- Email: [email protected]
Biases in signal evolution: learning makes a difference
Review
TRENDS in Ecology and Evolution
Vol.22 No.7
Biases in signal evolution: learning makes a difference Carel ten Cate1 and Candy Rowe2 1
Behavioural Biology, Institute of Biology, Leiden University, PO Box 9516, 2300 RA Leiden, The Netherlands Centre for Behaviour & Evolution, Division of Psychology, Newcastle University, Henry Wellcome Building, Framlington Place, Newcastle upon Tyne, NE2 4HH, UK
2
It is now well established that signal receivers have a key role in the evolution of animal communication: the suite of sensory and cognitive processes by which animals perceive and learn about their environment can have a significant impact on signal design. A crucial property of these information-processing mechanisms is the emergence of ‘receiver bias’ in the behavioural responses to signals. Whereas most research has focussed on receiver biases in the sensory system, more recent studies show that biases can also arise from learning about signals. Here, we highlight how learning-based biases can arise, and how these differ from biases emerging from sensory systems in their impact on signal evolution. Signal evolution by receiver biases A commonly held view is that signal evolution progresses through selection for costly signals that provide receivers with reliable information [1–4]. However, over recent years, this position has been eroded by studies showing that signal evolution can be driven by ‘receiver biases’ [5–10]. Receiver biases can be defined as byproducts of natural selection or as incidental non-selected consequences of the way in which sensory systems or brains are formed [4], predisposing receivers to respond more strongly to one signal than to another and even to prefer novel signals [9,10]. Here, we demonstrate that different types of receiver bias operate under different conditions and can have different implications for signal evolution. We focus on two opposite corners of what can be viewed as a continuous space determined by two different dimensions of the proximate mechanisms underlying receiver biases. The first dimension is whether biases result from the peripheral physiology of sensory systems or from higher cognitive brain processes. The second dimension is developmental phenotypic plasticity: some biases develop in the same way over a wide range of conditions, whereas others depend heavily on specific experience. We differentiate here between ‘sensory system biases’ (see Glossary), which we define as receiver biases arising from more peripheral and primary sensory processing mechanisms and also showing limited if any phenotypic plasticity, and ‘learning-based biases’, which we define as arising from central information processing involving plasticity generated by learning. In making this distinction, Corresponding author: ten Cate, C. ([email protected]). Available online 26 March 2007. www.sciencedirect.com
we create a false dichotomy within a continuum. However, we do so to illustrate how different proximate mechanisms can differ in their impact on signal evolution. Sensory system and learning-based biases Perhaps the most well known and widely cited example of a signal that might have evolved owing to a sensory system bias is the ‘chuck’ component of the male mating call in the Tu´ngara frog Physalaemus pustulosus. Results from comparative studies suggest that, before the chuck signal evolved, ancestral female frogs already had a bias in their auditory system to respond to particular sound features, which drove subsequent call evolution [5,6]. Other examples of traits that are likely to have evolved by exploiting sensory system biases are the swords of swordtail Xiphophorus fish [7,8], the vibratory signals of the water mite Neumania papillator [11,12], the red coloration of the male three-spined stickleback Gasterosteus aculeatus [13] and the blue and yellow coloration of two Lake Victoria cichlid species Pundamilia pundamilia
Glossary Area shift: a phenomenon similar to peak shift where the peak response remains highest for the training stimuli but the generalization gradient is asymmetric, resulting in higher responses to novel stimuli away from the S+ in the direction opposite from the S and vice versa. Discrimination learning: the process by which animals learn to discriminate among stimuli, either along a single stimulus dimension or using more complex multidimensional features. Generalization gradient: the degree to which an animal trained to respond to a particular stimulus will respond to novel stimuli that vary along a stimulus dimension shared with the training stimulus. Intensity dimension: stimuli differ along an intensity dimension when they stimulate the same receptors but to a different extent; for instance, two lights of the same wavelength differing in intensity; chemical substances differing only in concentration; or tones differing only in amplitude. Learning-based bias: a receiver bias that is generated through learning; for example, peak shift. Peak shift: a consequence of discrimination learning between an S+ (a positively rewarded stimulus) and an S (a negatively or neutrally rewarded stimulus) that differ along a stimulus dimension, leading to stronger responding to novel stimuli away from the S+ in a direction opposite from the S , and vice versa (Box 1). Rearrangement dimension: stimuli differ along a ‘rearrangement’ dimension when each stimulus addresses a different set of receptors, such as lights differing in wavelength; chemicals of different structure; or tones of different frequency. Whereas an intensity dimension can be considered a quantitative dimension, the rearrangement dimension is more qualitative. Sensory system bias: a receiver bias arising from peripheral sensory processing. Although this need not imply absence of environmental plasticity, we assume here that it does. Stimulus dimension: any aspect of a stimulus that can vary along an axis; for example, light intensity or angle of orientation.
0169-5347/$ – see front matter ß 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.tree.2007.03.006
Review
TRENDS in Ecology and Evolution
and Pundamilia nyererei [14]. These biases are assumed either to be byproducts of the sensory systems involved [7] or to arise in these systems as a result of selection for signal detection in other contexts. For example, the vibratory water movements in the water mite might have arisen from exploiting predator detection mechanisms [11,12], the red coloration of stickleback males might result from selection for preferences for red food items [13], and the blue and yellow colour in cichlids might be a consequence of a differential sensitivity to long and short wavelengths owing to adaptation to different ambient light conditions in both species [14]. Although we do not know whether these biases lack phenotypic plasticity, this is often implied from the finding that related species share similar biases and that the processing is peripheral. Here, we assume that they have limited plasticity. Learning-based biases arise from the cognitive mechanisms that process and store information. There is an increasing awareness that biases based on cognitive processes could be just as important for signal evolution as are sensory system biases [15–25]. Not all biases arising from cognitive processing are necessarily plastic. However, most interesting from our perspective are biases that arise from the way in which animals learn to recognize important environmental features. One mechanism that can generate a learning-based bias is that of ‘peak shift’, which emerges from learning to discriminate among stimuli [26,27]. Using peak shift as an example, we highlight the need to consider sensory system and learning-based biases as distinct selection pressures in signal evolution. Learning and generalization In many contexts, responding to signals involves learning. Avian predators, for instance, avoid prey with aposematic warning signals by learning to associate the coloration with the defences (e.g. Ref. [28]), songbirds learn to recognize the songs of conspecifics or local geographical variants of songs (e.g. Ref. [29]), and bumble bees learn to recognize the colour of the most profitable flowers [30]. Recognition involves remembering particular signals, but also discriminating them from other similar signals. Thus, predators can learn to discriminate between undefended and defended prey using visual patterns, songbirds learn to ignore the songs of other sympatric bird species, and bumble bees learn to avoid the colours of less profitable flowers. This discrimination learning process also determines how animals respond to novel stimuli; that is, how they generalize (e.g. Refs [26,27,31–33]). Generalization gradients resulting from discrimination learning have been extensively studied by psychologists in experimental laboratory settings, often using simple artificial stimuli, such as tilted lines, lights of different wavelengths or tones of different frequency (e.g. Refs [26,27,31–34]). In these experiments, animals are usually trained to discriminate between two stimuli that are differentially reinforced, where one stimulus indicates the presence of a reward (S+), and the other stimulus is neutral or indicates some punishment (S ). After training, animals are presented with novel stimuli that differ from their training stimuli to varying degrees, and their responses are measured. Although we might expect that the strongest www.sciencedirect.com
Vol.22 No.7
381
responses are seen towards the training stimuli, this is not always the case, and the strongest response might be given to stimuli that are more extreme on the dimension separating the training stimuli (Box 1). This shift in generalization gradient is known as a ‘peak shift’ because the peak response is shifted along the training dimension away from the original stimuli used in training. The shape of this generalization gradient differs depending on whether the training stimuli vary along an ‘intensity’ dimension (e.g. two lights of the same wavelength differing in intensity) or qualitatively along a ‘rearrangement’ dimension (e.g. two lights of the same intensity but of a different wavelength) (Box 1). A less extreme version of peak shift is ‘area shift’, where, although the peak response remains highest for the training stimuli, the generalization gradient is skewed, with novel stimuli on the S side of S+ receiving fewer responses than S+ and novel stimuli on the other side of S+ getting similar responses as S+ [34,35]. Apart from peak and area shift, other types of generalization processes can also bring about preferences for novel stimuli, such as ‘range effects’ (e.g. Ref. [34]), in which the location of the peak in responding depends on the range of stimuli used in testing. However, the relevance of these processes to signal evolution is still unclear. Peak shift is an example of a learning-based bias that can drive signal evolution. It is taxonomically widespread, occurring in vertebrates and invertebrates, and appears to be a general property of discrimination learning. Peak shift is also found for a range of different stimuli in different sensory modalities (e.g. vision, hearing or olfaction), and along many perceptual dimensions (e.g. such as the frequency, duration or amplitude of a sound signal). It is a fundamental property of discrimination learning; therefore, whenever animals learn to discriminate among closely similar signals, whether prey coloration, conspecific song, or floral colours, they might develop biases that can select for more extreme signals. However, does peak shift occur in more biologically relevant contexts similar to the ones the animal has to solve in every day life, or with more complex stimuli compared with the tilted lines or coloured keys found in traditional experimental psychology experiments? Peak shift in the natural world? There is increasing evidence for peak shifts in a more natural context. One example of this concerns a task that humans perform in daily life: face recognition. Humans show peak shift when they have to distinguish similarly looking faces [35]. This is also thought to give rise to the phenomenon where we are better at recognizing familiar individuals from caricatures exaggerating specific features than from drawings based on real facial features [36]. For animals, several recent studies have also used complex stimuli that either are close to natural ones or are biologically relevant, for instance using a learning task that is more similar to one they are likely to experience in the wild. Peak shift has, for example, been observed in the spatial orientation of pigeons Columbia livia and honey bees Apis mellifera in relation to artificial ‘landmarks’ [37,38], while sphinx moths Manduca sexta, which use
382
Review
TRENDS in Ecology and Evolution Vol.22 No.7
Box 1. The mechanism of peak shift There are three main theoretical hypotheses to explain the apparently counterintuitive phenomenon of peak shift. It was originally thought that animals learn the relationship between the positively rewarded stimulus (S+) and the negatively rewarded or neutral one (S ), rather than the absolute value of each. According to this ‘relational hypothesis’, animals might learn a rule such as ‘choose the brighter stimulus’. This might make them prefer a more extreme version of a rewarded stimulus when given the choice between the rewarded stimulus and the more extreme version. However, most experimental findings cannot be explained in this way [31,34,59,60]. An alternative hypothesis suggested that peak shift results from overlap between the generalization gradients around S+ and S . The interaction between the excitatory gradient around S+ and the inhibitory gradient around S could result in a combined gradient in which the strongest response is not given to S+, but is displaced, away from S [61]. This ‘conditioning-extinction theory’ has been propagated in the animal communication literature [15,40], but its underlying assumptions are not supported [31]. Finally, the ‘elemental’ or ‘receptor’ account proposes that the amount of feature sharing between novel and training stimuli will determine the response to novel stimuli [27,37,59,62]. If S+ and S
share some features, then stronger responses to novel than to training stimuli arise when novel stimuli share features of S+ that are not present in S . This can also explain the difference in generalization gradients between stimuli differing along ‘intensity’ and ‘rearrangement’ dimensions [19,27]. Stimuli differ along an intensity dimension when they stimulate the same receptors but to a different extent; for instance, two lights of the same wavelength differing in intensity, chemical substances differing only in concentration or tones differing only in amplitude. In these cases, the generalization gradient is often steep and, to some extent, monotonic (Figure I), indicating that the response to novel, more extreme, stimuli is sometimes considerable stronger than to the training stimuli. When the training stimuli differ along a ‘rearrangement’ dimension, each stimulus addresses a different set of receptors, such as lights differing in wavelength, chemicals of different structure or tones of different frequency. In these cases, the generalization gradient usually peaks close to the training stimulus and the peak response is not much stronger than when responding to familiar stimuli (Figure I; reviewed in Ref. [27]). Apart from the dimension on which the stimuli differ, another parameter that affects the shape and skew of the generalization is how close together S+ and S are; the closer they are, the stronger the shift [27].
Figure I. Idealized generalization gradients resulting from discrimination learning. The X-axis represents a range of stimuli differing on one particular dimension. When animals have been trained to respond differentially to two stimuli, one providing a positive (S+) and one a negative reward (S ), one would expect the animal to respond most strongly to training stimuli compared with novel stimuli (i.e. according to the grey line). However, this is only the case when S+ and S differ relatively strongly. If the stimuli are similar, the generalization gradient will show a peak shift (arrows) and be similar to the red gradient (for stimuli differing along a rearrangement dimension) or to the blue gradient (for stimuli differing along an intensity dimension) [27]. The vertical solid lines indicate the location of the S+ and S and of the peaks for a generalization gradient without peak shift. The dashed lines indicate the location of the peaks for a generalization gradient involving peak shift along a rearrangement dimension.
odours to find food, show area shift if conditioned with odours differing in the length of the carbon chain or the functional groups (alcohol or ketone) [39]. Peak shift has been proposed as driving force behind signal evolution [17], in particular for the contexts of food preference learning [15] and sexual communication [40]; indeed, the best evidence for peak shift in natural systems comes from these two areas. Many animals need to learn which food types are profitable sources of nutrients and which are less profitable, either in terms of handling costs www.sciencedirect.com
or energetic content. For example, bumble bees Bombus terrestris exhibit peak-shifted colour preferences following training on artificial ‘flower’ colour discrimination and such changes are suggested to drive the evolution of floral species that have orchids as batesian mimics [30]. Domestic chicks Gallus domesticus trained to discriminate between palatable and aposematic insect prey subsequently show a biased generalization against prey that have stronger or larger warning signals [41,42]. This suggests that peak shift might drive the evolution of
Review
TRENDS in Ecology and Evolution
aposematic signals, a suggestion supported by theoretical models [43–45]. Two recent studies suggest that peak shift can be responsible for the evolution of exaggerated sexual traits. One of these used artificially constructed zebra finch Taeniopygia guttata songs as discriminative stimuli. It showed that perceptual learning about songs varying along an intensity dimension (i.e. number of elements in a song) and songs varying along a rearrangement dimension (i.e. position of an odd element in a series of repeated elements) resulted in differently skewed generalization patterns, similar to what can be observed in discrimination studies using simple stimuli. Peak shift was present for the responses to novel songs with higher or lower numbers of elements than the training stimuli (i.e. along the intensity gradient) [46]. This finding suggests that it is easier to get song evolution driven by a receiver bias along an intensity
Vol.22 No.7
383
dimension than along a rearrangement dimension. The second study showed that male zebra finches developed peak-shifted sexual preferences for beak colouration through the learning process of sexual imprinting (Box 2; [21]). Crucially, these examples show that the findings obtained in highly simplified experimental conditions appear to hold under more complex situations involving tasks closer to the ones the animals are likely to encounter in nature. As there are many natural contexts in which animals learn about signals, and because peak shift is a widespread emergent property of discrimination learning, other cases are likely to be common. For instance, the example presented in Box 2 suggests that peak shift can also be found in other contexts involving learning about features of conspecifics or closely similar species, such as when two highly similar species overlap in their
Box 2. Super mates: peak shift and exaggerated sexual signals Many birds [63], as well as some mammals [64] and fish [65], acquire their mating preferences by learning about their parents and taking these as a model for their later sexual preference, a process known as ‘sexual imprinting’. At the same time, sexual preferences are often skewed and can be combined with sexual dimorphism, resulting in preferences for mates with exaggerated traits. Could such skewed mating preferences be a side effect of imprinting, owing to a peak shift? If offspring are raised by a sexually dimorphic species with biparental care, the presence of two parents differing in appearance and behaviour provides a context for discrimination learning. Whether this can induce peak-shifted mate preferences [40] was tested in an experiment in which young zebra finch Taeniopygia guttata males were raised by white zebra finch parents that differed in beak colour only (Figure I) [21]. In one experimental group, the beak of the father was painted red and that of the mother orange using nail varnish. In a second group, the reverse occurred. In earlier studies [66,67], males raised in this way went on to prefer females that were similar in appearance to the mother, suggesting that mothers act as
an ‘S+’ and fathers as an ‘S ’. In the current experiment, the males could choose between eight females differing in beak colour. Both groups preferred females with a beak in a more extreme colour than their mother (i.e. more intense red when the mother’s beak was red and more yellow orange when the mother’s beak colour was orange). Hence, the parental beak colouration induced directional colour biases in the mate preference of male offspring. The experiment demonstrates that peak shift can arise from a naturally occurring and crucial learning process, such as sexual imprinting. Although the evolution of sexual dimorphism cannot be reconstructed with such an experiment, it does show how a slight initial sexual dimorphism can give rise to the emergence of a bias for novel signals that exaggerate the dimorphism. It is the difference between the two stimuli involved in the learning process that determines both the dimension of the preference (beak colour in this case), as well as the direction of the bias (more extreme than the mother’s beak). Had the parents not differed in beak colour, but along another dimension (e.g. some plumage characteristic), then the bias would have been different.
Figure I. Peak shift in the sexual preferences of zebra finch males. Arrows indicate the parental beak colours (mother orange, father red for one group; mother red, father orange for the other). White bars and grey bars indicate the preferences (median and interquartile ranges) for females having different beak colours. Males show stronger preferences for females with a more extreme beak colour compared with that of their mother. Modified with permission from Ref. [21]; drawing of finches by Machteld Verzijden. The beak colours in the figure provide an indication, but not an exact match, of the colours used in the experiment. www.sciencedirect.com
384
Review
TRENDS in Ecology and Evolution Vol.22 No.7
geographical range. In such a situation, one can imagine that learning to discriminate between the two species could induce a peak-shifted recognition of conspecifics that might drive character displacement as an emergent property of the learning process (Box 3). This and other contexts might provide fruitful areas for further exploration. The same holds for the question of whether generalization gradients are subject to selection [23,30, 33]. Bumble bees, for instance, respond with an adaptive peak shift to changes in payoffs of selecting different stimuli [30], indicating that the shape of generalization gradients is not always the same, but varies depending on the consequences of making wrong discriminations or the current payoffs during learning [21,26]. It is possible that selection operates on such plasticity and that the fitness consequences of the generalization pattern affect the speed and direction of signal evolution.
Comparing sensory system and learning-based biases Sensory system and learning-based biases are similar in that they both predispose an animal to be most responsive to signals that have not been experienced before and that might guide subsequent evolution. But to what extent do learning-based biases differ from those arising in sensory systems? Sensory system biases have their origins in other contexts and exist before the evolution of the signal. For example, selection for a preference for a particular food colour, as demonstrated in several species [13,47], could lead to a sensory system with a general, non-learned sensitivity to that type of stimulation. The more this bias depends on peripheral sensory physiology or processing, the more likely it is that the bias shows limited, if any, developmental plasticity. Hence, it will be similar from one generation to the next. If, in a mating context, a potential
Box 3. Peak shift, species discrimination and character displacement Sympatric species usually exhibit greater species mating discrimination than do allopatric species. This is often ascribed to the evolutionary process of reinforcement; the adaptive strengthening of mate discrimination to avoid detrimental hybridization (e.g. Ref. [68]), which might be the outcome of several mechanisms selecting for genetically based preferences for particular types of mates (e.g. Refs [18,69]). Although it is clear that changes in discriminative ability can arise in this way, incorporating learning as a mechanism involved in species recognition suggests an alternative scenario in which the increase in discriminative ability arises as an emergent byproduct. Adult learning about conspecific and heterospecific traits is of widespread importance in birds and can result in stronger species discrimination in sympatric areas (reviewed in Ref. [70]). Such learning can provide a mechanism for a flexible adaptation to the current situation [70]. We suggest that skewed generalization makes the impact of learning greater than already anticipated. Suppose we are dealing with two allopatric, closely similar bird species (Figure I), differing in wing colour only. Each allopatric species learns the characteristics of it own species, resulting in an
average preference for mates with an appearance around the population mean (Figure Ia). When the species suddenly become sympatric locally (e.g. by crossing a geographical barrier), individuals might discover that they sometimes interact with conspecifics and sometimes with birds that are responding in an unusual way and also look slightly different: the other species. This context can result in discrimination learning that might give rise to biases in which each species will preferentially respond to those conspecifics that are more distinct from the other species (Figure Ib). The presence of sympatry will itself thus give rise to greater species mating discrimination in sympatry than in allopatry as an emergent property. One can imagine that this might drive character displacement. A particular feature is that the effect will only be induced and, hence, present in those parts of the population where the two species overlap. The exact dynamics of the process will be different depending upon whether the discrimination operates along a rearrangement dimension, leading to a limited peak shift (as in Figure I), or a dimension based upon intensity, which may give rise to more open-ended and stronger preference shifts [27].
Figure I. Hypothetical representation of traits and learned preferences of two closely similar species in allopatry (a) and sympatry (b). In allopatry, the trait (signal) distribution (solid line) and the recognition of this trait (hatched line) show a close match. In sympatry, mutual interactions will lead to a peak shift of the recognition gradients away from the actual traits, as indicated by the arrows, providing scope for selection on traits to become more distinct. www.sciencedirect.com
Review
TRENDS in Ecology and Evolution
partner provides a signal that tunes in to the bias, this results in selection for that novel signal and subsequent signal evolution. Evolutionary changes in the bias will primarily be the consequence of changes in the selection pressures in the context that gave rise to the bias. In this case, a new food type of a different colour could generate selection for new signals of that colour. The nature or direction of this bias will only be affected by the current signals if the colour preference in the signalling context becomes independent of a general perceptual visual preference. By contrast, biases through peak shift arise within the specific signalling context. Whereas generalization is a general property of the nervous system of animals, learning about a signal is not all that is required for a skewed generalization gradient to emerge. It also requires that two or more signals are closely related, otherwise, learning will not give rise to skewed generalization in the first place: the more similar two stimuli are, the more exaggerated the biases will be. In a signalling context, this could lead to stronger selection pressures acting on signals. To provide a consistent directional selection, a high degree of similarity in the learning experiences of individuals within a population is also needed. Such a shared bias that would also be stable over generations might arise, for instance, when predatory species all learn to discriminate among the same colour patterns of edible and non-edible prey items. Peak shift is a function of the discrimination process: its nature (i.e. a monotonic open-ended preference when stimuli differ along an intensity dimension, or a peaked one with a rearrangement dimension), direction and strength all depend upon the current contrast between the stimuli and, hence, will be plastic. As the direction of the signal evolution will depend upon the signalling niche in which it evolves, individuals that differ in their experiences will also develop different learning-based biases, unlike with sensory biases. This could lead to receivers selecting to diversify signals across populations rather than to drive them in the same direction, which could lead to speciation between different populations. Because the bias is generated by the current contrast between stimuli or signals, it might be less likely to induce novel traits. As the bias will only arise from discrimination learning involving closely similar stimuli, it is not something that, on its own, can generate large differences. However it might be important in the initial process of exaggerating small differences before other selection processes take over. Finally, if the signal changes as a consequence of selection, the bias will follow. Hence, the bias is not only contingent upon the present signal contrast, but also evolves in concert with this contrast. This plasticity can enable rapid evolutionary change. Evolutionary implications Thus, there are several differences in the origin and plasticity of sensory system and learning-based biases, but why does this distinction matter? The plasticity over time of learning-based biases is key to their function and, given their widespread occurrence, they could influence many signalling scenarios. Many studies have shown that incorporating the plasticity that learning mechanisms www.sciencedirect.com
Vol.22 No.7
385
provide in evolutionary models can alter the dynamics of evolutionary processes such as speciation and sexual selection compared with when only genetic mechanisms are considered [48–55]. Nevertheless, most evolutionary models that do incorporate learning in recognition systems assume either that learned preferences match the stimulus about which the learning occurred, or invoke a non-learned bias to account for or induce changes in preferences [48,51–55]. Incorporating skewed generalization will add a mechanism that might be responsible for driving evolutionary changes in many contexts involving learning, enabling rapid and diversifying changes in signal characteristics that are less likely to be accomplished by sensory system biases. Demonstrating trait evolution by learning-based biases A final point that we want to make concerns the research strategy assessing the potential impact of learning-based biases on signal evolution. This strategy is, to some extent, necessarily different from assessing the impact of sensory system biases. Studies on sensory exploitation, as in the Tu´ngara frogs [5,6], swordtails [7,8,56] and others [12,13,47], use comparative studies within a phylogenetic framework to demonstrate that preference for a trait was likely to be present in an ancestor of the current species, before emergence of the trait itself. For a preference emerging from the learning process, this approach has limited value. A characteristic of learning-based biases is that they are transient. As changing conditions between generations affect the nature, shape or even the presence of biases, they might be less likely to leave a phylogenetically traceable effect. But it is not impossible. If species A has a more elaborated signal than its sister species B, and if species B learns about the signal and shows a biased preference in the direction of a signal similar to A owing to peak shift, this suggests that the elaborated signal in species A is a consequence of selection driven by skewed generalization. Nevertheless, we might have to revert to other methods to ascertain whether certain traits or signals arose as a consequence of exploiting a generalization skew. One approach is to demonstrate that skewed generalization gradients arise in a context where one can hypothesize their role at present or in the past, and with stimuli of the relevant type. So, whereas the example in Box 2 does not demonstrate how the current sexual dimorphism in the zebra finch has arisen or whether it is still driven by peak shift, it does demonstrate that peak shift can occur after imprinting on a slightly sexually dimorphic trait. Hence, it illustrates the potential of peak shift to contribute to the differentiation of sexually selected traits. Another type of experiment indicating the evolutionary potential of a skew in generalization is one in which a trait is allowed to ‘evolve’ in response to behavioural choices made by an animal. An example of this is a series of experiments in which colours or other traits on a computer screen could change (‘evolve’) in response to choices made by domestic chickens trained to discriminate between them [57,58]. Such evidence can be supplemented with mathematical models [50] that test whether the presumed starting conditions, combined with characteristics of the learning process, result in reliable evolutionary changes.
386
Review
TRENDS in Ecology and Evolution Vol.22 No.7
Conclusions Biases can drive signal evolution, but it is clear that the origin and nature of a bias matters as to what, when and how signals evolve. With our example of peak shift, we have deliberately chosen an example of a learning-based bias that is different from a sensory system bias to highlight the contrasts. Not all cognitive biases will be like this, and not all cases of sensory biases will be as developmentally rigid as we presented them. However, by providing such a strong contrast, we hope to have demonstrated that receiver biases resulting from different proximate origins can result in different outcomes and can be relevant in different contexts. These matters are wide open for further research, and we indicated some areas that we consider profitable. In our view, we have only just begun to scratch the surface of how cognitive processes might affect the nature, direction and speed of signal evolution. But even so, we feel justified to conclude that learning does make a difference. Acknowledgements We thank Christina Halpin, Sue Healy, Rob Lachlan, Hans Slabbekoorn, Machteld Verzijden, Jeri Wright and three referees for helpful comments and/or discussion. C.t.C. was a Lorentz Fellow at NIAS (Wassenaar, NL) during the final revision.
References 1 Zahavi, A. (1975) Mate selection: a selection for a handicap. J. Theor. Biol. 53, 205–214 2 Grafen, A. (1990) Biological signals as handicaps. J. Theor. Biol. 144, 517–546 3 Andersson, M. (1994) Sexual Selection, Princeton University Press 4 Searcy, W.A. and Nowicki, S. (2005) The Evolution of Communication, Princeton University Press 5 Ryan, M.J. et al. (1990) Sexual selection for sensory exploitation in the frog Physalaemus pustulosus. Nature 343, 66–67 6 Ryan, M.J. and Rand, A.S. (1993) Sexual selection and signal evolution: the ghost of biases past. Philos. Trans. R. Soc. B 340, 187–195 7 Basolo, A.L. (1990) Female preference predates the evolution of the sword in female swordtail fish. Science 250, 808–810 8 Basolo, A.L. (1995) Phylogenetic evidence for the role of pre-existing bias in sexual selection. Proc. R. Soc. B 259, 307–311 9 Endler, J.A. and Basolo, A.L. (1998) Sensory ecology, receiver biases and sexual selection. Trends Ecol. Evol. 13, 415–420 10 Ryan, M.J. (1998) Sexual selection, receiver biases, and the evolution of sex differences. Science 281, 1999–2003 11 Proctor, H.C. (1991) Courtship in the water mite Neumania-papillator – males capitalize on female adaptations for predation. Anim. Behav. 42, 589–598 12 Proctor, H.C. (1993) Sensory exploitation and the evolution of male mating-behavior – a cladistic test using water mites (Acari, Parasitengona). Anim. Behav. 44, 745–752 13 Smith, C. et al. (2004) A receiver bias in the origin of three-spined stickleback mate choice. Proc. R. Soc. B 271, 949–955 14 Maan, M.E. et al. (2006) Sensory drive in cichlid speciation. Am. Nat. 167, 947–954 15 Guilford, T. and Dawkins, M.S. (1991) Receiver psychology and the evolution of animal signals. Anim. Behav. 42, 1–14 16 Guilford, T.C. and Dawkins, M.S. (1993) Receiver psychology and the design of animal signals. Trends Neurosci. 16, 430–436 17 Staddon, J.E.R. (1975) A note on the evolutionary significance of ‘supernormal’ stimuli. Am. Nat. 109, 541–545 18 Enquist, M. and Arak, A. (1998) Neural representation and the evolution of signal form. In Cognitive Ecology (Dukas, R., ed.), pp. 21–87, University of Chicago Press 19 Enquist, M. and Ghirlanda, S. (2005) Neural Networks and Animal Behavior, Princeton University Press 20 Rowe, C. and Skelhorn, J. (2004) Avian psychology and communication. Proc. R. Soc. B 271, 1435–1442 www.sciencedirect.com
21 ten Cate, C. et al. (2006) Sexual imprinting can induce sexual preferences for exaggerated parental traits. Curr. Biol. 16, 1128–1132 22 Lynn, S.K. (2005) Learning to avoid aposematic prey. Anim. Behav. 70, 1121–1226 23 Lynn, S.K. (2006) Cognition and evolution: learning and the evolution of sex traits. Curr. Biol. 16, R421–R423 24 Bateson, M. and Healy, S.D. (2005) Comparative evaluation and its implications for mate choice. Trends Ecol. Evol. 20, 659–664 25 Chittka, L. and Raine, N.E. (2006) Recognition of flowers by pollinators. Curr. Opin. Plant Biol. 9, 428–435 26 Purtle, R.B. (1973) Peak shift: a review. Psychol. Bull. 80, 408–421 27 Ghirlanda, S. and Enquist, M. (2003) A century of generalization. Anim. Behav. 66, 15–36 28 Guilford, T. (1990) Predator psychology and the evolution of prey coloration. In Natural Enemies: The Population Biology of Predators, Parasites and Diseases (Crawley, M.J., ed.), pp. 377–494, Blackwell Scientific 29 Riebel, K. (2003) The ‘mute’ sex revisited; vocal production and perception learning in female songbirds. Adv. Stud. Behav. 33, 49–86 30 Lynn, S.K. et al. (2005) Peak shift discrimination learning as a mechanism of signal evolution. Evolution 59, 1300–1305 31 Mackintosh, N.J. (1983) Conditioning and Associative Learning, Clarendon Press 32 Mackintosh, N.J. (1995) Categorization by people and pigeons: the twenty-second Bartlett memorial lecture. Quart. J. Exp. Psychol. B 48, 193–214 33 Cheng, K. (2002) Generalisation: mechanistic and functional explanations. Anim. Cogn. 5, 33–40 34 Thomas, D.R. et al. (1991) Peak shift revisited: a test of alternative interpretations. J. Exp. Psychol. 17, 130–140 35 Spetch, M.L. et al. (2004) Peak shift but not range effects in recognition of faces. Learn. Motiv. 35, 221–241 36 Lewis, M.B. and Johnston, R.A. (1999) Are caricatures special? Evidence of peak shift in face recognition. Eur. J. Cogn. Psychol. 11, 105–117 37 Cheng, K. et al. (1997) Spatial peak shift and generalization in pigeons. J. Exp. Psychol. Anim. Behav. Process. 23, 469–481 38 Cheng, K. et al. (2006) Small-scale spatial orientation in pigeons. Behav. Process. 72, 115–127 39 Daly, K.C. et al. (2001) The generalization of an olfactory-based conditioned response reveals unique but overlapping odour representations in the moth Manduca sexta. J. Exp. Biol. 204, 3085–3095 40 Weary, D.M. et al. (1993) A product of discrimination learning may lead to female preferences for elaborate males. Evolution 47, 333–336 41 Gamberale, G. and Tullberg, B.S. (1996) Evidence for a peak-shift in predator generalization among aposematic prey. Proc. R. Soc. B. 263, 1329–1334 42 Gamberale-Stille, G. and Tullberg, B.S. (1999) Experienced chicks show biased avoidance of stronger signals: an experiment with natural colour variation in live aposematic prey. Evol. Ecol. 13, 579– 589 43 Leimar, O. et al. (1986) Evolutionary stability of aposematic colouration and prey unprofitability: a theoretical analysis. Am. Nat. 128, 469–490 44 Yachi, S. and Higashi, M. (1998) The evolution of warning signals. Nature 394, 882–884 45 Leimar, O. and Tuomi, J. (1998) Synergistic selection and graded traits. Evol. Ecol. 12, 59–71 46 Verzijden, M.N. et al. (2007) Song discrimination learning in zebra finches induces highly divergent responses to novel songs. Proc. R. Soc. B 274, 295–301 47 Madden, J.R. and Tanner, K. (2003) Preferences for coloured bower decorations can be explained in a nonsexual context. Anim. Behav. 65, 1077–1083 48 Laland, K. (1994) On the evolutionary consequences of sexual imprinting. Evolution 48, 477–489 49 ten Cate, C. (2000) How learning mechanisms might affect evolutionary processes. Trends Ecol. Evol. 15, 179–181 50 Kamo, M. et al. (2002) The evolution of signal form: effects of learned versus inherited recognition. Proc. R. Soc. B 269, 1765–1771 51 Aoki, K. et al. (2001) Models of sexual selection on a quantitative genetic trait when preference is acquired by sexual imprinting. Evolution 55, 25–32
Review
TRENDS in Ecology and Evolution
52 Ellers, J. and Slabbekoorn, H. (2003) Song divergence and male dispersal among bird populations: a spatially explicit model testing the role of vocal learning. Anim. Behav. 65, 671–681 53 Beltman, J.B. et al. (2004) Learning and the colonization of new niches: a first step towards speciation. Evolution 58, 35–46 54 Lachlan, R.F. and Servedio, M.R. (2004) Song learning accelerates allopatric speciation. Evolution 58, 2049–2063 55 Verzijden, M.N. et al. (2005) Female mate-choice behavior and sympatric speciation. Evolution 59, 2079–2108 56 Moretz, J.A. and Morris, M.R. (2006) Phylogenetic analysis of a signal of aggressive intent in northern swordtail fishes. Am Nat 168, 336–349 57 Jansson, L. and Enquist, M. (2003) Receiver biases for colourful signals. Anim. Behav. 66, 965–971 58 Jansson, L. and Enquist, M. (2005) Testing the receiver bias hypothesis empirically with ‘virtual evolution’. Anim. Behav. 70, 865–875 59 Wills, S. and Mackintosh, N.J. (1998) Peak shift on an artificial dimension. Quart. J. Exp. Psychol. B 51, 1–31 60 McLaren, I.P.L. and Mackintosh, N.J. (2002) Associative learning and elemental stimulus representation: II. Generalization and discrimination. Anim. Learn. Behav. 30, 177–200 61 Spence, K. (1937) The differential response in animals to stimuli varying within a single dimension. Psychol. Rev. 44, 430–444
Vol.22 No.7
62 Blough, D.S. (1975) Steady state data and a quantitative model of operant generalization and discrimination. J. Exp. Psychol. Anim. Behav. Proc. 1, 3–21 63 ten Cate, C. and Vos, D.R. (1999) Sexual imprinting and evolutionary processes in birds: a reassessment. Adv. Study Behav. 28, 1–31 64 Vos, D.R. (1994) Sex recognition in zebra finch males results from early experience. Behaviour 128, 1–14 65 Vos, D.R. (1995) The role of sexual imprinting for sex recognition in zebra finches: a difference between males and females. Anim. Behav. 50, 645–653 66 Kendrick, K.M. et al. (1998) Mothers determine sexual preferences. Nature 395, 229–230 67 Verzijden, M.N. and ten Cate, C. (2007) Early learning influences species assortative mating preferences in Lake Victoria cichlid fish. Biol. Lett. 3, 134–136 68 Noor, M.A.F. (1999) Reinforcement and other consequences of sympatry. Heredity 83, 503–508 69 Pfennig, K.S. (1998) The evolution of mate choice and the potential for conflict between species and mate recognition. Proc. R. Soc. B 265, 1743–1748 70 Irwin, D.E. and Price, T. (1999) Sexual imprinting, learning and speciation. Heredity 82, 347–354
Elsevier celebrates two anniversaries with a gift to university libraries in the developing world In 1580, the Elzevir family began their printing and bookselling business in the Netherlands, publishing works by scholars such as John Locke, Galileo Galilei and Hugo Grotius. On 4 March 1880, Jacobus George Robbers founded the modern Elsevier company intending, just like the original Elzevir family, to reproduce fine editions of literary classics for the edification of others who shared his passion, other ‘Elzevirians’. Robbers co-opted the Elzevir family printer’s mark, stamping the new Elsevier products with a classic symbol of the symbiotic relationship between publisher and scholar. Elsevier has since become a leader in the dissemination of scientific, technical and medical (STM) information, building a reputation for excellence in publishing, new product innovation and commitment to its STM communities. In celebration of the House of Elzevir’s 425th anniversary and the 125th anniversary of the modern Elsevier company, Elsevier donated books to ten university libraries in the developing world. Entitled ‘A Book in Your Name’, each of the 6700 Elsevier employees worldwide was invited to select one of the chosen libraries to receive a book donated by Elsevier. The core gift collection contains the company’s most important and widely used STM publications, including Gray’s Anatomy, Dorland’s Illustrated Medical Dictionary, Essential Medical Physiology, Cecil Essentials of Medicine, Mosby’s Medical, Nursing and Allied Health Dictionary, The Vaccine Book, Fundamentals of Neuroscience, and Myles Textbook for Midwives. The ten beneficiary libraries are located in Africa, South America and Asia. They include the Library of the Sciences of the University of Sierra Leone; the library of the Muhimbili University College of Health Sciences of the University of Dar es Salaam, Tanzania; the library of the College of Medicine of the University of Malawi; and the University of Zambia; Universite du Mali; Universidade Eduardo Mondlane, Mozambique; Makerere University, Uganda; Universidad San Francisco de Quito, Ecuador; Universidad Francisco Marroquin, Guatemala; and the National Centre for Scientific and Technological Information (NACESTI), Vietnam. Through ‘A Book in Your Name’, these libraries received books with a total retail value of approximately one million US dollars.
For more information, visit www.elsevier.com www.sciencedirect.com
387
TRENDS in Ecology and Evolution
Vol.22 No.7
Biases in signal evolution: learning makes a difference Carel ten Cate1 and Candy Rowe2 1
Behavioural Biology, Institute of Biology, Leiden University, PO Box 9516, 2300 RA Leiden, The Netherlands Centre for Behaviour & Evolution, Division of Psychology, Newcastle University, Henry Wellcome Building, Framlington Place, Newcastle upon Tyne, NE2 4HH, UK
2
It is now well established that signal receivers have a key role in the evolution of animal communication: the suite of sensory and cognitive processes by which animals perceive and learn about their environment can have a significant impact on signal design. A crucial property of these information-processing mechanisms is the emergence of ‘receiver bias’ in the behavioural responses to signals. Whereas most research has focussed on receiver biases in the sensory system, more recent studies show that biases can also arise from learning about signals. Here, we highlight how learning-based biases can arise, and how these differ from biases emerging from sensory systems in their impact on signal evolution. Signal evolution by receiver biases A commonly held view is that signal evolution progresses through selection for costly signals that provide receivers with reliable information [1–4]. However, over recent years, this position has been eroded by studies showing that signal evolution can be driven by ‘receiver biases’ [5–10]. Receiver biases can be defined as byproducts of natural selection or as incidental non-selected consequences of the way in which sensory systems or brains are formed [4], predisposing receivers to respond more strongly to one signal than to another and even to prefer novel signals [9,10]. Here, we demonstrate that different types of receiver bias operate under different conditions and can have different implications for signal evolution. We focus on two opposite corners of what can be viewed as a continuous space determined by two different dimensions of the proximate mechanisms underlying receiver biases. The first dimension is whether biases result from the peripheral physiology of sensory systems or from higher cognitive brain processes. The second dimension is developmental phenotypic plasticity: some biases develop in the same way over a wide range of conditions, whereas others depend heavily on specific experience. We differentiate here between ‘sensory system biases’ (see Glossary), which we define as receiver biases arising from more peripheral and primary sensory processing mechanisms and also showing limited if any phenotypic plasticity, and ‘learning-based biases’, which we define as arising from central information processing involving plasticity generated by learning. In making this distinction, Corresponding author: ten Cate, C. ([email protected]). Available online 26 March 2007. www.sciencedirect.com
we create a false dichotomy within a continuum. However, we do so to illustrate how different proximate mechanisms can differ in their impact on signal evolution. Sensory system and learning-based biases Perhaps the most well known and widely cited example of a signal that might have evolved owing to a sensory system bias is the ‘chuck’ component of the male mating call in the Tu´ngara frog Physalaemus pustulosus. Results from comparative studies suggest that, before the chuck signal evolved, ancestral female frogs already had a bias in their auditory system to respond to particular sound features, which drove subsequent call evolution [5,6]. Other examples of traits that are likely to have evolved by exploiting sensory system biases are the swords of swordtail Xiphophorus fish [7,8], the vibratory signals of the water mite Neumania papillator [11,12], the red coloration of the male three-spined stickleback Gasterosteus aculeatus [13] and the blue and yellow coloration of two Lake Victoria cichlid species Pundamilia pundamilia
Glossary Area shift: a phenomenon similar to peak shift where the peak response remains highest for the training stimuli but the generalization gradient is asymmetric, resulting in higher responses to novel stimuli away from the S+ in the direction opposite from the S and vice versa. Discrimination learning: the process by which animals learn to discriminate among stimuli, either along a single stimulus dimension or using more complex multidimensional features. Generalization gradient: the degree to which an animal trained to respond to a particular stimulus will respond to novel stimuli that vary along a stimulus dimension shared with the training stimulus. Intensity dimension: stimuli differ along an intensity dimension when they stimulate the same receptors but to a different extent; for instance, two lights of the same wavelength differing in intensity; chemical substances differing only in concentration; or tones differing only in amplitude. Learning-based bias: a receiver bias that is generated through learning; for example, peak shift. Peak shift: a consequence of discrimination learning between an S+ (a positively rewarded stimulus) and an S (a negatively or neutrally rewarded stimulus) that differ along a stimulus dimension, leading to stronger responding to novel stimuli away from the S+ in a direction opposite from the S , and vice versa (Box 1). Rearrangement dimension: stimuli differ along a ‘rearrangement’ dimension when each stimulus addresses a different set of receptors, such as lights differing in wavelength; chemicals of different structure; or tones of different frequency. Whereas an intensity dimension can be considered a quantitative dimension, the rearrangement dimension is more qualitative. Sensory system bias: a receiver bias arising from peripheral sensory processing. Although this need not imply absence of environmental plasticity, we assume here that it does. Stimulus dimension: any aspect of a stimulus that can vary along an axis; for example, light intensity or angle of orientation.
0169-5347/$ – see front matter ß 2007 Elsevier Ltd. All rights reserved. doi:10.1016/j.tree.2007.03.006
Review
TRENDS in Ecology and Evolution
and Pundamilia nyererei [14]. These biases are assumed either to be byproducts of the sensory systems involved [7] or to arise in these systems as a result of selection for signal detection in other contexts. For example, the vibratory water movements in the water mite might have arisen from exploiting predator detection mechanisms [11,12], the red coloration of stickleback males might result from selection for preferences for red food items [13], and the blue and yellow colour in cichlids might be a consequence of a differential sensitivity to long and short wavelengths owing to adaptation to different ambient light conditions in both species [14]. Although we do not know whether these biases lack phenotypic plasticity, this is often implied from the finding that related species share similar biases and that the processing is peripheral. Here, we assume that they have limited plasticity. Learning-based biases arise from the cognitive mechanisms that process and store information. There is an increasing awareness that biases based on cognitive processes could be just as important for signal evolution as are sensory system biases [15–25]. Not all biases arising from cognitive processing are necessarily plastic. However, most interesting from our perspective are biases that arise from the way in which animals learn to recognize important environmental features. One mechanism that can generate a learning-based bias is that of ‘peak shift’, which emerges from learning to discriminate among stimuli [26,27]. Using peak shift as an example, we highlight the need to consider sensory system and learning-based biases as distinct selection pressures in signal evolution. Learning and generalization In many contexts, responding to signals involves learning. Avian predators, for instance, avoid prey with aposematic warning signals by learning to associate the coloration with the defences (e.g. Ref. [28]), songbirds learn to recognize the songs of conspecifics or local geographical variants of songs (e.g. Ref. [29]), and bumble bees learn to recognize the colour of the most profitable flowers [30]. Recognition involves remembering particular signals, but also discriminating them from other similar signals. Thus, predators can learn to discriminate between undefended and defended prey using visual patterns, songbirds learn to ignore the songs of other sympatric bird species, and bumble bees learn to avoid the colours of less profitable flowers. This discrimination learning process also determines how animals respond to novel stimuli; that is, how they generalize (e.g. Refs [26,27,31–33]). Generalization gradients resulting from discrimination learning have been extensively studied by psychologists in experimental laboratory settings, often using simple artificial stimuli, such as tilted lines, lights of different wavelengths or tones of different frequency (e.g. Refs [26,27,31–34]). In these experiments, animals are usually trained to discriminate between two stimuli that are differentially reinforced, where one stimulus indicates the presence of a reward (S+), and the other stimulus is neutral or indicates some punishment (S ). After training, animals are presented with novel stimuli that differ from their training stimuli to varying degrees, and their responses are measured. Although we might expect that the strongest www.sciencedirect.com
Vol.22 No.7
381
responses are seen towards the training stimuli, this is not always the case, and the strongest response might be given to stimuli that are more extreme on the dimension separating the training stimuli (Box 1). This shift in generalization gradient is known as a ‘peak shift’ because the peak response is shifted along the training dimension away from the original stimuli used in training. The shape of this generalization gradient differs depending on whether the training stimuli vary along an ‘intensity’ dimension (e.g. two lights of the same wavelength differing in intensity) or qualitatively along a ‘rearrangement’ dimension (e.g. two lights of the same intensity but of a different wavelength) (Box 1). A less extreme version of peak shift is ‘area shift’, where, although the peak response remains highest for the training stimuli, the generalization gradient is skewed, with novel stimuli on the S side of S+ receiving fewer responses than S+ and novel stimuli on the other side of S+ getting similar responses as S+ [34,35]. Apart from peak and area shift, other types of generalization processes can also bring about preferences for novel stimuli, such as ‘range effects’ (e.g. Ref. [34]), in which the location of the peak in responding depends on the range of stimuli used in testing. However, the relevance of these processes to signal evolution is still unclear. Peak shift is an example of a learning-based bias that can drive signal evolution. It is taxonomically widespread, occurring in vertebrates and invertebrates, and appears to be a general property of discrimination learning. Peak shift is also found for a range of different stimuli in different sensory modalities (e.g. vision, hearing or olfaction), and along many perceptual dimensions (e.g. such as the frequency, duration or amplitude of a sound signal). It is a fundamental property of discrimination learning; therefore, whenever animals learn to discriminate among closely similar signals, whether prey coloration, conspecific song, or floral colours, they might develop biases that can select for more extreme signals. However, does peak shift occur in more biologically relevant contexts similar to the ones the animal has to solve in every day life, or with more complex stimuli compared with the tilted lines or coloured keys found in traditional experimental psychology experiments? Peak shift in the natural world? There is increasing evidence for peak shifts in a more natural context. One example of this concerns a task that humans perform in daily life: face recognition. Humans show peak shift when they have to distinguish similarly looking faces [35]. This is also thought to give rise to the phenomenon where we are better at recognizing familiar individuals from caricatures exaggerating specific features than from drawings based on real facial features [36]. For animals, several recent studies have also used complex stimuli that either are close to natural ones or are biologically relevant, for instance using a learning task that is more similar to one they are likely to experience in the wild. Peak shift has, for example, been observed in the spatial orientation of pigeons Columbia livia and honey bees Apis mellifera in relation to artificial ‘landmarks’ [37,38], while sphinx moths Manduca sexta, which use
382
Review
TRENDS in Ecology and Evolution Vol.22 No.7
Box 1. The mechanism of peak shift There are three main theoretical hypotheses to explain the apparently counterintuitive phenomenon of peak shift. It was originally thought that animals learn the relationship between the positively rewarded stimulus (S+) and the negatively rewarded or neutral one (S ), rather than the absolute value of each. According to this ‘relational hypothesis’, animals might learn a rule such as ‘choose the brighter stimulus’. This might make them prefer a more extreme version of a rewarded stimulus when given the choice between the rewarded stimulus and the more extreme version. However, most experimental findings cannot be explained in this way [31,34,59,60]. An alternative hypothesis suggested that peak shift results from overlap between the generalization gradients around S+ and S . The interaction between the excitatory gradient around S+ and the inhibitory gradient around S could result in a combined gradient in which the strongest response is not given to S+, but is displaced, away from S [61]. This ‘conditioning-extinction theory’ has been propagated in the animal communication literature [15,40], but its underlying assumptions are not supported [31]. Finally, the ‘elemental’ or ‘receptor’ account proposes that the amount of feature sharing between novel and training stimuli will determine the response to novel stimuli [27,37,59,62]. If S+ and S
share some features, then stronger responses to novel than to training stimuli arise when novel stimuli share features of S+ that are not present in S . This can also explain the difference in generalization gradients between stimuli differing along ‘intensity’ and ‘rearrangement’ dimensions [19,27]. Stimuli differ along an intensity dimension when they stimulate the same receptors but to a different extent; for instance, two lights of the same wavelength differing in intensity, chemical substances differing only in concentration or tones differing only in amplitude. In these cases, the generalization gradient is often steep and, to some extent, monotonic (Figure I), indicating that the response to novel, more extreme, stimuli is sometimes considerable stronger than to the training stimuli. When the training stimuli differ along a ‘rearrangement’ dimension, each stimulus addresses a different set of receptors, such as lights differing in wavelength, chemicals of different structure or tones of different frequency. In these cases, the generalization gradient usually peaks close to the training stimulus and the peak response is not much stronger than when responding to familiar stimuli (Figure I; reviewed in Ref. [27]). Apart from the dimension on which the stimuli differ, another parameter that affects the shape and skew of the generalization is how close together S+ and S are; the closer they are, the stronger the shift [27].
Figure I. Idealized generalization gradients resulting from discrimination learning. The X-axis represents a range of stimuli differing on one particular dimension. When animals have been trained to respond differentially to two stimuli, one providing a positive (S+) and one a negative reward (S ), one would expect the animal to respond most strongly to training stimuli compared with novel stimuli (i.e. according to the grey line). However, this is only the case when S+ and S differ relatively strongly. If the stimuli are similar, the generalization gradient will show a peak shift (arrows) and be similar to the red gradient (for stimuli differing along a rearrangement dimension) or to the blue gradient (for stimuli differing along an intensity dimension) [27]. The vertical solid lines indicate the location of the S+ and S and of the peaks for a generalization gradient without peak shift. The dashed lines indicate the location of the peaks for a generalization gradient involving peak shift along a rearrangement dimension.
odours to find food, show area shift if conditioned with odours differing in the length of the carbon chain or the functional groups (alcohol or ketone) [39]. Peak shift has been proposed as driving force behind signal evolution [17], in particular for the contexts of food preference learning [15] and sexual communication [40]; indeed, the best evidence for peak shift in natural systems comes from these two areas. Many animals need to learn which food types are profitable sources of nutrients and which are less profitable, either in terms of handling costs www.sciencedirect.com
or energetic content. For example, bumble bees Bombus terrestris exhibit peak-shifted colour preferences following training on artificial ‘flower’ colour discrimination and such changes are suggested to drive the evolution of floral species that have orchids as batesian mimics [30]. Domestic chicks Gallus domesticus trained to discriminate between palatable and aposematic insect prey subsequently show a biased generalization against prey that have stronger or larger warning signals [41,42]. This suggests that peak shift might drive the evolution of
Review
TRENDS in Ecology and Evolution
aposematic signals, a suggestion supported by theoretical models [43–45]. Two recent studies suggest that peak shift can be responsible for the evolution of exaggerated sexual traits. One of these used artificially constructed zebra finch Taeniopygia guttata songs as discriminative stimuli. It showed that perceptual learning about songs varying along an intensity dimension (i.e. number of elements in a song) and songs varying along a rearrangement dimension (i.e. position of an odd element in a series of repeated elements) resulted in differently skewed generalization patterns, similar to what can be observed in discrimination studies using simple stimuli. Peak shift was present for the responses to novel songs with higher or lower numbers of elements than the training stimuli (i.e. along the intensity gradient) [46]. This finding suggests that it is easier to get song evolution driven by a receiver bias along an intensity
Vol.22 No.7
383
dimension than along a rearrangement dimension. The second study showed that male zebra finches developed peak-shifted sexual preferences for beak colouration through the learning process of sexual imprinting (Box 2; [21]). Crucially, these examples show that the findings obtained in highly simplified experimental conditions appear to hold under more complex situations involving tasks closer to the ones the animals are likely to encounter in nature. As there are many natural contexts in which animals learn about signals, and because peak shift is a widespread emergent property of discrimination learning, other cases are likely to be common. For instance, the example presented in Box 2 suggests that peak shift can also be found in other contexts involving learning about features of conspecifics or closely similar species, such as when two highly similar species overlap in their
Box 2. Super mates: peak shift and exaggerated sexual signals Many birds [63], as well as some mammals [64] and fish [65], acquire their mating preferences by learning about their parents and taking these as a model for their later sexual preference, a process known as ‘sexual imprinting’. At the same time, sexual preferences are often skewed and can be combined with sexual dimorphism, resulting in preferences for mates with exaggerated traits. Could such skewed mating preferences be a side effect of imprinting, owing to a peak shift? If offspring are raised by a sexually dimorphic species with biparental care, the presence of two parents differing in appearance and behaviour provides a context for discrimination learning. Whether this can induce peak-shifted mate preferences [40] was tested in an experiment in which young zebra finch Taeniopygia guttata males were raised by white zebra finch parents that differed in beak colour only (Figure I) [21]. In one experimental group, the beak of the father was painted red and that of the mother orange using nail varnish. In a second group, the reverse occurred. In earlier studies [66,67], males raised in this way went on to prefer females that were similar in appearance to the mother, suggesting that mothers act as
an ‘S+’ and fathers as an ‘S ’. In the current experiment, the males could choose between eight females differing in beak colour. Both groups preferred females with a beak in a more extreme colour than their mother (i.e. more intense red when the mother’s beak was red and more yellow orange when the mother’s beak colour was orange). Hence, the parental beak colouration induced directional colour biases in the mate preference of male offspring. The experiment demonstrates that peak shift can arise from a naturally occurring and crucial learning process, such as sexual imprinting. Although the evolution of sexual dimorphism cannot be reconstructed with such an experiment, it does show how a slight initial sexual dimorphism can give rise to the emergence of a bias for novel signals that exaggerate the dimorphism. It is the difference between the two stimuli involved in the learning process that determines both the dimension of the preference (beak colour in this case), as well as the direction of the bias (more extreme than the mother’s beak). Had the parents not differed in beak colour, but along another dimension (e.g. some plumage characteristic), then the bias would have been different.
Figure I. Peak shift in the sexual preferences of zebra finch males. Arrows indicate the parental beak colours (mother orange, father red for one group; mother red, father orange for the other). White bars and grey bars indicate the preferences (median and interquartile ranges) for females having different beak colours. Males show stronger preferences for females with a more extreme beak colour compared with that of their mother. Modified with permission from Ref. [21]; drawing of finches by Machteld Verzijden. The beak colours in the figure provide an indication, but not an exact match, of the colours used in the experiment. www.sciencedirect.com
384
Review
TRENDS in Ecology and Evolution Vol.22 No.7
geographical range. In such a situation, one can imagine that learning to discriminate between the two species could induce a peak-shifted recognition of conspecifics that might drive character displacement as an emergent property of the learning process (Box 3). This and other contexts might provide fruitful areas for further exploration. The same holds for the question of whether generalization gradients are subject to selection [23,30, 33]. Bumble bees, for instance, respond with an adaptive peak shift to changes in payoffs of selecting different stimuli [30], indicating that the shape of generalization gradients is not always the same, but varies depending on the consequences of making wrong discriminations or the current payoffs during learning [21,26]. It is possible that selection operates on such plasticity and that the fitness consequences of the generalization pattern affect the speed and direction of signal evolution.
Comparing sensory system and learning-based biases Sensory system and learning-based biases are similar in that they both predispose an animal to be most responsive to signals that have not been experienced before and that might guide subsequent evolution. But to what extent do learning-based biases differ from those arising in sensory systems? Sensory system biases have their origins in other contexts and exist before the evolution of the signal. For example, selection for a preference for a particular food colour, as demonstrated in several species [13,47], could lead to a sensory system with a general, non-learned sensitivity to that type of stimulation. The more this bias depends on peripheral sensory physiology or processing, the more likely it is that the bias shows limited, if any, developmental plasticity. Hence, it will be similar from one generation to the next. If, in a mating context, a potential
Box 3. Peak shift, species discrimination and character displacement Sympatric species usually exhibit greater species mating discrimination than do allopatric species. This is often ascribed to the evolutionary process of reinforcement; the adaptive strengthening of mate discrimination to avoid detrimental hybridization (e.g. Ref. [68]), which might be the outcome of several mechanisms selecting for genetically based preferences for particular types of mates (e.g. Refs [18,69]). Although it is clear that changes in discriminative ability can arise in this way, incorporating learning as a mechanism involved in species recognition suggests an alternative scenario in which the increase in discriminative ability arises as an emergent byproduct. Adult learning about conspecific and heterospecific traits is of widespread importance in birds and can result in stronger species discrimination in sympatric areas (reviewed in Ref. [70]). Such learning can provide a mechanism for a flexible adaptation to the current situation [70]. We suggest that skewed generalization makes the impact of learning greater than already anticipated. Suppose we are dealing with two allopatric, closely similar bird species (Figure I), differing in wing colour only. Each allopatric species learns the characteristics of it own species, resulting in an
average preference for mates with an appearance around the population mean (Figure Ia). When the species suddenly become sympatric locally (e.g. by crossing a geographical barrier), individuals might discover that they sometimes interact with conspecifics and sometimes with birds that are responding in an unusual way and also look slightly different: the other species. This context can result in discrimination learning that might give rise to biases in which each species will preferentially respond to those conspecifics that are more distinct from the other species (Figure Ib). The presence of sympatry will itself thus give rise to greater species mating discrimination in sympatry than in allopatry as an emergent property. One can imagine that this might drive character displacement. A particular feature is that the effect will only be induced and, hence, present in those parts of the population where the two species overlap. The exact dynamics of the process will be different depending upon whether the discrimination operates along a rearrangement dimension, leading to a limited peak shift (as in Figure I), or a dimension based upon intensity, which may give rise to more open-ended and stronger preference shifts [27].
Figure I. Hypothetical representation of traits and learned preferences of two closely similar species in allopatry (a) and sympatry (b). In allopatry, the trait (signal) distribution (solid line) and the recognition of this trait (hatched line) show a close match. In sympatry, mutual interactions will lead to a peak shift of the recognition gradients away from the actual traits, as indicated by the arrows, providing scope for selection on traits to become more distinct. www.sciencedirect.com
Review
TRENDS in Ecology and Evolution
partner provides a signal that tunes in to the bias, this results in selection for that novel signal and subsequent signal evolution. Evolutionary changes in the bias will primarily be the consequence of changes in the selection pressures in the context that gave rise to the bias. In this case, a new food type of a different colour could generate selection for new signals of that colour. The nature or direction of this bias will only be affected by the current signals if the colour preference in the signalling context becomes independent of a general perceptual visual preference. By contrast, biases through peak shift arise within the specific signalling context. Whereas generalization is a general property of the nervous system of animals, learning about a signal is not all that is required for a skewed generalization gradient to emerge. It also requires that two or more signals are closely related, otherwise, learning will not give rise to skewed generalization in the first place: the more similar two stimuli are, the more exaggerated the biases will be. In a signalling context, this could lead to stronger selection pressures acting on signals. To provide a consistent directional selection, a high degree of similarity in the learning experiences of individuals within a population is also needed. Such a shared bias that would also be stable over generations might arise, for instance, when predatory species all learn to discriminate among the same colour patterns of edible and non-edible prey items. Peak shift is a function of the discrimination process: its nature (i.e. a monotonic open-ended preference when stimuli differ along an intensity dimension, or a peaked one with a rearrangement dimension), direction and strength all depend upon the current contrast between the stimuli and, hence, will be plastic. As the direction of the signal evolution will depend upon the signalling niche in which it evolves, individuals that differ in their experiences will also develop different learning-based biases, unlike with sensory biases. This could lead to receivers selecting to diversify signals across populations rather than to drive them in the same direction, which could lead to speciation between different populations. Because the bias is generated by the current contrast between stimuli or signals, it might be less likely to induce novel traits. As the bias will only arise from discrimination learning involving closely similar stimuli, it is not something that, on its own, can generate large differences. However it might be important in the initial process of exaggerating small differences before other selection processes take over. Finally, if the signal changes as a consequence of selection, the bias will follow. Hence, the bias is not only contingent upon the present signal contrast, but also evolves in concert with this contrast. This plasticity can enable rapid evolutionary change. Evolutionary implications Thus, there are several differences in the origin and plasticity of sensory system and learning-based biases, but why does this distinction matter? The plasticity over time of learning-based biases is key to their function and, given their widespread occurrence, they could influence many signalling scenarios. Many studies have shown that incorporating the plasticity that learning mechanisms www.sciencedirect.com
Vol.22 No.7
385
provide in evolutionary models can alter the dynamics of evolutionary processes such as speciation and sexual selection compared with when only genetic mechanisms are considered [48–55]. Nevertheless, most evolutionary models that do incorporate learning in recognition systems assume either that learned preferences match the stimulus about which the learning occurred, or invoke a non-learned bias to account for or induce changes in preferences [48,51–55]. Incorporating skewed generalization will add a mechanism that might be responsible for driving evolutionary changes in many contexts involving learning, enabling rapid and diversifying changes in signal characteristics that are less likely to be accomplished by sensory system biases. Demonstrating trait evolution by learning-based biases A final point that we want to make concerns the research strategy assessing the potential impact of learning-based biases on signal evolution. This strategy is, to some extent, necessarily different from assessing the impact of sensory system biases. Studies on sensory exploitation, as in the Tu´ngara frogs [5,6], swordtails [7,8,56] and others [12,13,47], use comparative studies within a phylogenetic framework to demonstrate that preference for a trait was likely to be present in an ancestor of the current species, before emergence of the trait itself. For a preference emerging from the learning process, this approach has limited value. A characteristic of learning-based biases is that they are transient. As changing conditions between generations affect the nature, shape or even the presence of biases, they might be less likely to leave a phylogenetically traceable effect. But it is not impossible. If species A has a more elaborated signal than its sister species B, and if species B learns about the signal and shows a biased preference in the direction of a signal similar to A owing to peak shift, this suggests that the elaborated signal in species A is a consequence of selection driven by skewed generalization. Nevertheless, we might have to revert to other methods to ascertain whether certain traits or signals arose as a consequence of exploiting a generalization skew. One approach is to demonstrate that skewed generalization gradients arise in a context where one can hypothesize their role at present or in the past, and with stimuli of the relevant type. So, whereas the example in Box 2 does not demonstrate how the current sexual dimorphism in the zebra finch has arisen or whether it is still driven by peak shift, it does demonstrate that peak shift can occur after imprinting on a slightly sexually dimorphic trait. Hence, it illustrates the potential of peak shift to contribute to the differentiation of sexually selected traits. Another type of experiment indicating the evolutionary potential of a skew in generalization is one in which a trait is allowed to ‘evolve’ in response to behavioural choices made by an animal. An example of this is a series of experiments in which colours or other traits on a computer screen could change (‘evolve’) in response to choices made by domestic chickens trained to discriminate between them [57,58]. Such evidence can be supplemented with mathematical models [50] that test whether the presumed starting conditions, combined with characteristics of the learning process, result in reliable evolutionary changes.
386
Review
TRENDS in Ecology and Evolution Vol.22 No.7
Conclusions Biases can drive signal evolution, but it is clear that the origin and nature of a bias matters as to what, when and how signals evolve. With our example of peak shift, we have deliberately chosen an example of a learning-based bias that is different from a sensory system bias to highlight the contrasts. Not all cognitive biases will be like this, and not all cases of sensory biases will be as developmentally rigid as we presented them. However, by providing such a strong contrast, we hope to have demonstrated that receiver biases resulting from different proximate origins can result in different outcomes and can be relevant in different contexts. These matters are wide open for further research, and we indicated some areas that we consider profitable. In our view, we have only just begun to scratch the surface of how cognitive processes might affect the nature, direction and speed of signal evolution. But even so, we feel justified to conclude that learning does make a difference. Acknowledgements We thank Christina Halpin, Sue Healy, Rob Lachlan, Hans Slabbekoorn, Machteld Verzijden, Jeri Wright and three referees for helpful comments and/or discussion. C.t.C. was a Lorentz Fellow at NIAS (Wassenaar, NL) during the final revision.
References 1 Zahavi, A. (1975) Mate selection: a selection for a handicap. J. Theor. Biol. 53, 205–214 2 Grafen, A. (1990) Biological signals as handicaps. J. Theor. Biol. 144, 517–546 3 Andersson, M. (1994) Sexual Selection, Princeton University Press 4 Searcy, W.A. and Nowicki, S. (2005) The Evolution of Communication, Princeton University Press 5 Ryan, M.J. et al. (1990) Sexual selection for sensory exploitation in the frog Physalaemus pustulosus. Nature 343, 66–67 6 Ryan, M.J. and Rand, A.S. (1993) Sexual selection and signal evolution: the ghost of biases past. Philos. Trans. R. Soc. B 340, 187–195 7 Basolo, A.L. (1990) Female preference predates the evolution of the sword in female swordtail fish. Science 250, 808–810 8 Basolo, A.L. (1995) Phylogenetic evidence for the role of pre-existing bias in sexual selection. Proc. R. Soc. B 259, 307–311 9 Endler, J.A. and Basolo, A.L. (1998) Sensory ecology, receiver biases and sexual selection. Trends Ecol. Evol. 13, 415–420 10 Ryan, M.J. (1998) Sexual selection, receiver biases, and the evolution of sex differences. Science 281, 1999–2003 11 Proctor, H.C. (1991) Courtship in the water mite Neumania-papillator – males capitalize on female adaptations for predation. Anim. Behav. 42, 589–598 12 Proctor, H.C. (1993) Sensory exploitation and the evolution of male mating-behavior – a cladistic test using water mites (Acari, Parasitengona). Anim. Behav. 44, 745–752 13 Smith, C. et al. (2004) A receiver bias in the origin of three-spined stickleback mate choice. Proc. R. Soc. B 271, 949–955 14 Maan, M.E. et al. (2006) Sensory drive in cichlid speciation. Am. Nat. 167, 947–954 15 Guilford, T. and Dawkins, M.S. (1991) Receiver psychology and the evolution of animal signals. Anim. Behav. 42, 1–14 16 Guilford, T.C. and Dawkins, M.S. (1993) Receiver psychology and the design of animal signals. Trends Neurosci. 16, 430–436 17 Staddon, J.E.R. (1975) A note on the evolutionary significance of ‘supernormal’ stimuli. Am. Nat. 109, 541–545 18 Enquist, M. and Arak, A. (1998) Neural representation and the evolution of signal form. In Cognitive Ecology (Dukas, R., ed.), pp. 21–87, University of Chicago Press 19 Enquist, M. and Ghirlanda, S. (2005) Neural Networks and Animal Behavior, Princeton University Press 20 Rowe, C. and Skelhorn, J. (2004) Avian psychology and communication. Proc. R. Soc. B 271, 1435–1442 www.sciencedirect.com
21 ten Cate, C. et al. (2006) Sexual imprinting can induce sexual preferences for exaggerated parental traits. Curr. Biol. 16, 1128–1132 22 Lynn, S.K. (2005) Learning to avoid aposematic prey. Anim. Behav. 70, 1121–1226 23 Lynn, S.K. (2006) Cognition and evolution: learning and the evolution of sex traits. Curr. Biol. 16, R421–R423 24 Bateson, M. and Healy, S.D. (2005) Comparative evaluation and its implications for mate choice. Trends Ecol. Evol. 20, 659–664 25 Chittka, L. and Raine, N.E. (2006) Recognition of flowers by pollinators. Curr. Opin. Plant Biol. 9, 428–435 26 Purtle, R.B. (1973) Peak shift: a review. Psychol. Bull. 80, 408–421 27 Ghirlanda, S. and Enquist, M. (2003) A century of generalization. Anim. Behav. 66, 15–36 28 Guilford, T. (1990) Predator psychology and the evolution of prey coloration. In Natural Enemies: The Population Biology of Predators, Parasites and Diseases (Crawley, M.J., ed.), pp. 377–494, Blackwell Scientific 29 Riebel, K. (2003) The ‘mute’ sex revisited; vocal production and perception learning in female songbirds. Adv. Stud. Behav. 33, 49–86 30 Lynn, S.K. et al. (2005) Peak shift discrimination learning as a mechanism of signal evolution. Evolution 59, 1300–1305 31 Mackintosh, N.J. (1983) Conditioning and Associative Learning, Clarendon Press 32 Mackintosh, N.J. (1995) Categorization by people and pigeons: the twenty-second Bartlett memorial lecture. Quart. J. Exp. Psychol. B 48, 193–214 33 Cheng, K. (2002) Generalisation: mechanistic and functional explanations. Anim. Cogn. 5, 33–40 34 Thomas, D.R. et al. (1991) Peak shift revisited: a test of alternative interpretations. J. Exp. Psychol. 17, 130–140 35 Spetch, M.L. et al. (2004) Peak shift but not range effects in recognition of faces. Learn. Motiv. 35, 221–241 36 Lewis, M.B. and Johnston, R.A. (1999) Are caricatures special? Evidence of peak shift in face recognition. Eur. J. Cogn. Psychol. 11, 105–117 37 Cheng, K. et al. (1997) Spatial peak shift and generalization in pigeons. J. Exp. Psychol. Anim. Behav. Process. 23, 469–481 38 Cheng, K. et al. (2006) Small-scale spatial orientation in pigeons. Behav. Process. 72, 115–127 39 Daly, K.C. et al. (2001) The generalization of an olfactory-based conditioned response reveals unique but overlapping odour representations in the moth Manduca sexta. J. Exp. Biol. 204, 3085–3095 40 Weary, D.M. et al. (1993) A product of discrimination learning may lead to female preferences for elaborate males. Evolution 47, 333–336 41 Gamberale, G. and Tullberg, B.S. (1996) Evidence for a peak-shift in predator generalization among aposematic prey. Proc. R. Soc. B. 263, 1329–1334 42 Gamberale-Stille, G. and Tullberg, B.S. (1999) Experienced chicks show biased avoidance of stronger signals: an experiment with natural colour variation in live aposematic prey. Evol. Ecol. 13, 579– 589 43 Leimar, O. et al. (1986) Evolutionary stability of aposematic colouration and prey unprofitability: a theoretical analysis. Am. Nat. 128, 469–490 44 Yachi, S. and Higashi, M. (1998) The evolution of warning signals. Nature 394, 882–884 45 Leimar, O. and Tuomi, J. (1998) Synergistic selection and graded traits. Evol. Ecol. 12, 59–71 46 Verzijden, M.N. et al. (2007) Song discrimination learning in zebra finches induces highly divergent responses to novel songs. Proc. R. Soc. B 274, 295–301 47 Madden, J.R. and Tanner, K. (2003) Preferences for coloured bower decorations can be explained in a nonsexual context. Anim. Behav. 65, 1077–1083 48 Laland, K. (1994) On the evolutionary consequences of sexual imprinting. Evolution 48, 477–489 49 ten Cate, C. (2000) How learning mechanisms might affect evolutionary processes. Trends Ecol. Evol. 15, 179–181 50 Kamo, M. et al. (2002) The evolution of signal form: effects of learned versus inherited recognition. Proc. R. Soc. B 269, 1765–1771 51 Aoki, K. et al. (2001) Models of sexual selection on a quantitative genetic trait when preference is acquired by sexual imprinting. Evolution 55, 25–32
Review
TRENDS in Ecology and Evolution
52 Ellers, J. and Slabbekoorn, H. (2003) Song divergence and male dispersal among bird populations: a spatially explicit model testing the role of vocal learning. Anim. Behav. 65, 671–681 53 Beltman, J.B. et al. (2004) Learning and the colonization of new niches: a first step towards speciation. Evolution 58, 35–46 54 Lachlan, R.F. and Servedio, M.R. (2004) Song learning accelerates allopatric speciation. Evolution 58, 2049–2063 55 Verzijden, M.N. et al. (2005) Female mate-choice behavior and sympatric speciation. Evolution 59, 2079–2108 56 Moretz, J.A. and Morris, M.R. (2006) Phylogenetic analysis of a signal of aggressive intent in northern swordtail fishes. Am Nat 168, 336–349 57 Jansson, L. and Enquist, M. (2003) Receiver biases for colourful signals. Anim. Behav. 66, 965–971 58 Jansson, L. and Enquist, M. (2005) Testing the receiver bias hypothesis empirically with ‘virtual evolution’. Anim. Behav. 70, 865–875 59 Wills, S. and Mackintosh, N.J. (1998) Peak shift on an artificial dimension. Quart. J. Exp. Psychol. B 51, 1–31 60 McLaren, I.P.L. and Mackintosh, N.J. (2002) Associative learning and elemental stimulus representation: II. Generalization and discrimination. Anim. Learn. Behav. 30, 177–200 61 Spence, K. (1937) The differential response in animals to stimuli varying within a single dimension. Psychol. Rev. 44, 430–444
Vol.22 No.7
62 Blough, D.S. (1975) Steady state data and a quantitative model of operant generalization and discrimination. J. Exp. Psychol. Anim. Behav. Proc. 1, 3–21 63 ten Cate, C. and Vos, D.R. (1999) Sexual imprinting and evolutionary processes in birds: a reassessment. Adv. Study Behav. 28, 1–31 64 Vos, D.R. (1994) Sex recognition in zebra finch males results from early experience. Behaviour 128, 1–14 65 Vos, D.R. (1995) The role of sexual imprinting for sex recognition in zebra finches: a difference between males and females. Anim. Behav. 50, 645–653 66 Kendrick, K.M. et al. (1998) Mothers determine sexual preferences. Nature 395, 229–230 67 Verzijden, M.N. and ten Cate, C. (2007) Early learning influences species assortative mating preferences in Lake Victoria cichlid fish. Biol. Lett. 3, 134–136 68 Noor, M.A.F. (1999) Reinforcement and other consequences of sympatry. Heredity 83, 503–508 69 Pfennig, K.S. (1998) The evolution of mate choice and the potential for conflict between species and mate recognition. Proc. R. Soc. B 265, 1743–1748 70 Irwin, D.E. and Price, T. (1999) Sexual imprinting, learning and speciation. Heredity 82, 347–354
Elsevier celebrates two anniversaries with a gift to university libraries in the developing world In 1580, the Elzevir family began their printing and bookselling business in the Netherlands, publishing works by scholars such as John Locke, Galileo Galilei and Hugo Grotius. On 4 March 1880, Jacobus George Robbers founded the modern Elsevier company intending, just like the original Elzevir family, to reproduce fine editions of literary classics for the edification of others who shared his passion, other ‘Elzevirians’. Robbers co-opted the Elzevir family printer’s mark, stamping the new Elsevier products with a classic symbol of the symbiotic relationship between publisher and scholar. Elsevier has since become a leader in the dissemination of scientific, technical and medical (STM) information, building a reputation for excellence in publishing, new product innovation and commitment to its STM communities. In celebration of the House of Elzevir’s 425th anniversary and the 125th anniversary of the modern Elsevier company, Elsevier donated books to ten university libraries in the developing world. Entitled ‘A Book in Your Name’, each of the 6700 Elsevier employees worldwide was invited to select one of the chosen libraries to receive a book donated by Elsevier. The core gift collection contains the company’s most important and widely used STM publications, including Gray’s Anatomy, Dorland’s Illustrated Medical Dictionary, Essential Medical Physiology, Cecil Essentials of Medicine, Mosby’s Medical, Nursing and Allied Health Dictionary, The Vaccine Book, Fundamentals of Neuroscience, and Myles Textbook for Midwives. The ten beneficiary libraries are located in Africa, South America and Asia. They include the Library of the Sciences of the University of Sierra Leone; the library of the Muhimbili University College of Health Sciences of the University of Dar es Salaam, Tanzania; the library of the College of Medicine of the University of Malawi; and the University of Zambia; Universite du Mali; Universidade Eduardo Mondlane, Mozambique; Makerere University, Uganda; Universidad San Francisco de Quito, Ecuador; Universidad Francisco Marroquin, Guatemala; and the National Centre for Scientific and Technological Information (NACESTI), Vietnam. Through ‘A Book in Your Name’, these libraries received books with a total retail value of approximately one million US dollars.
For more information, visit www.elsevier.com www.sciencedirect.com
387