner R: Sequence learning by honeybees.
64.
Fischbach KF, Dittrich APM: The optic lobe of Drosophila melanogaster. I. A Golgi analysis of wild-type structure. Cell Tissue Res 1989, 258:441-475.
65.
O'Carroll D: Feature-detecting neurons in dragonflies. Nature 1993, 362:541-543.
66.
Eyding D: Lemen und Kurzzeltged~chtnis belm operanten Konditionleren auf visuelle Muster bei strukturellen und biochemischen Lernmutanten yon Drosophila rnelanogaster [Diplom thesis]. WLirzburg: Universit~t WLirzburg; 1994. [Title translation: Operant learning and short-term memory for visual patterns in structural and biochemical learning mutants of
67.
Wittig T, Dill M, Wolf R, Heisenberg M: Visual pattern discrimination learning in Drosophila melanogaster without mushroom bodies. In Learning and Memory. Proceedings of the 23rd G~ttingen Neurobiology Conference, vol I. Edited by Eisner N, Menzel R. Stuttgart: Georg Thieme Verlag; 1995:4.
68.
Tinbergen N: The Study of Instinct. Oxford: Clarendon Press; 1951.
J Comp Physiol A 1993, 172:693-706. 48.
Yon Helmholtz H: Handbuch der physiologischen Optik. Leipzig: L. Voss; 1867. [Title translation: Handbook of physiological optics.[
49.
Collett TS, Land MF: Visual spatial memory in a hoverfly. J Comp Physiol 1975, 100:59-84.
50.
Wolf R, Heisenberg M: On the fine structure of yaw torque in visual flight orientation of Drosophila melanogaster. II. A temporally and spatially variable weighting function for the visual field ('visual attention'). J Comp Physiol 1980, 140:69-80.
51.
Van Hateren IH, Srinivasan MV, Wait PB: Pattern recognition in bees: orientation discrimination. J Comp Physiol A 1990, 167:649-654.
52.
Wehner R: The generalization of directional visual stimuli in the honey bee, Apis mellifera. J Insect Physiol 1971, 17:1579-1591.
53. 54.
Heisenberg M, Buchner E: The role of retinola cell types in visual behavior of Drosophila melanogaster. J Comp Physiol 1977, 117:127-162.
Cheng K, Collett TS, Pickhard A, Wehner R: The use of visual landmarks by honeybees: bees weight landmarks according to their distance from the goal. J Comp Physiol A 1987, 161:469-475.
Georg Thieme Verlag; 1995:31.
47.
Kaiser W, Liske E: Die optomotorischen Reaktionen von fixiert fliegenden Bienen bei Reizung mit Speldrallichtern. J Comp Physiol 1974, 89:391-408. [Title translation: Optomotor responses of fixed flying bees stimulated with spectral lights.]
Zhang SW, Srinivasan MV, Collett T: Convergent processing in honeybee vision: multiple channels for the recognition of shape, proc Nail Acad Sci USA 1995, 92:3029-3031. The site of pattern recognition in the brain, that is, the memory template that is deposited and later activated in the matching process, can be reached by at least three independent input routes defined by the stimulus conditions: isoluminant color contrast, monochromatic luminance contrast and motion parallax contrast between figure and ground. Studies like this may eventually help to pin down the neuronal correlate of the memory template.
Menzel R, Joerges J, M011er U: Computational phenomena in long distance orientation of honey bees. In Learning and
Memory. Proceedings of the 23rd G~ttingen Neurobiology Conference, vol I. Edited by Eisner N, Menzel R. Stuttgart: 46.
55.
Srinivasan MV, Zhang SW, Rolfe B: Is pattern vision in insects mediated by 'cortical' processing? Nature 1993, 362:539-540. Menzel R: Spectral response of movement detecting and 'sustaining' fibres in the optic lobe of the bee. J Comp Physiol 1973, 82:135-150.
Drosophila melanogaster.]
M Heisenberg, Theodor-Boveri-lnstitut fiir Biowissenschaften, Lehrstuhl ftir Genetik, University of Wiirzburg, Am Hubland, D-97074 Wiirzburg, Germany. E-mail: [email protected]
Introduction Since the discovery and first characterization of pattern recognition in honeybees by Karl von Frisch and his school in the early 1900s (see [1-4]), this field has progressed slowly. In the post-war era of cybernetics, studying visual course control in insects seemed more appealing, because, with their robust motion-dependent behavioral responses, insects could be conceived as simple input-output systems (see e.g. [5,6]). As a practical matter, patterns did ¢ot drive .behavioral responses as reliably as motion stimuli. On the theoretical side, the nature of the sensory-motor interface in pattern recognition was a vexing problem that defied model building. By the mid-70s, however, analysis of filmed flight manoeuvres of flies [7,8] strongly suggested that visual guidance by motion cues is an optional and not a strictly deterministic affair, and that behavioral programs can be recruited by trial and error [9-12]. The growing interest in insect pattern recognition in the 80s, I believe, had to do with the rediscovery that behavior is a genuinely active process. The importance of behavioral operations in pattern recognition will be the main focus of this review. Landmark orienlation and flower constancy Most of what we know about pattern recognition in insects derives from the analysis of two behaviors, landmark orientation and flower identification. The former behavior refers to insects' ability to use visual landmarks to guide their excursions and to locate their nest or other important sites (see Fig. 1). The latter behavior concerns bees' and other insects' ability to identify flower species by their shapes. In both cases, visual images must be stored and the engrams must be used upon retrieval to trigger a certain behavioral disposition or program. It is likely that distinct motivational states underlie landmark and flower recognition, but it is not clear how much the two
differ. Fast flying insects are evolutionarily much older than flowers; therefore, Landmark recognition should be 'older' than flower recognition. Hence, the latter may be a specialization of the former. Controlled laboratory experiments are often not able to determine to which of the two pattern recognition categories the tested behavior belongs. Nevertheless, the distinction should be kept in mind.
Real image or parametrization? Sensory recognition may be described as the match between an engram of a previous event and the neural correlate of the actual one. In visual pattern recognition, the engram might be conceived as an actual image (i.e. a kind of 'photograph' of the object or scene). For matching, then, the stored image and the actual image would have to be somehow superimposed (see e.g. [13-16]). How are these images stored? Are they 'real' images or parametrizations? Storage of real images may not be very economical. An animal walking or flying around with a comprehensive album of pixel-by-pixel representations of its familiar territory and food sources might be overloaded by irrelevant information. Alternatively, images might be stored as values for a set of parameters, such as size, luminance, contour length, orientation of edges, retinal position of the salient features, etc. [3,17-19]. These parameters could be envisaged as the co-ordinates of a multidimensional perceptual space in which each pattern is defined by a specific point in that multidimensional space [20,21]. A template consisting of a set of parameters might seem less demanding on storage space than a real-image template. However, this may depend on the specific task to be solved. (It may be worth pointing out that any task that can be solved by real-image matching can also be solved by parametrization, because it's always possible to chose a set of parameters that allows for reconstruction of the relevant features of the image. On the other hand, some tasks that can be solved by parametrization, such
Abbreviations Q~size of overlap; R--actual retinal image; sf--similarity function. © Current Biology Ltd ISSN 0959-4388
475
476
Sensory systems
(a)
.2
(b)
Fig. 1. Insects are guided by landmarks
to important locations. (a) In a classic experiment, Tinbergen [68] marked
the entrance of a wasp's nest with a
© 1995 Current Opinion in Neurobiology
as generalizations and invariance formation, can not be solved by matching real images.) Insects often use landmarks to find a previously held orientation or position in space (see Fig. 1). H o w do they do this? One way would be to store the retinal co-ordinates o f some of the features o f the scene that they want to return to as a frame o f reference, while maintaining a specific position or orientation. To re-locate the scene of interest, the insect would have to place itself in the same orientation and position s o that the particular values o f the retinal co-ordinates for the stored features (i.e. the template) match those o f the actual image; this process is called retinotopic matching. The retinal co-ordinates to be matched might be stored as a real image in a retinotopic array o f visual elements or, alternatively, as a set o f parameter values elsewhere in the brain. Below, I will briefly review some o f the experimental evidence showing that insects do indeed store retinal co-ordinates of salient features in a scene
circle made of pine cones. After the wasp had become accustomed to the layout, (b) the cones were moved to a new location while the wasp was out hunting. When the wasp returned, it searched in the center of the pine cone circle for the nest. Modified from [68].
and later match them to the actual image. (See [22] for a review o f the literature up to about 1978.) Insects use retinotopic matching to recognize visual patterns
As early as 1969, Wehner [13] proposed retinotopic matching as the basis o f pattern recognition in insects. Training bees to discriminate between black and white sectors of varying size and orientation on large circular disks, Wehner [14] showed that the relative size o f the overlapping and non-overlapping areas of the training and test patterns was a good predictor o f the degree to which an animal would treat the two patterns as similar. The bees always fixated the center o f the disk before landing so that about the same part o f the retina was stimulated during the training and test trials [15]. Moreover, no interocular transfer or retinal transfer between ventral and dorsal eye regions could be detected in experiments in which the bees' eyes were partially or
Pattern recognition in insects I-leisenberg 477 totally occluded [23]. Gould [24] trained bees on more flower-like patterns, and arrived at similar conclusions. Since the late 70s, retinotopic matching has been postulated to also be the basis for landmark learning in insects (for bees [25-28]; for ants [29]). Typically, individual bees were trained to feed in a place surrounded by several conspicuous landmarks. Once a bee became accustomed to the surroundings, the food was removed and the bee's search pattern was recorded. In some tests, the size or relative position of the landmarks was also changed. The bee usually expected the food to be where the actual image of the landmarks matched most closely the distribution and size of the landmarks seen from the previous feeding position (established during training). Cartwright and Collett [27,28] addressed the difficult problem of how bees make this retinotopic match. They proposed that bees might generate a filtered version of the actual image, containing only the low spatial frequencies, and that the bees would start searching with this modified template. The authors [27,28] designed computer models that were designed to simulate the bees' approach to the site of an optimal match. They programmed the 'computer bees' to minimize the overall deviations between the edges in the template and the actual image. To simplify matters, they had the computer bees always approach the site from the same direction; this behavioral constraint was later found to actually apply to live bees (see below).
A simple similarity function in tethered flies More recently, retinotopic matching has been studied in the fly Drosophila melanogaster under more restrained laboratory conditions, which make it possible to interfere with the normal visual consequences of flight manoeuvres [30]. Flies were attached to a transducers measuring torque and positioned in a flight simulator, in which the yaw torque of the fly determined the angular velocity o f the surrounding panorama. In this artificial situation, patterns could be presented at fixed elevations in the fly's visual field. The flies were trained to keep certain directions of flight with respect to the panorama and to avoid others [30]. They readily learned to discriminate nearly all pairs of different patterns tested, and even discriminated identical patterns if these were presented at slightly different heights. Neither retinal transfer for vertical displacements of the patterns nor interocular transfer was observed [31]. Interestingly, a vertical shift of as little as 9 ° between training and test patterns abolished retrieval of the memory template [32]. Recognition emerged as a simple procedure of measuring the similarity between the memory template and the actual figure. Flies seemed to compare the size of the overlap (Q) between the template and the actual image with the size of the actual retinal image (P,) under conditions of optimal match (similarity function if=Q/P,). No special analysis of shape was observed under these conditions [33].
This leaves little doubt that insects can use retinotopic matching in orientation and, more importantly, it gives the concept of retinotopic matching a concrete, quantitative meaning. Yet, the task studied here is much simpler than that achieved by the computer bees [27]. The latter are guided by the discrepancies between the template and the actual scene in their translatory movements towards the place where the template is optimally matched, and, presumably, where it had been inscribed. In the flight simulator, only horizontal rotatory movement occurs and the fly can find the optimal match by a simple rule: 'revert the direction of turning if the value of the similarity function decreases" One does not have to design a computer fly model to confirm that this algorithm solves the task. Needless to say, the algorithm generated from the flight simulator work only provides for angular orientation. The corresponding behavior in free flight or walking would substitute for, or add to, compass orientation. A conclusive experiment showing that freely moving insects align their direction of locomotion on the basis of landmark memory has been reported by Htlldobler [34], who investigated the effect of the canopy pattern on the homing direction of ants foraging on the forest floor. On their relatively short excursions, these animals apparently refer to the silhouette of the canopy against the sky instead of celestial cues. H/511dobler [34] suggested that the ants leaving the nest take a 'snapshot picture' of their surroundings and use this as a reference vector for the trip. Future experiments will have to show whether in ants the similarity function (if=Q/P,), as derived for tethered flies, is, indeed, used for angular orientation in situations in which the sun compass is not available. The function would be of little use if finding the match required translatory displacement. Therefore, it does not conform with the types of computer models mentioned above [25,27].
Memories are linked to special behavior patterns How can one bridge the gap between behavior in the flight simulator and that of freely moving animals? For nearly a century it has been known that bees and wasps, when leaving a site to which they need to return, perform special flight manoeuvres, called orientation flights, that seem to be crucial for them to find their way back [35]. A bee, for instance, that leaves the hive for the first time, turns around to face the entrance of the hive and circles around it in arcs of increasing size and distance. Only bees that are familiar with the terrain leave the hive in a straight course; however, they resume orientation flights whenever something has changed in the spatial lay out around the hive, or the site of interest (summarized in [22,36]). In the past years, interest in these flight manoeuvres has surged, and remarkable progress has been made in their understanding [37-41,42"',43°']. Collett and Lehrer [41] report that in the orientation flights of the wasp Vespula vulgaris, the end points of successive
478
Sensorysystems arcs are strikingly aligned, suggesting that these might be the moments when images are stored. In agreement with Zeil [38], they observed similarities in the flight patterns of the departures and subsequent arrivals [41]. They speculated that the manoeuvres might serve not only to record the apparent size and angular position of the landmarks (as in a snapshot), but may, by means of motion parallax, also record the distance of nearby landmarks. This has now been elegantly confirmed for ground-nesting bees [42"] and for honeybees [40,43"]. Bees were trained to feed at a fixed distance from a black cylinder [42"']. Some bees were allowed to see the cylinder only during arrival, others only during departure. In the test, the size or distance to the feeder, or both, could be changed. It turned out that on their approach, the bees recorded only the angular size of the cylinder, but in their orientation flights on the first few departures, they also measured its distance. Experienced bees relied predominantly on angular size ([43"']; for wasps, see [38]). In a further set of experiments, Collett and Baron [44"] showed that their bees from Sussex, UK, have a preference for heading south in their final approach towards a feeding site and in the orientation flights leaving it. It would make sense that bees store a coherent (compass-based) memory templates for both arrival and departure. In any case, approaching and leaving a feeding site with a single preferred orientation reduces the number of snapshots required and facilitates retinotopic matching. Apparently, in a familiar surround, piloting - - that is, keeping a certain orientation with respect to a stable reference vector and fixating a target - - brings insects close enough to their goal that retinotopic matching can do the rest. Menzel and co-workers [45] have now demonstrated that bees released at an unfamiliar site choose a direct route towards the hive by interpolating between two learned homing vectors. On the other hand, orientation flights seem to be required to gain familiarity as they are a means of identifying close-by landmarks, which can be used as targets for piloting and for the final search [46]. What exactly bees and wasps learn during these manoeuvres is still not clear. Whether they store a complex distribution and sequence of motion parallax cues in relation to their own movements or whether they take a series of snapshots from different angles remains to be determined. I will briefly return to this problem later. As mentioned above, storage of specific visual memories is linked to certain motor patterns. Even though, under most circumstances, motion parallax cues are just a part of the visual guidance system, during orientation flights, they control the formation of, or become themselves, memory templates. Similarly, retrieval is linked to behavior. Collett et al. [47] have provided suggestive evidence that bees on their foraging routes learn sequences of behavioral operations, each one consisting of a set of detailed instructions about what direction to take and how far to travel. The bee seems to
activate different memory templates in a sequence- and context-dependent manner. It is most intriguing to discover that insects segrnent a scene into landmarks that are close to and landmarks that are far away from a goal. This implies that they restrict certain kinds of processing to particular features in the visual field. As much as they prefer close landmarks in their final search of the goal [42",43"',46], they might use only distant ones as reference for their flight direction. Segmentation of the scene is reminiscent of selective attention in humans [48], and has been observed long ago in hoverflies [49], as well as in the tethered Drosophila at the torque meter ([50]; tk Wolf, M Heisenberg, unpublished data).
Insects can discriminate patterns by certain parameters The notion of retinotopic matching is preceded by 30 years of intense research focused on identifying which parameters insects use to discriminate patterns. Color and luminance contrast were seen as obvious independent parameters contributing to this task. In addition, contour density (figural intensity, high spatial frequencies) and a less well defined parameter, 'figural quality', were also proposed. This early phase of the investigation will not be discussed here (but see [22]). Recently, it has been unambiguously shown that bees can be made to discriminate patterns solely on the orientation of the patterns' contours [51,52]. In these experiments, the bees had to chose between two random gratings oriented at right angles to each other (e.g. horizontal and vertical). Each bee was successively exposed during the training to 10 pairs of gratings differing in the widths of the black and white bars. Bees quickly learned to choose the patterns with the right orientation, even if they were confronted with two patterns they had not encountered before. Hence, they could not have relied on the similarity function (sf=Q/R) for this task. Therefore, insects must possess special feature detectors for the orientation of edges. What these findings do not show is to which degree recognition by contour orientation is independent of the retinal co-ordinates o f the patterns. In other words, is retinotopic matching required even in cases when contour orientation is the only discriminating feature? The same question applies to the discrimination of patterns differing only by color or luminance contrast [21]. Retention of the retinal co-ordinates may be relevant to discrimination tasks, at they ot~en have an orientational component.
Orientation detectors are sensitive to stationary patterns Srinivasan et al. [53] ventured that sensitivity to contour orientation might be due to the intrinsic directionality
Pattern recognition in insects
o f the well known motion detectors [6] stimulated by the bees' own movements relative to the stationary test patterns. As with earlier experiments [51], the authors [53] used 10 different pairs of orthogonally oriented random gratings, but this time they flashed them in a random sequence every 0.5s for only 2 ms on two computer screens so as to eliminate any motion or pseudo-motion stimuli. Still, after some training, the bees were able to consistently choose the orientation paired with the reward. This stimulation procedure probably excludes the above-mentioned directional motion detectors, but the new orientation detectors might still require jittery movement of edges as input. The suggestion that stationary images, not the 'motion signature', of the scenery are stored raises tantalizing questions about the memory template. How can the notion of a snapshot be reconciled with the following three observations: first, that motion parallax seems to be of primary importance during the orientation flights; second, that young worker bees need several minutes to get acquainted with the landmarks around their hive; and, third, that in the flight simulator, the panorama is in permanent motion. How many independent snapshots can an insect store and how can snapshots be elaborated in successive exposures? How are snapshots tied in with the behavioral options they stimulate? How can several slightly different snapshots be combined with the same behavioral option? We will not have to wait long for answers to some of these questions.
A separate processing stage for pattern recognition Like mammals, insects have distinct input channels for motion and color. Motion-dependent behavior is mediated by only one type of receptor, the green receptors [54], and is insensitive to hue contrast [55-57]. Conversely, the input channels to color vision do not mediate motion responses [57]. Having demonstrated that pattern recognition in the bee did not require directional motion [53], Zhang and Srinivasan [58"] discovered that bees still could be primed to use their motion detectors as input channels for pattern recognition. Without prior experience, bees are unable to learn to discriminate textured figures in front of a similarly textured background, although they can detect the edges by motion parallax, as was shown independently [59-61]. If, however, bees are first trained with the same figures in black and white, they subsequently also recognize the textured ones. After this sensitization, they are readily trained to discriminate textured figures in a textured background, even if they had not seen them before. An elegant explanation for this priming effect was proposed by Collett [62"]. If a bee flies straight towards the patterns, these are well camouflaged because motion parallax is small. They would stand out much more clearly if the bee would take a meandering flight trajectory. It should be possible
Heisenberg
to test this explanation by observing bees during their approach. The parallel organization in the insect visual system allowed Zhang et ai. [63"'] to study the relation between the various input channels and pattern recognition. Patterns presented during training as isoluminant color contrast could still be discriminated if in the test they were discernible only by the green receptors (luminance contrast) or by motion parallax. The same general resuk was obtained for all other combinations of input channels during training and testing. Hence, one has to assume that all three input channels - - the color, luminance contrast and movement detection - - feed into an independent station for pattern recognition.
Conclusions and outlook What makes insect pattern recognition so exciting at present is that after eight decades, this field is now approaching the physiological level. Behavioral analysis has become sufficiently fine-grained to permit pertinent conjectures about the underlying neural substrate. We may now speculate where in the insect brain the memory templates for visual pattern recognition are localized. Among the obvious regions to be considered are the optic lobes and mushroom bodies [27]. I n the optic lobes, anatomically obvious candidates for 'storage cells' (e.g. large sets of similar-looking small neurons in each column) have not been detected [64]. O'CarroU [65] has recently described orientation-sensitive neurons in the dragonfly lobula. Such cells, if present in the bee, might be part of the above-mentioned orientation detectors [53]. One would have to search for the templates downstream of these cells [63"']. How about the mushroom bodies? At least for landmark recognition, as assessed in the flight simulator with Drosophila, these are ruled out since they can be genetically and chemically ablated without negative consequences for this behavior [66,67]. In bees, the mushroom bodies, with their massive visual input, might be more important for visual pattern recognition than in flies. Further behavioral analysis is required before a search for the memory templates has a fair chance of success. What this brief account tries to convey to the reader is that visual pattern recognition, which was formerly considered to be a 'cognitive' process of considerable complexity, now appears to be a task-related recruitment of behavioral operations combined with a few potentially simple neuronal processing steps in the visual input. In the brain sciences, behavioral research plays a crucial role. For phylogenetical reasons, higher brain functions can only be adequately understood in the context of a detailed behavioral and micro-behavioral analysis.
Acknowledgements 1 thank M Dill for reading and commenting on the manuscript, and TS Collett, M Lehrer, B Konacher and MV Srinivasan for generously contributing unpublished data.
479
480
Sensory systems
References and recommended reading Papers of particular interest, published within the annual period of review, have been highlighted as: • of special interest •of outstanding interest
Equivalence between size and luminance differences in the context of honeybee vision.] 19.
Horridge GA, Zhang SW: Pattern vision of bees: flower like patterns with no predominant orientation. J Insect Physiol 1995, in press.
20.
Ronacher B: Beitrag einzelner Parameter zum wahmehmungs~ gem~ll~en Unterschied yon zusammengesetzten Reizen hei der Honigblene. Biol Cybem 1979, 32:77-83. [Title translation: Contributions of individual parameters to the perceptual discrimination of modular (composed) stimuli.]
1.
Von Frisch K: Der Farhensinn und Formensinn der Bienen. Zool Jb Abt allg Zool Physiol 1915, 35:1-182. [Title translation: The senses of color and form in the honeybee.]
2.
Hertz M: Die Organisation des optischen feldes hei der Biene I. Z vergl Physiol 1929, 8:693-748. [Title translation: Organization of the optical field in the honeybee.]
21.
Hertz M: Uber figurale Intensit~iten und Qualit~ten in der optischen Wahrnehmung der Biene. Biol Zbl 1933, 53:10-40. [Title translation: On figural intensity and quality in the optical perception of the honeybee.]
Ronacher B: Pattern recognition in honeybees: multidimensional scaling reveals a city-block metric, vision Res 1992, 32:1837-1843.
22.
Wehner R: Spatial vision in arthroflods. In Handbook of Sensory Physiology, vol VII/6C: Vision in Invertebrates. Edited by Autrum H. Berlin: Springer; 1981:287-616.
23.
Wehner R, Flatt h Visual fixation in freely flying bees. Z Naturforsch 1977, 32:469-471.
24.
Gould JL: How do bees remember flower shape? Science 1985, 227:1492-1494.
25.
Anderson AM: Shape perception in the honey bee. Anim Behav 1977, 25:67-79.
26.
Cartwright BA, Collett TS: How honey bees know their distance from a near-by visual mark. J Comp Physiol 1979, 161:335-355.
3.
4.
5.
6.
Zerrahn G: Formdressur und Formunlerscheidung hei der Honigbiene. Z vergl Physiol 1933, 20:117-150. [Title translation: Form conditioning and form discrimination in the honeybee.] Hassenstein B, Reichardt W: Systemtheoretische Analyse der Zeit-, Reihenfolgen- und Vorzeichenauswertung bei der Bewegungsperzeption des Ribselldifers Chlorophanus. Z Natufforsch 1956, 116:513-524. [Title translation: Theoretical systems analysis of the evaluation of time, sequence and sign of motion perception in the beetle Chlorophanos.] EgelhaafM, Borst A: Movement detection in arthropods. In
Reviews of Oculomotor Research, vol 5: Visual Motion and Its Role in the Stabilisation of Gaze. Edited by Miles FA, Wallman J. Amsterdam: Elsevier; 1993:53-77.
27.
Cartwright BA, Collett TS: Landmark learning in bees. ] Comp Physiol 1983, 151:521-543.
28.
7.
Land MF: Head movement of flies during visually guided flight. Nature 1973, 243:299-300.
Cartwright BA, Collett TS: Landmark maps for honeybees. Biol Cybern 1987, 57:85-93.
29.
8.
Collett TS, Land MF: Visual control of flight behaviour in the hoverfly, Syritta pipiens. J Comp Physiol 1975, 99:1-66.
Wehner R, R~ber F: Visual spatial memory in desert ants, Calaglyphis bicolor (Hymenoptera, Formicidae). Experimentia 1979, 35:1569-1571.
9.
Heisenberg M, Wolf R: On the fine structure of yaw torque in visual flight orientation of Drosophila melanoBaster. J Cornp Physiol 1979, 130:113-130.
30.
Wolf R, Heisenberg M: Basic organisation of operant behaviour as revealed in Drosophila flight orientation. J Comp Physiol A 1991, 169:699-705.
10.
Heisenberg M, Wolf R: Vision in Drosophila. Berlin: Springer Verlag; 1984. [Braitenberg V (Ed): Studies of Brain Function, vol 12.]
31.
11.
Heisenberg M, Wolf R: The sensory-motor link in motiondependent flight control of flies. In Reviews of Oculomotor Research, vol 5: visual Motion and Its Role in the Stabilisation of Gaze. Edited by Miles FA, Wallman J. Amsterdam: Elsevier; 1993:265-283.
Dill M: Unterscheidung und Wiedererkennung visuelier Muster bet der Tanfliege Drosophila melanogaster [PhD thesis]. W0rzburg: Universit~it W(Jrzburg; 1995. [Title translation: Discrimination and recognition of visual patterns in the fly Drosophila melanogaster.]
32.
Dill M, Wolf R, Heisenberg M: Visual pattern recognition in Drosophila involves retinotopic matching. Nature 1993, 365:751-753.
33.
Dill M, Heisenberg M: Visual pattern memory without shape recognition. Philos Trans R So(: Lond [Biol] 1995, in press.
34.
HSIIdobler B: Canopy orientation: a new kind of orientation in ants. Science 1980, 210:86-88.
35.
Peckham GW, Peckham EG: On the instincts and habits of solitary wasps. Wisconsin Geol Nat I-list Survey 1898, 2:1-148.
12.
Wolf R, Heisenberg M: Visual orientation in Drnsophi/a is an operant behavior. Nature 1986, 323:154-156.
13.
Wehner R: Der Mecbanismus der optischen Winkelmessong hei der Biene (Apis mellMca). Zool Anz Supp 1969, 33:586-592. [Title translation: Mechanisms of measurement of angles in the honeybee pattern vision.l
14.
Wehner R: Dorsoventral asymmetry in the visual field of the bee, Apis mellifica. J Comp Physiol 1972, 77:256-277.
36.
15.
Wehner R: Pattern modulation and pattern detection in the visual systems of Hymenoptera. In Information Processing in the Visual System of Arthropods. Edited by Wehner R. Berlin: Springer; 1972:183-194.
Von Frisch, K: Tanzsprache und Orientierung der Bienen. Berlin: Springer; 1965. [Title translation: Dance language and orientation of the honeybee.]
37.
Zeil J: Orientation flights of solitary wasps (Cerceri~ Sphecldae; Hymenoptera) I. Description of flight. J Comp Physiol A 1993, 172:189-205.
16.
Cruse H: Versuch einer quantitativen Besc'hreibun8 des Formensehens der Honighiene. Kybernetik 1972, 11:185-200. [Title translation: On the quantitative description of form perception in the honeybee.]
38.
Zeil J: Orientation flights of solitary wasps (Cereeris; Sphecidae; Hymenoptera) II. Similarities helween orientation and return flights and the use of motion parallax. J Comp Physiol A 1993, 172:207-222.
17.
Wehner R: Pattern recognition. In The Compound Eye and vision of Insects. Edited by Horridge GA. Oxford: Clarendon Press; 1975:75-114.
39.
18.
Ronacher B: ,~quivalenz zwischen Grid, n- und Helligkeitsunterschieden im Rahmen der visoellen Wahrnehmung der Honigblene. Biol Cybern 1979, 32:63-75. [Title translation:
Lehrer M: Locomotion does more than bring the bee to new places. In The Behaviour and Physiology of Bees. Edited by Goodman LJ, Fisher RC. Wallingford: C.A.B. International; 1991:185-202.
40.
Lehrer M: Why do bees turn back and look? J Comp Physiol A 1993, 172:549-563.
Pattern recognition in insects Heisenberg 481 41.
Collett TS, Lehrer M: Looking and learning: a spatial pattern in the orientation flight of the wasp Vespula vulgaris. Proc R Soc Lond [Biol] 1993, 252:129-134.
42. •"
Br0nnert U, Kelber A, Zeil j: Ground-nesting bees determine the location of their nests relative to a landmark by other than angular size cues. J Comp Physiol A 1994, 175:363-369. In two earlier papers [37,38], Zeil provided the first thorough account of orientation flights ('turn-back-and-look', [40]). In the tradition of Coilett and Land (see [8]), he presented a detailed frame-by-frame film analysis showing that wasps move laterally in semi-circles of increasing diameter around their nest entrance and the closest landmarks. He also found that a wasp's flight path at the next arrival is sufficiently similar to the previous orientation flight that one could imagine that the animal is storing the 'motion signature' of the relevant objects during that manoeuvre. In this present study [42ee], Zeil and co-workers demonstrate that wild bees estimate the distance of their nest entrance relative to a familiar landmark largely independently of the angular size at which the landmark appears from that position. The result does not directly prove, but supports, the idea that distance might be measured by motion parallax. 43. "
Lehrer M, Collett TS: Approaching and departing bees learn different cues to the distance of a landmark. J Comp Physiol A 1994, 175:171-177. It is during orientation flights that honeybees learn about the distance of landmarks from the goal. During arrival, they memorize only their angular size. This is an excellent demonstration that memories are attached to specific behavioral operations. 44. •
Collett TS, Baron l: Biological compasses and the coordinate frame of landmark memories in honeybees. Nature 1994, 368:137-140. The bees studied by the authors in Sussex, UK consistently approach a feeding site from the north, using the sun compass and the earth's magnetic field. This is a remarkable finding underscoring the general notion that in many situations a behavioral activity dramatically facilitates the computational task of the brain. 45.
56.
Kaiser W: The relationship between movement detection and colour vision in insects. In The Compound Eye and Vision of Insects. Edited by Horridge GA. Oxford: Clarendon Press;
1975:359-377. 57.
58.
Zhang SW, Srinivasan MV: Prior experience enhances paltem discrimination in insect vision. Nature 1994, 368:330-332. his is an especially beautiful example of visual 'priming'. A memory template of a pattern induces the bee to attend to motion parallax cues revealing the same pattern. In other words, the matching memory template pushes the weak motion parallax cues above threshold (see also
151,53]). 59.
Lehrer M, Srinivasan MV, Zhang SW, Horridge GA: Motion cues provide the bee's visual world with a third dimension. Nature 1988, 332:356-357.
60.
Srinivasan MV, Lehrer M, Horridge GA: Visual figure-ground discrimination in the honeybee: the role of motion parallax at boundaries. Proc R Soc Lond [Biol] 1990, 238:331-350.
61.
Lehrer M, Srinivasan MV: Object detection by honeybees: why do they land on edges. J Comp Physiol A 1993, 173:23-32.
Collett TS: Invertebrate vision: bees learn how to look. Curr Biol 1994, 4:717-719. brief review recounts several telling examples in which insects solve certain tasks by activating the specific behavioral operation that provides them with the needed visual input. 62. his
63. ""
Collett TS, Fry SN, Wehner R: Sequence learning by honeybees.
64.
Fischbach KF, Dittrich APM: The optic lobe of Drosophila melanogaster. I. A Golgi analysis of wild-type structure. Cell Tissue Res 1989, 258:441-475.
65.
O'Carroll D: Feature-detecting neurons in dragonflies. Nature 1993, 362:541-543.
66.
Eyding D: Lemen und Kurzzeltged~chtnis belm operanten Konditionleren auf visuelle Muster bei strukturellen und biochemischen Lernmutanten yon Drosophila rnelanogaster [Diplom thesis]. WLirzburg: Universit~t WLirzburg; 1994. [Title translation: Operant learning and short-term memory for visual patterns in structural and biochemical learning mutants of
67.
Wittig T, Dill M, Wolf R, Heisenberg M: Visual pattern discrimination learning in Drosophila melanogaster without mushroom bodies. In Learning and Memory. Proceedings of the 23rd G~ttingen Neurobiology Conference, vol I. Edited by Eisner N, Menzel R. Stuttgart: Georg Thieme Verlag; 1995:4.
68.
Tinbergen N: The Study of Instinct. Oxford: Clarendon Press; 1951.
J Comp Physiol A 1993, 172:693-706. 48.
Yon Helmholtz H: Handbuch der physiologischen Optik. Leipzig: L. Voss; 1867. [Title translation: Handbook of physiological optics.[
49.
Collett TS, Land MF: Visual spatial memory in a hoverfly. J Comp Physiol 1975, 100:59-84.
50.
Wolf R, Heisenberg M: On the fine structure of yaw torque in visual flight orientation of Drosophila melanogaster. II. A temporally and spatially variable weighting function for the visual field ('visual attention'). J Comp Physiol 1980, 140:69-80.
51.
Van Hateren IH, Srinivasan MV, Wait PB: Pattern recognition in bees: orientation discrimination. J Comp Physiol A 1990, 167:649-654.
52.
Wehner R: The generalization of directional visual stimuli in the honey bee, Apis mellifera. J Insect Physiol 1971, 17:1579-1591.
53. 54.
Heisenberg M, Buchner E: The role of retinola cell types in visual behavior of Drosophila melanogaster. J Comp Physiol 1977, 117:127-162.
Cheng K, Collett TS, Pickhard A, Wehner R: The use of visual landmarks by honeybees: bees weight landmarks according to their distance from the goal. J Comp Physiol A 1987, 161:469-475.
Georg Thieme Verlag; 1995:31.
47.
Kaiser W, Liske E: Die optomotorischen Reaktionen von fixiert fliegenden Bienen bei Reizung mit Speldrallichtern. J Comp Physiol 1974, 89:391-408. [Title translation: Optomotor responses of fixed flying bees stimulated with spectral lights.]
Zhang SW, Srinivasan MV, Collett T: Convergent processing in honeybee vision: multiple channels for the recognition of shape, proc Nail Acad Sci USA 1995, 92:3029-3031. The site of pattern recognition in the brain, that is, the memory template that is deposited and later activated in the matching process, can be reached by at least three independent input routes defined by the stimulus conditions: isoluminant color contrast, monochromatic luminance contrast and motion parallax contrast between figure and ground. Studies like this may eventually help to pin down the neuronal correlate of the memory template.
Menzel R, Joerges J, M011er U: Computational phenomena in long distance orientation of honey bees. In Learning and
Memory. Proceedings of the 23rd G~ttingen Neurobiology Conference, vol I. Edited by Eisner N, Menzel R. Stuttgart: 46.
55.
Srinivasan MV, Zhang SW, Rolfe B: Is pattern vision in insects mediated by 'cortical' processing? Nature 1993, 362:539-540. Menzel R: Spectral response of movement detecting and 'sustaining' fibres in the optic lobe of the bee. J Comp Physiol 1973, 82:135-150.
Drosophila melanogaster.]
M Heisenberg, Theodor-Boveri-lnstitut fiir Biowissenschaften, Lehrstuhl ftir Genetik, University of Wiirzburg, Am Hubland, D-97074 Wiirzburg, Germany. E-mail: [email protected]