Honeybees can discriminate between Monet and Picasso paintings
- First Online:
- Cite this article as:
- Wu, W., Moreno, A.M., Tangen, J.M. et al. J Comp Physiol A (2013) 199: 45. doi:10.1007/s00359-012-0767-5
Honeybees (Apis mellifera) have remarkable visual learning and discrimination abilities that extend beyond learning simple colours, shapes or patterns. They can discriminate landscape scenes, types of flowers, and even human faces. This suggests that in spite of their small brain, honeybees have a highly developed capacity for processing complex visual information, comparable in many respects to vertebrates. Here, we investigated whether this capacity extends to complex images that humans distinguish on the basis of artistic style: Impressionist paintings by Monet and Cubist paintings by Picasso. We show that honeybees learned to simultaneously discriminate between five different Monet and Picasso paintings, and that they do not rely on luminance, colour, or spatial frequency information for discrimination. When presented with novel paintings of the same style, the bees even demonstrated some ability to generalize. This suggests that honeybees are able to discriminate Monet paintings from Picasso ones by extracting and learning the characteristic visual information inherent in each painting style. Our study further suggests that discrimination of artistic styles is not a higher cognitive function that is unique to humans, but simply due to the capacity of animals—from insects to humans—to extract and categorize the visual characteristics of complex images.
Vision is one of the most important sensory modalities for the perception of biologically relevant stimuli. It is one of the major senses of insects like honeybees, and there is abundant evidence for the honeybee’s ability to quickly learn colours, shapes and patterns (von Frisch 1914, 1967; Zhang et al. 1995; Srinivasan 2010). However, simple visual cues rarely exist in nature: during their daily foraging trips, honeybees have to rely on a variety of complex visual cues from their environment in order to navigate, such as constellations of landmarks, multifaceted landscapes, and flowering trees (Collett 1996; Collett and Collett 2002; Collett et al. 2003; Steffan-Dewenter and Kuhn 2003; Dyer et al. 2008). This requires sophisticated visual processing and learning abilities. Indeed, bees have been shown to discriminate complex forest scenes (Dyer et al. 2008), be capable of categorizing images from natural scenes such as different flower shapes (Zhang et al. 2004), and most surprisingly, human faces (Dyer et al. 2005; Dyer and Vuong 2008; Avarguès-Weber et al. 2010). Furthermore, bees have been shown to display numerical processing abilities, solve delayed-matching-to-sample tasks, learn abstract rules and concepts, and transfer these to novel stimuli and tasks, even to different sensory modalities (Srinivasan et al. 1998; Giurfa et al. 2001; Giurfa 2007; Gross et al. 2009; Avarguès-Weber et al. 2011, 2012). These are remarkable capabilities for an insect, comparable to those of vertebrates. In spite of their small brain, honeybees have the capacity to process and learn complex visual information, which in turn facilitates efficient navigation and assists foraging in their ever-changing visual environment.
Although numerous studies have demonstrated that bees can learn much more than just simple patterns, colours and shapes, the cues that honeybees use to solve complex visual tasks are still a matter of debate. Some models assume that bee vision and visual learning is determined by mechanistic hardwired circuits, and that bees rely only on low-level feature detectors and elemental cues (Horridge 2000, 2005, 2009a, b). In this scenario bees learn combinations of coinciding elemental cues as retinotopic label for a particular image and generalize between images containing similar cues. This simple elemental processing, however, cannot explain how bees use previously acquired information to solve novel tasks, categorize novel stimuli that significantly differ in low-level cues, and transfer abstract concepts to novel domains. Therefore, other models suggest that honeybee vision and visual learning is a plastic system based on multiple mechanisms (Dyer and Griffiths 2012). Depending on the visual task at hand, honeybees may rely on simple, elemental processing if sufficient, however, with increasing complexity of the task and continued visual experience, honeybees can learn to move to non-elemental processing, such as configural type processing and rule-learning, and can access top-down information to solve novel tasks (Giurfa et al. 2003; Stach et al. 2004; Stach and Giurfa 2005; Giurfa 2007; Avarguès-Weber et al. 2010; Dyer 2012).
To further investigate the cues and mechanisms underlying honeybee visual learning, we asked whether the honeybee’s ability to discriminate between complex stimuli could be extended to the discrimination of paintings, which humans distinguish on the basis of artistic style—that is, Claude Monet paintings from the Impressionist period and Pablo Picasso paintings from the Cubist period. Previous work with birds has already demonstrated that the capacity for discrimination of artistic style is not limited to humans: Pigeons can learn to distinguish Monet from Picasso paintings, generalize to novel paintings by the same artist and even to paintings by other artists from the same period (Watanabe 2001; Watanabe et al. 1995). If honeybees were similarly able to distinguish multiple paintings by Monet and Picasso and then transfer this discrimination to novel paintings by the same artists, it would suggest that they are sensitive to the visual characteristics that are common to each style. With each painting being unique and differing in countless visual details from others even by the same artist, honeybees are unlikely to achieve generalization to novel paintings, if they rely only on simple elemental processing and retinotopic image matching.
Here, we investigate for the first time whether discrimination of paintings and generalization of artistic styles can be achieved by an insect that has a brain the size of a grass seed containing less than one million neurons. In a series of experiments, we tested whether honeybees could discriminate Monet from Picasso paintings at all; whether bees could learn to discriminate more than one painting pair at the same time; and whether bees could generalize their discrimination to novel paintings.
Materials and methods
Experiments used free-flying, individually marked honeybees (Apis mellifera) and were conducted in an indoor flight facility with controlled temperature and illumination at the Queensland Brain Institute, Australia, in accordance with the national guidelines and regulations.
For each experiment, two groups of 25 individually marked honeybees were trained separately to discriminate between a pair (or pairs, depending on experiment) of Monet and Picasso paintings. One group of bees was trained to Monet rewarded, the second group was trained to Picasso rewarded. Only one bee at a time was allowed to enter the tunnel, and the next bee was only allowed in once the first bee had been released from the chamber. This prevented any potential olfactory or social cues being released by a bee sitting on the feeder influencing the choice of the next bee. The feeder was exchanged for a clean feeder every 20 min (after each block, see below) to reduce the possibility of potential olfactory cues, such as pheromones deposited on the feeder that could influence the bees’ choices.
Training was conducted in blocks of 20 min. To prevent side preferences, the rewarded image was presented on the right side for 10 min, then on the left side for 10 min in random order. Each bee visited the tunnel at least once, usually 3–4 times during a block, but only the bees’ first choices with the rewarded image presented on the left and on the right side per block were used for analysis. The results for all bees of a group were pooled for a block, and the mean percentage of correct choices for each block (or set of five blocks, depending on experiment) was calculated. Data were analyzed using ANOVA and Fisher post hoc tests.
Mean luminance (lux) of the Monet and Picasso images used for discrimination training and testing for generalization to novel paintings in Apis mellifera honeybees
Training pair 1
Training pair 2
Training pair 3
Training pair 4
Training pair 5
Novel pair 1
Novel pair 2
Novel pair 3
Novel pair 4
Discrimination of Monet from Picasso
Discrimination of multiple paintings
After demonstrating that honeybees are capable of discriminating a Monet painting from a Picasso painting, the next question was whether bees could extend this capacity and learn to discriminate multiple painting pairs at the same time. When presented with five different pairs of Monet and Picasso paintings (Fig. 2a), honeybees learned to discriminate all five pairs, with percentage of correct choices increasing significantly over 5 days of training for all pairs (ANOVAday effect Monet rewarded F4,120 = 38.39, p < 0.001; Picasso rewarded F4,120 = 31.34, p < 0.001) (Fig. 3b). There were no differences in how well bees performed over the five different training pairs (ANOVApair effect Monet rewarded F4,120 = 1.49, p < 0.208; Picasso rewarded F4,120 = 0.51, p < 0.725).
Discrimination of paintings in greyscale
To investigate whether bees used colour as discrimination cue, we repeated the experiment using greyscale versions of the paintings. When the same groups of bees were presented with the greyscale versions of the five training pairs, their discrimination performance was as good as for the colour versions of the paintings, that is there was no effect of colour (ANOVAcolour effect Monet rewarded F1,48 = 2.589, p = 0.114; Picasso rewarded F1,48 = 1.495, p = 0.227) (Fig. 3d left). Again, there was no effect of training pair (ANOVApair effect Monet rewarded F4,20 = 1.49, p = 0.244; Picasso rewarded F4,20 = 0.365, p = 0.831).
Generalization to novel paintings
Lastly, we investigated whether honeybees could transfer knowledge about the visual structure that sets Monet’s paintings apart from Picasso’s to new images they had never encountered before. To this end, the same two groups of bees that had successfully learnt to discriminate five Monet–Picasso pairs, were shown four unrewarded, novel pairs of Monet and Picasso paintings (Fig. 2b), interspersed with blocks of rewarded training paintings (Fig. 2a). Honeybees continued to discriminate all training pairs of Monet and Picasso paintings, but did not seem to generalize to the novel pairs (Fig. 3c) (ANOVApair effect Monet rewarded F8,5 = 18.31, p < 0.003; Picasso rewarded F8,5 = 9.99, p < 0.011). For the Monet rewarded group, post hoc comparisons indicated that performance for all novel pairs was significantly lower than performance for the training pairs. However, for the Picasso rewarded group, the bees’ performance for novel pair 2 did not differ from training pair 3 (Fisher post hoc test; p = 0.060), suggesting that bees were able to generalize to novel pair 2. Notably, for both groups the percentage of correct choices for novel pairs was above chance (i.e., above 50 %) in six out of the eight tests, indicating that a weak generalization may have occurred.
We also presented the bees with novel paintings in greyscale, using the same procedure described above. Again the honeybees performed well when they discriminated the greyscale versions of the five training pairs, and their performance declined when they were presented with novel paintings (Fig. 3d, right). However, generalization to the novel pairs was better when the paintings were presented in greyscale, with only marginal or no significant difference between training pairs and most novel pairs (ANOVApair effect Monet rewarded F8,5 = 5.09, p < 0.045; Picasso rewarded F8,5 = 2.69, p < 0.145). Post hoc tests revealed that for the Monet rewarded group bees’ performance, for the greyscale novel pairs 1, 2, and 4, did not differ significantly from training pairs 3 and 5 (Fisher post hoc tests; range of p values 0.057–0.140). For the Picasso rewarded group, bees’ performance for all novel pairs could statistically not be separated from at least two of the training pairs (Fisher post hoc test; range of p-values 0.056–0.245).
There have been previous studies investigating how bees respond to artistic paintings, showing that naïve bumblebees are attracted to paintings displaying flowers (Chittka and Walker 2006, 2007). However, our study is the first to investigate bees’ ability to discriminate and generalize between artistic styles. We show that honeybees can distinguish between Monet and Picasso paintings, and that they even learn to discriminate several painting pairs simultaneously. Considering the complexity of the paintings and the fact that these stimuli are of no biological relevance to honeybees, our results illustrate the extent of bees’ visual capacities and impressive pattern recognition abilities (Gould 1985, 1986; Chittka et al. 2003; Giurfa 2007). It suggests that honeybees can learn to discriminate between many other complex images irrespective of biological relevance, and supports earlier studies showing that honeybees have the capacity to learn and distinguish multiple complex stimuli (Zhang and Srinivasan 2004; Reinhard et al. 2006; Srinivasan 2010).
What cues do honeybees use for painting discrimination?
The painting pairs were matched for luminance, where each Monet painting had similar mean luminance to its respective Picasso partner both as colour and greyscale versions. Also, the mean luminance for all pairs was in the same range, apart from training pair 5, which was slightly dimmer in the colour version (Table 1). One might argue that these measurements were based on human perceptual function and visual perception of absolute luminance between bees and humans may differ. Although bees might indeed perceive the images brighter or darker than we do, it does not change the relative measure, which demonstrates that the two images of a pair had similar mean luminance irrespective of their absolute perceived luminance. Thus, the bees would have difficulty discriminating between paintings on the basis of luminance alone.
Colour seems an obvious cue for discriminating between paintings. We therefore had matched the Monet and Picasso paintings in each pair according to their overall colour appearance—as much as possible considering the complexity of the images. However, with each painting being a unique piece of art, and therefore differing in countless colour details from its partner, the honeybees could potentially rely on specific colour cues within each painting when distinguishing one from another. Bees have the capacity to store multiple complex memories of signal combinations at any one time during their foraging trips, such as combinations of colours, scents and locations (Reinhard et al. 2006); hence it is conceivable that they might have simply memorized different colour cues and colour combinations for each individual painting and relied exclusively on this information during discrimination. However, the experiments using greyscale versions of the paintings showed that bees easily transferred their acquired knowledge from the colour pairs to the greyscale versions, therefore colour per se is not crucial for discrimination. Our findings are in line with earlier work showing that discrimination and categorization of landscapes, flowers and plant stems were not compromised when the colours of the stimuli were removed (Zhang et al. 2004).
There is the possibility that bees may use other elemental cues to discriminate Monet from Picasso paintings, such as salient edges, which bees learn very well (Horridge 2007). Monet paintings in general have less salient edges than Picasso’s. However, the specific Monet images we used for this experiment all display salient edges both vertical and horizontal, particularly strong in training pairs 1, 2, and 5 (Fig. 2a). Also, considering the number and complexity of salient edges in the images, and the fact that each painting has different salient edges, it seems unlikely that bees use the complex arrangement of edges in each image as retinotopic label to identify and distinguish between the paintings.
The fact, that bees can learn to discriminate not one but several painting pairs simultaneously, indicates that bees may learn about the categorical structure of the paintings rather than the specific cues of single exemplars. This hypothesis is supported by the finding that the bees were able to generalize their discrimination to novel paintings at least to some extent.
Generalization to novel paintings
Generalization is a fundamental cognitive capacity that allows classifying or categorizing similar stimuli according to shared characteristics, treating similar stimuli as equivalents, and thus responding to them in the same manner (Zentall et al. 2008). Generalization across visual stimuli is a well-known ability in honeybees (Wehner 1971; Zhang et al. 2004; Stach et al. 2004; Lehrer and Campman 2005; Stach and Giurfa 2005; Gross et al. 2009; Horridge 2009a). It enables bees to successfully forage in an ever-changing environment, since it allows adaptive responses to novel objects. During foraging, honeybees learn the characteristics of rewarding flowers and use this knowledge not only to recognize the same flowers, but also respond to new ones with similar characteristics (Giurfa and Lehrer 2001). That is, honeybees can form categories of a type of flower or object based on a range of similar characteristics that are shared among the members of the same category. Indeed, Zhang et al. (2004) have shown that honeybees can be trained to group different naturally occurring objects, such as landscapes, plant stems and flower shapes into distinct categories, and then discriminate novel visual objects according to these categories even if the novel objects greatly differ in their individual features.
Our study showed that honeybees also discriminated some of the novel paintings they had never encountered before, although statistically the evidence for generalization was not strong. This suggests that honeybees have some ability to learn about the general visual structure that sets Monet’s paintings apart from Picasso’s, and generalize this knowledge to novel paintings by the same artists—in particular when the paintings are presented in greyscale. Colour may have affected the bees’ ability to generalize to novel paintings, as the novel paintings differed in countless colour details from training paintings by the same artists. The new colour cues may have captured the bees’ attention and distracted them from recognizing the shared characteristics. The absence of colour cues in the greyscale versions facilitated generalization to novel paintings.
The reason why generalization to novel paintings in bees was not more pronounced may lie in the training regime. Individual training length and procedure (absolute vs. differential conditioning) are known to improve discrimination and generalization of visual stimuli in honeybees, with bees moving from feature-extraction mechanisms to configural type processing with increasing experience (Giurfa et al. 2003; Stach et al. 2004; Dyer et al. 2005; Dyer and Vuong 2008; Stach and Giurfa 2005; Giurfa 2007; Avarguès-Weber et al. 2010). It is possible that a different protocol, where individual honeybees are presented with more exemplars of Monet and Picasso paintings over a longer period of time, and are tested with only one novel pair per day, would improve generalization performance to novel paintings. Indeed, similar experiments using pigeons have shown that hundreds of exemplars of a category and weeks of training are needed to achieve significant generalization to novel paintings (Watanabe et al. 1995). Due to the limited life span of insects, it is however difficult to conduct equivalent experiments in honeybees.
What cues do honeybees use for categorizing paintings?
As each painting is a unique piece of art composed of lines, shapes, edges and colours that are arranged differently, bees could not have used painting-specific information for categorization and generalization to novel paintings. Which cues then could honeybees use to characterize a ‘painting style’? Honeybees are thought to use a range of stimulus features such as symmetry or orientation of objects, as well as layout of the features for categorization and generalization of visual stimuli (Stach et al. 2004; Benard et al. 2006). It is conceivable that bees learned the characteristics that are shared across the paintings of one category, such as configurations of shapes (e.g., shapes predominantly found in the image centre vs. shapes predominantly found in the image periphery vs. shapes evenly distributed across the image), orientation of objects (e.g., predominantly horizontal vs. predominantly vertical), or salient lines and edges (e.g., high load of salient edges vs. low load of salient edges). However, establishing these categories would require all Monet images used in this study to be significantly different from the Picasso images with respect to such characteristics, which was not the case (Fig. 2). Indeed, when the paintings were convoluted using horizontal and vertical filters to determine whether the majority of features and boundaries within Monet paintings compared to Picasso paintings were oriented in a particular way, we found that both artistic styles contained similar orientational information, with countless vertical, horizontal and diagonal structures in each (Online Resources 1 and 2). Due to the complexity of the paintings, there was no distinct orientational information inherent in all Picasso paintings compared to all Monet paintings, which the bees could have used as global feature for discrimination and categorization of artistic style.
Our study does not yet provide a conclusive explanation of how bees solve the task of discriminating and categorizing complex images. Based on our data, however, we can assume that honeybees do not rely on particular elemental features such as luminance, colour, salient edges, orientation or spatial frequency content. It is more likely that they use feature extraction and/or configural processing to learn the visual characteristics that are shared across the paintings of one category, which is consistent with the way honeybees are thought to process human faces, forest scenes and landscapes, and solve novel visual tasks (Stach et al. 2004; Zhang et al. 2004; Stach and Giurfa 2005; Dyer et al. 2005, 2008; Dyer and Vuong 2008; Avarguès-Weber et al. 2010). Thus, our study contributes to the growing body of evidence that insects like honeybees have the ability to learn, retain, classify and process visual information in a way that is not predicted by simple mechanistic or elemental responses to stimuli (Dyer 2012). Of course, the fact that bees can discriminate paintings does not imply that bees actually recognize an artistic style per se or interpret art content at a semantic level in a similar way to humans. But it suggests that discrimination of artistic styles is not a higher cognitive function that is unique to humans, but may simply be due to the capacity of animals, from insects to humans, to extract and categorize the visual characteristics of complex images. The fact that bees (or pigeons) can distinguish art is surprising only to the extent that one believes their discriminations are based on local elemental features within each painting. Artistic style, however, is based on information that is shared across paintings. Future work will show whether bees are indeed sensitive to redundant visual information that is consistent across several images, which is captured by covariance measures such as SVD.
We thank Chin Y. J. Yuen for help with the experiments, and Adrian Dyer and Allen Cheung for advice on the image analyses. W.W. was funded through an Australian Postgraduate Award by the Australian Government and an Australian Research Council Discovery grant to JT (DP0985830). A.M.M. was funded by a FAPESP doctorate scholarship (08/50576-8), Brazil.