Learning complex stimuli with a honey bee brain: contrasting consequences on performance and flight behaviour from degrading configural versus featural information

Learning complex stimuli with a honey bee brain: contrasting consequences on performance and flight behaviour from degrading configural versus featural information | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Learning complex stimuli with a honey bee brain: contrasting consequences on performance and flight behaviour from degrading configural versus featural information Marie-Genevieve Guiraud, Adam D. Hines, Mengjie Jin, Andrew B. Barron This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9487302/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 5 You are reading this latest preprint version Abstract Honeybees ( Apis mellifera ) impress with their visual learning abilities despite their miniature brains and relatively coarse vision. A striking demonstration of bee cognition shows that they can rapidly learn to discriminate human faces, despite these stimuli being ecologically irrelevant. To probe the mechanisms underlying this ability, we assessed bees’ capacity to learn and categorise faces selected from the Chicago Face Database (CFD). Once trained, bees were further tested with stimuli that degraded either the featural or configural information in the face stimuli. The generalisation test examined whether trained bees could successfully discriminate novel faces from the two trained categories: degrading featural information did not degrade configural information in faces. The inverted test offered bees the two faces used in training but rotated by 180°: degrading configural information but not featural information. In all tests, we recorded bees’ performance and flight path. Bees successfully generalised their learning to novel faces, and performance in the generalisation test correlated with the amount of training. Most trained bees failed the inverted test, however. In all tests, bees’ flights were faster and more linear than naïve bees. We propose that during training bees learned a flight strategy that aided their discrimination of faces, but as a consequence, bee learning of complex stimuli is more sensitive to configural changes that changed the spatial ordering of features than featural changes. Apis mellifera active vision visual cognition insect cognition 2D tracking face recognition Figures Figure 1 Figure 2 Figure 3 Figure 4 Introduction The honey bee has become an important model for visual recognition and visual cognition (Giurfa et al. 2003, Srinivasan 2010 ). Compared to humans and most mammals, honey bees have limited eye resolution and minute brains, yet bees can discriminate shapes, colours, spatial configurations and even human faces under some conditions (Dyer et al. 2005 , Avarguès-Weber et al. 2010 ). This paradox—high-level visual performance with minimal neural resources—has fuelled interest in identifying the mechanisms by which bees acquire, encode and use visual information. The face case is especially interesting because face discrimination is a subtle multi-element comparison that, for humans, involves both featural (individual features) and configural information (the spatial relationships among those features; Maurer et al. 2002 , Richler et al. 2014) and it is important to note that for bees, learning faces is outside of their cognitive ecology. In this study, we explored how bees learn these complex visual stimuli by testing their sensitivity to manipulations that degraded either the configural or the featural information in faces. Bees show striking abilities in visual learning and categorisation (Giurfa 2013, Chittka 2017 ). Bees can discriminate complex natural scenes (Zhang et al. 2004 ), learn abstract relationships such as same/different, and encode global relational properties of visual patterns even when composed of simple geometric elements (Avarguès-Weber et al. 2010 , 2011 , 2012 ). Although insect vision lacks both a retina capable of high spatial resolution and a cortex for higher-order visual object recognition (Kanwisher et al. 1997 , Kanwisher 2000 ), bees nonetheless achieve robust visual recognition using compact neural circuits and highly efficient active-vision strategies—namely, the integration of movement, selective sampling and iterative evidence accumulation (Srinivasan & Lehrer 1988 , Lehrer 1994 , Baird et al. 2013 ). Dyer et al. ( 2005 ) first showed that honey bees can discriminate individual human faces when trained on whole-face images. This study raised interest in how bees might be solving complex discriminations such as this. Face perception has been considered a hallmark of primate cognition (Kanwisher et al. 1997 , Tsao & Livingstone 2008 ). Humans and other primates possess highly specialized neural systems for detecting, encoding and individuating faces, relying on both featural information (eyes, nose, mouth) and the spatial relations among features—so-called configural or holistic processing (Kanwisher 2000 , Maurer et al. 2002 ). The classic face-inversion effect (Yin 1969 ), in which inversion of faces slows recognition more than inversion of most objects, exemplifies the sensitivity of primates to the orientation-dependent spatial relations of facial features. Extensive developmental, neurophysiological and psychophysical research demonstrates that experience strongly shapes the strategies humans use to extract identity-relevant cues (Carey & Diamond 1997 , Peterson & Eckstein 2013 ). In Dyer’s original study (Dyer et al. 2005 ), five bees were first pre-trained to associate a rewarded human face with sucrose against a schematic distractor, before being trained to discriminate a target face from other similar face images. Following differential conditioning, bees successfully discriminated the rewarded face from novel distractor faces, indicating generalisation across exemplars. However, performance dropped markedly when both target and distractor faces were presented inverted, suggesting that recognition relied on orientation-dependent visual cues. This finding can be interpreted as being consistent with sensitivity to configural information in faces, although inversion effects alone cannot distinguish between configural and feature-based strategies. Accordingly, this interpretation has been questioned (Pascalis et al. 2006 ). Configural processing encompasses different levels of relational encoding, from basic first-order arrangements of features to complex metric and holistic representations, which are typically considered more demanding than simple feature-based learning. In this context, and in the absence of dedicated face-specific neural circuits comparable to those described in primates, it remains unclear to what extent bees rely on genuinely configural representations as opposed to alternative, simpler strategies. A follow-up study by Avarguès-Weber et al. (2013) showed that honeybees trained on minimalistic, schematic “face-like” stimuli can learn first-order relational (configural) properties between stimulus elements. Bees trained to recognise a canonical arrangement of features were able to generalise this learning to novel stimuli, discriminating face-like configurations from scrambled arrangements composed of the same elements (Avarguès-Weber et al. 2013). This provides strong evidence that bees can encode spatial relationships between features rather than relying solely on local cues. However, these results are limited to first-order relational structure and much remains to be understood about the mechanisms that allow bees to rapidly learn and exploit such structured visual information. Bees’ movements can yield information on how bees learn. Bees do not passively receive visual input and instead continuously sample a scene through structured flight manoeuvres that shape what information reaches the eye (Srinivasan 2010 , Guiraud et al. 2018 , 2024 , 2025 , MaBouDi et al. 2023 ). When approaching complex stimuli, bees exhibit characteristic saccade-like turns, scanning loops, and speed modulations that guide visual processing to facilitate evidence accumulation during decision-making (Braun et al. 2012 , Lehrer 1994 ). Bees fly close to specific regions, follow contours, oscillate laterally, or sweep across edges in ways that maximise contrast detection and minimise processing costs (Lehrer et al. 1995 ; Dittmar et al. 2010 ). These movements generate temporal sequences of visual features that the mushroom bodies can associate with reinforcement (Menzel & Giurfa, 2001 ). Detailed analyses of bee flight during pattern discrimination tasks showed that trained bees shifted from slow, rotational sampling during unfamiliar stimuli to faster and more directed trajectories once they recognise a target (MaBouDi et al. 2025 b ). Therefore, investigation of bees’ flight behaviours can reveal whether or not a stimulus is familiar. Here, we asked whether bees trained on a single pair of human faces selected from different categories within the Chicago Face Database (CFD) would (i) generalise their learning to a test in which both alternatives were novel faces, (ii) be sensitive to inversion (using the same two stimuli used in training but inverted at 180°), which disrupts the trained spatial arrangement of features, and (iii) show corresponding changes in flight behaviour indicative of a learned active-sampling strategy. By combining performance measures with flight kinematics, we aimed to determine whether bees could solve these discriminations, and whether successful recognition in conditions where the featural (i) and configural information (ii) was degraded is accompanied by distinct signatures of learned visual sampling. Material and methods Experiments were conducted over three consecutive summers (2021–2024). A single hive was placed within a 160 m 2 enclosed shade house for 1-2 months, depending on the hive’s health. Foragers from this hive accessed a communal gravity feeder containing 30% (w/w) sucrose solution positioned 1.5 m from the hive. Individual bees were trained from the gravity feeder to fly into a testing arena measuring 60 cm (L) × 60 cm (W) × 40 cm (H) (Fig. 1). Six vertically mounted Eppendorf tubes (experimental feeders) were affixed to the rear wall beneath six visual stimuli. Each feeder dispensed 15 μl of 50% (w/w) sucrose solution—below the bee’s crop capacity—to encourage multiple feeder visits within a trial in the testing arena. Stimuli were real human faces (self-reported sex and ethnicity) in shades of grey against white background from the Chicago Face Database (CFD) (https://www.chicagofaces.org). For the training and tests, we used faces from two different categories in the CFD: self-reported women faces from self-rated Asian ethnicity and self-reported men faces from self-rated Latino ethnicity. Individual pre-training protocol To recruit single foragers to the testing arena, a returning bee at the gravity feeder was offered a 50% (w/w) sucrose-soaked cotton swab. Once the bee began feeding, the experimenter gently transferred it to one of six Eppendorf-tube feeding stations in the testing arena, simultaneously marking the bee’s dorsal abdomen with a unique coloured dot (Posca pens, Uni-Ball, Japan). During pre-training, the stimuli wall displayed six uniformly grey, oval-shaped discs (similar in size to the faces). Only one bee was trained at a time. This procedure was repeated until the bee learnt to fly directly to the feeders in the testing arena. Typically, a bee needed two or three transfers to the testing arena before she returned of her own volition. The experimenter ensured that the selected individual visited each of the six feeders in the testing arena at least once. This phase familiarised the bee with the spatial layout of the feeders. All feeders contained an identical sucrose reward at this stage. Unmarked bees were removed from the testing arena. Training Once bees consistently visited all six feeder locations during pre-training, we initiated differential conditioning with two different faces drawn from the CFD. In each trial, six faces were displayed – two distinct faces, one from each of our selected categories from the CFD, three copies of each - on the rear wall in randomised positions. Bees in group A received 15 μL of 50% (w/w) sucrose solution when landing on feeders beneath Asian women (CS+) faces and 15 μL of saturated quinine solution (0.12% w/w) when landing on feeders beneath Latino men (CS–) faces. Training group B was trained with the reciprocal contingencies. Randomising stimuli positions each trial ensured that bees relied on visual features rather than location. To eliminate residual odours and chemical cues between trials, the arena and stimuli were thoroughly wiped with 70% ethanol. Bees are known to leave pheromone-laden footprints or scent marks during foraging and exploration, which can influence the behaviour of subsequent individuals if not removed (Chittka et al. 1999, Goulson et al. 2000). Ethanol is widely used in bee cognition studies as an effective cleaning agent to remove such olfactory traces. To ensure that ethanol itself did not interfere with the bees’ behaviour, a minimum of 2–3 minutes was observed after cleaning, allowing sufficient time for complete evaporation before the next bee was introduced (Giurfa et al. 1999). Each trial encompassed the full duration of a bee’s visit to the testing arena, beginning with its entry into the arena and concluding with its departure. During this period, a bee could land on multiple feeders, with each landing recorded as a choice. Individual bees typically made three to five landings per trial. Training continued until each bee achieved ≥ 80% correct choices (landing on the rewarded faces) across its most recent 20 landings. Bees typically required 25–35 trials to reach criterion, which would take 1-2 days of training. Testing After reaching the training criterion, we conducted three unrewarded test types: learning test (LT), inverted test (IT), and generalisation test (GT; Fig. 2). In all tests, all six feeders offered distilled water. In the learning test, bees were presented with the original six stimuli. This test assessed whether bees had genuinely learned to use the previously rewarded face stimulus to perform the discrimination. In the inverted test, the same six faces were shown inverted (180° rotation) to assess sensitivity to configural changes. A drop in performance compared to the learning test would indicate reliance on canonical face orientation in training. In the generalisation test, a new self-assigned gender and ethnicity Asian woman and Latino man face was presented matched for greyscale and relative size to evaluate whether bees could extend the learned discrimination to new exemplars. Each test lasted 2 minutes, during which we recorded the number of landings on the previously rewarded (correct) versus previously punished faces (incorrect). To sustain motivation, unrewarded tests were interleaved with refreshment trials using the original training contingencies (sucrose on CS+; quinine on CS–). Bees were required to achieve ≥ 80 % correct in a refreshment trial before attempting the next unrewarded test, typically performing one to three refresher trials between tests. By comparing performance across these tests, we quantified both retention of the learned discrimination and the flexibility of visual generalisation under altered presentation conditions. A total of 33 bees completed at least the training phase. Sample size varied across analyses because not all individuals completed all training blocks and subsequent tests. Sixteen bees completed the training and the first learning test; fourteen completed the training, the first learning test, and the inverted test; and twelve completed the training and all three tests. Video analysis: 2D bee tracking We recorded flight trajectories and scanning behaviour using an iPhone X camera (1,280 × 720 px at 120 fps) opposite the stimulus wall. Recording began as soon as the marked bee entered the arena and continued until departure, generating one video file per trial. For the bee tracking experiments, we utilised the Segment Anything 2 (SAM2) model and SAMURAI for zero-shot visual tracking and motion-aware memory to track bees over several frames (Yang et al. 2024, Ravi et al. 2024). The SAM2 model used was the SAM2.1 Hiera Large for the highest accuracy segmentation and was fine-tuned on 100 examples of bees in the experimentation arena. Ground truth segmentation masks were annotated manually using ImageJ. The fine-tuned code, model including the images, masks, and bounding boxes used for fine-tuning are available at https://huggingface.co/AdamHines/samurai-beetracker. Individual frames for segmentation were derived from the experimental recordings at a rate of 24 frames per second and divided into 1,000 frame chunks for processing to circumvent computational memory limitations (Ravi et al. 2024). Initial frames per chunk were selected and a bounding box was manually selected around the bee for the starting segmentation. Once selected, the SAMURAI system performed motion-aware tracking across the rest of the frames (Yang et al. 2024). In order to carry segmentation information from one chunk to the next, the bounding box from the last frame of the preceding chunk was used to initialise the segmentation of the first frame of the next chunk. Once all chunks were processed, the entire trajectory was stitched together for analysis. Statistical analysis All statistical analyses were conducted in R (v4.x). Learning performance was analysed using a Generalised Linear Mixed Model (GLMM; package lme4 ) with a binomial error structure, where the response variable was the number of correct versus incorrect choices per block of 10 trials (successes/failures). Fixed effects included training block and rewarded stimulus category (stimulus × protocol), and bee identity (Bee_ID) was included as a random intercept to account for repeated measures and inter-individual variability. Model assumptions and rank deficiency were evaluated through inspection of residuals and by verifying collinearity among predictors. Post-hoc comparison of learning rates between women-rewarded and men-rewarded groups was assessed by comparing a reduced (Block + Group) and a full (Block × Group) GLMM using likelihood-ratio tests. For unrewarded tests, bees’ performance (proportion of correct choices) was compared against chance level (0.5) using Wilcoxon signed-rank tests, as normality assumptions were not met. Differences between treatment groups (female-rewarded vs male-rewarded) were evaluated using Wilcoxon rank-sum tests or Kruskal–Wallis tests when more than two groups were compared. For each test type (learning, inverted, generalisation), effect sizes, test statistics, and exact p -values were reported. Relationships between training length (total number of training trials) and test performance were examined using Pearson correlations with pairwise deletion of missing values. Flight-tracking data obtained from the SAM2/SAMURAI segmentation pipeline were analysed separately. We reconstructed flights within a narrow region of interest around each of the face stimuli. This was defined as a rectangle centred on the face and extending beyond it by approx. 40 pixels in each direction. For each flight segment within the region of interest of the face stimulus, the mean translational speed (px/s) and angular velocity (rad/s) were extracted. Differences in speed and angular velocity between experimental conditions (naïve, learning test, inversion test, generalisation test) were first analysed using a one-way ANOVA when assumptions of normality and variance homogeneity were met. Otherwise, a Kruskal–Wallis test was used. Significant omnibus tests were followed by Tukey’s HSD (for ANOVA) or Dunn’s post-hoc comparisons with Holm correction (for non-parametric tests). Data are reported as means ± SEM unless stated otherwise. All tests were two-tailed and significance thresholds were set at α = 0.05. Results Bees’ performance in face recognition task All 33 bees included in the training analysis learnt the task to criterion (Fig. 3a). Training of bees stopped when an individual made 80% (or more) correct choices within the last 20 choices, therefore the number of training choices differed for each bee. Bees progressively increased their proportion of correct choices across training, as reflected by a significant positive effect of training block in the GLMM analysis (Table S1; Fig. 3a). The response variable was the number of correct choices within each block of 10 trials, modelled as a binomial response (successes versus failures). The model revealed a strong effect of training block on number of correct choices per block (estimate = 0.0168 ± 0.0037, z = 4.55, p = 5.34 × 10⁻⁶), indicating performance improvement with training. Neither training group (A vs. B) nor its interaction with training block explained additional variance (Table S1). An additional Wilcoxon test showed the significant difference between the two last bouts (Wilcoxon tests: p < 0.01, p < 0.001). Bees performed significantly above chance during the unrewarded learning test (Fig. 2c), confirming that they remembered the rewarded stimulus in an unrewarded test (Wilcoxon signed-rank test: V = 325, P = 6.36 × 10⁻⁶; Fig. 3b and Table S3). There was no significant difference in LT performance between bees trained in groups A or B (Wilcoxon rank-sum test: W = 60.5, P = 0.435). In the unrewarded inverted test IT (Fig. 2d) bees did not exhibit a preference for the previously rewarded stimulus (Wilcoxon signed-rank test: V = 101, P = 0.405; Fig. 3c). Performance did not differ between training groups (Kruskal–Wallis test: N = 14, H = 0.161, P = 0.688; Table S4). These results indicate that bees were unable to recognise the trained face when inverted. In the unrewarded generalisation test (GT) (Fig. 2e), bees showed a significant preference for the face from the same CFD category they were trained on (Wilcoxon signed-rank test: V = 98, P = 0.015; Fig. 3b). Performance did not differ between training groups (Kruskal–Wallis test: N = 12, H = 0.0017, P = 0.967; Table S5). Bees’ individual differences and correlations Bees exhibited marked individual variation (Fig. 3b) in their learning performance and in their ability to generalise facial information, despite the clear group-level patterns observed in the training and test phases. During acquisition, some individuals rapidly reached high accuracy, whereas others required substantially more trials to improve, leading to a broad range of total training lengths (from ~60 to >300 choices; Table S7). These differences in training were not predictive of performance in the Learning Test (LT: r = –0.33, P = 0.113; Table S6) or in the Inverted Test (IT: r = 0.18, P = 0.415; Table S6), where performance remained tightly clustered around the group means. However, training length was positively correlated with performance in the Generalisation Test (GT: r = 0.56–0.62, P = 0.012–0.004; Table S6), indicating that individuals exposed to more training trials developed more robust or more flexible face representations. These behavioural divergences were also reflected in individual outcomes: some bees completed very few choices during trials ( e.g., Bee_18, 60 choices), while others exceeded 300, and several individuals achieved perfect accuracy during the Learning Test (Bee_10, Bee_21, Bee_25). In contrast, a few bees excelled specifically in the Inverted Test (Bee_8, Bee_33), suggesting differential sensitivity to stimulus rotation. The most pronounced variability emerged in the Generalisation Test, where a subset of “high-level generalisers” ( e.g., Bee_3: 100%; Bee_6: 80%; Bee_4 and Bee_24: 75%) displayed exceptionally strong performance, whereas others remained at or near chance (e.g., Bee_13, Bee_14, Bee_16, Bee_33, Bee_9). Flight and scanning behaviour To determine how bees sampled the face stimuli during decision making, we reconstructed their flight trajectories and quantified their movement dynamics across all test conditions (Fig. 4a and b). Flight paths obtained using a semantic-segmentation-based tracking pipeline revealed systematic differences between naïve bees and bees that had undergone learning (Fig. 4a-b; see Methods). Naïve bees approached the face stimuli slowly with highly rotational flight, exhibiting the lowest speeds (mean = 69.37 px/s, n = 12) and the highest angular velocities (mean = –2.55 rad/s, n = 12; Fig. 4a–b) as they scanned the novel stimuli. There is a global effect of the conditions on flight speed (Kruskal-Wallis global test H (3)=13.55, p<0.003) but not for angular velocity (Kruskal-Wallis global test H (3)=6.056, p=0.1094). After learning (during the learning test LT), bees showed a marked shift in scanning behaviour. Flight speed increased significantly (mean = 98.6 px/s, n = 12; Kruskal–Wallis with Dunn’s post hoc test LT vs . N: z= 2.668, p = 0.008), and angular velocity was strongly reduced (mean = –0.49 rad/s, n = 12; Kruskal–Wallis with Dunn’s post hoc test LT vs . N: z= 2.403, p = 0.016), indicating straighter, and faster trajectories. In the Inverted Test IT, flight speed increased further (mean = 127.07 px/s), significantly higher than in naïve bees (Kruskal–Wallis with Dunn’s post hoc test IT vs . N: z= 3.479, p = 0.0005), although not significantly higher than the learning test LT (Kruskal–Wallis with Dunn’s post hoc test LT vs. IT: z= 0.794, p = 0.427; Fig. 4b). Angular velocity remained similar to the learning condition (Kruskal–Wallis with Dunn’s post hoc test LT vs. IT, z= -0.802, p = 0.422), but was still lower—although not significantly—than in naïve bees (Kruskal–Wallis with Dunn’s post hoc test IT vs. N, z= 1.584, p = 0.11), In the Generalisation Test GT, bees flew at speeds comparable to the learning test LT (mean = 97.09 px/s) and significantly faster than naïve bees (Kruskal–Wallis with Dunn’s post hoc test GT vs. N: z= 2.342, p = 0.019). Angular velocity in this condition was low and did not differ from the learning test LT or the inverted test IT (Kruskal–Wallis with Dunn’s post hoc test GT vs. LT: z= -0.982, p = 0.325 and GT vs. IT: z= -0.163, p = 0.871; Fig. 4b). Discussion Our findings build on a growing body of work showing that bees can perform sophisticated visual discriminations despite small brains and coarse visual resolution (Guiraud et al. 2018 , 2022 , 2024 , 2025 , MaBouDi et al., 2023 , 2025 a,b ). Consistent with previous studies showing that bees can recognise complex visual patterns, including faces and face-like stimuli, as holistic configurations (Dyer et al. 2005 ; Avarguès-Weber et al. 2010 ), bees rapidly learned to discriminate between two different human faces. As a population, bees performed significantly above chance in the unrewarded learning test (Fig. 3 b). All tested bees chose the face rewarded in training more than 50% in the test, and some bees chose the rewarded face 100% in test (Fig. 3 b and Table S7). We also noted clear changes in bees’ flight behaviour as they learned the visual discrimination task. Naïve bees flew slowly and with highly rotational trajectories, consistent with exploratory sampling during unfamiliar visual tasks (Guiraud et al. 2025 ; MaBouDi et al. 2025 a,b ). After training to an 80% performance criterion, bees exhibited faster, more linear flight (Fig. 4 a-b, Supp. Figure 1 ), possibly reflecting a transition to more directed sampling and stable evidence accumulation. This change mirrors findings by MaBouDi et al. ( 2023 ) showing that bees gather diagnostic information through dynamic, efficient scanning routines when familiar with a visual task. The changes in flight behaviour we saw were unlikely to be adaptations to the flight arena because we considered flights within a small region around each face only and not globally within the arena. Once trained with just a single pair of faces, bees were able to generalise their learned face information to successfully discriminate between two novel faces (GT, Fig. 3 b). Although novel faces differed from those experienced during training, bees significantly preferred the face from the same CFD category as the one rewarded in training (Fig. 3 b). In our generalisation test, both faces were new to the bee. This differs from Dyer et al. ( 2005 ) who tested whether bees could distinguish a learned rewarded face from a novel distractor. We believe this is the first test of whether bees can distinguish between two novel faces based on similarity to trained exemplars. Generalisation is a robust feature of bee visual learning and has been documented for naturalistic scenes and visual objects, as well as for abstract and relational pattern learning (Zhang et al. 2004 , Dyer et al. 2008 , Avarguès-Weber et al. 2011 , 2012 ). Because both test stimuli were novel, successful performance in the generalisation test is difficult to explain as mere recall of one rewarded training image. While bees performed significantly above chance in the generalisation test, their performance was less robust than in the learning test (Fig. 3 b). Generalisation performance varied between individuals, with a subset achieving exceptional accuracy ( e.g. Bee_3, Bee_6, Bee_4, see Table S7). Performance in the generalisation test — but not in the learning or inverted tests (Fig. S2) — was positively correlated with the training level, indicating that extended training facilitates generalisation of learned complex stimuli in bees. Flight speed and angular velocity in the generalisation test did not differ from the learning test, indicating that bees used the same learned flight strategy in the learning test and generalisation test (Fig. 4 a and b). Our generalisation test degraded featural information in the learned face discrimination but did not alter configural information (Maurer et al. 2002 , Richler et al. 2014). Even though honey bees were given relatively short training on a single pair of faces and were not trained to generalise learning to a category, it is noteworthy that bees could still generalise face learning to other faces from the same CFD category. Our inversion test degraded configural information but not featural information (Maurer et al. 2002 , Richler et al. 2014), and in this test bees performed at chance levels. Therefore, we conclude that for bees, face inversion abolished discrimination. In humans and other primates, performance in face recognition tasks is slowed and decreases when faces are inverted, but it is not eliminated (Yin 1969 , Diamond & Carey 1986 ). Clearly, changing the spatial configuration of features had an especially strong effect on bees’ ability to recognise learned faces. Bees have the capacity to learn configurations as it was shown by Avargues- Weber et al. (2011) who used abstract stimuli to isolate and test the capacity of bees to learn configurations. Our inverted test offered bees intact featural information but despite this the inversion reduced performance to chance levels. Dyer et al. ( 2005 ) also reported that inversion eliminated recognition of a learned rewarding face in honey bees. We noted that in the inversion test bees’ flight speed and angular velocity remained comparable to that seen in the learning test rather than resembling the slower, more rotational flights of naïve individuals (Fig. 4 a). This dissociation suggests that bees retained and deployed a learnt movement strategy for inspecting face stimuli, but that this strategy was no longer effective when the faces were inverted. The most plausible interpretation is that, during training, bees established an active sensorimotor routine that sampled facial information in a particular spatial and temporal sequence. When the stimuli were turned upside down, the same learnt routine would not match the learnt arrangements; preventing successful matching and disrupting the configural structure on which the recognition depended. Inversion would disrupt a learnt sensorimotor-perceptual loop, and we argue this is why inversion, but not generalisation, had such a dramatic effect on bee face recognition. Our data support theories of active vision: that bees use learned movements to aid learning and recognition of complex stimuli. Our findings reinforce emerging views that bees use efficient sampling strategies and robust neural representations to solve perceptual challenges normally associated with vertebrates (MaBouDi et al. 2025 a,b , Avarguès-Weber & Giurfa,2013, Guiraud et al. 2018 ). Interestingly, the inversion condition elicited the fastest flights of all, a pattern previously described in other forms of perceptual dissonance in bees ( e.g. Guiraud et al. 2025 ), potentially reflecting heightened uncertainty or the use of a default “rapid pass” strategy when trained relational cues fail. In contrast, generalisation flights closely resembled those from the learned condition, suggesting that bees recognised enough structural similarity to maintain an efficient scanning mode. Together, these results support the hypothesis that bees rely on active vision — combining selective scanning motions with efficient neural coding — to process visually complex stimuli. The fact that bees can learn, differentiate, and generalise human faces despite coarse spatial resolution indicates that high-level visual recognition does not require large brains, but rather efficient computation paired with informative sampling strategies. This aligns with recent work showing that small networks can replicate key aspects of insect visual decision-making (MaBouDi et al. 2023 , 2025 a,b ), and suggests that simple, energy-efficient architectures may suffice for tasks traditionally associated with higher cognition. Finally, our findings have significant implications for the comparative study of vision and for bio-inspired artificial intelligence. Bees’ ability to extract structural features from complex patterns, even across novel identities, offers a natural model for efficient object recognition systems in robotics — an area where active vision strategies inspired by insects are increasingly influential (Serres et al. 2018, MaBouDi et al. 2025 a , Ruzzante et al. 2024). These results highlight the potential for integrating insect-derived principles of sampling, generalisation, and face recognition learning into lightweight computational frameworks. Conclusion In conclusion, honey bees successfully learnt to discriminate complex human face stimuli, generalised this learning to novel exemplars, but failed when the learnt faces were inverted as a group. Together with the associated changes in flight trajectories, these results suggest that successful recognition depended not only on what bees learnt about the stimuli, but also on how they actively sampled them. Our findings therefore support the view that efficient active vision, rather than high neural complexity alone, can sustain robust recognition of visually complex patterns. Statements and Declarations Acknowledgements This research was supported by Australian Research Council (DP230100006), Australian Research Council (DP210100740), Templeton World Charity Foundation TWCF-2020-20539 and the European Union’s Horizon Europe research and innovation programme under the Marie Skłodowska-Curie grant agreement No. 101268923 (FACETS). Statement of Animal Ethics This study involved honey bees ( Apis mellifera ) and consisted exclusively of non-invasive behavioural experiments. No surgical, immobilising, or otherwise invasive procedures were performed. According to the applicable institutional and local regulations governing research on invertebrates at the time and place at which the study was conducted, formal approval by an animal ethics committee was not required for this work. All animals were handled in accordance with standard good practice for behavioural research on insects, and all efforts were made to minimise disturbance and stress during training and testing. Conflict of Interest Statement No competing interests declared. Data Availability Statement The datasets generated and analysed during the current study are available from the corresponding author upon reasonable request. The code and model resources used for bee tracking are available as described in the Methods section. Author contributions MGG and ABB conceived the study. MGG designed the protocol. MGG and MJ acquired the data. MGG curated the data. ADH performed video tracking. MGG analysed and presented the data. MGG drafted the manuscript. MGG and ABB revised the manuscript. All authors contributed to the finalisation of the manuscript. References Avarguès-Weber A., Portelli G., Bénard J., Dyer A. G., & Giurfa, M. (2010). Configural processing enables discrimination and categorisation of face-like stimuli in honeybees. Journal of Experimental Biology, 213, 593–601. https://doi.org/10.1242/jeb.039263 Avarguès-Weber A., Dyer A. G., & Giurfa M. (2011). Conceptualization of above and below relationships by an insect. Proceedings of the Royal Society B: Biological Sciences , 278(1707), 2299–2306. https://doi.org/10.1098/rspb.2010.1891 Avarguès-Weber A., Dyer A. G., Combe M., & Giurfa M. (2012). Simultaneous mastering of two abstract concepts by the miniature brain of bees. Proceedings of the National Academy of Sciences, 109(19), 7481–7486. https://doi.org/10.1073/pnas.1202576109 Avarguès-Weber A., & Giurfa M. (2013). Conceptual learning by miniature brains. Proceedings of the Royal Society B: Biological Sciences , 280(1772), 20131907. https://doi.org/10.1098/rspb.2013.1907 Baird E., Boeddeker N., Ibbotson M.R., & Srinivasan M. V. (2013). A universal strategy for visually guided landing. Proceedings of the National Academy of Sciences, 110, 18686–18691. https://doi.org/10.1073/pnas.1314311110 Barron A.B., Halina M., Klein C. (2023). Transitions in cognitive evolution. Proceedings of the Royal Society B: Biological Sciences, 290, 20230671. https://doi.org/10.1098/rspb.2023.0671 Braun E., Dittmar L., Boeddeker N., Egelhaaf M. (2012). Prototypical components of honeybee homing flight behaviour depend on the visual appearance of objects surrounding the goal. Frontiers in Behavioural Neuroscience, 6, 1. https://doi.org/10.3389/fnbeh.2012.00001 Carey S., & Diamond R. (1997). Are faces perceived as configurations more by adults than by children? Visual Cognition, 4, 253–274. https://doi.org/10.1080/13506289408402302 Chittka L., Thomson J.D., & Waser N.M. (1999). Flower constancy, insect psychology, and plant evolution. Naturwissenschaften , 86, 361–377. https://doi.org/10.1007/s001140050636 Chittka L., Dyer A. G., Bock F., & Dornhaus A. (2003). Bees trade off foraging speed for accuracy. Nature. https://doi.org/10.1038/424388a Chittka L. (2017). Bee cognition. Current Biology, 27(19), R1049–R1053. https://doi.org/10.1016/j.cub.2017.08.008 Dittmar L., Stürzl W., Baird E., Boeddeker N., & Egelhaaf M. (2010). Goal seeking in honeybees: Matching optic flow to distance. Journal of Experimental Biology, 213, 2916–2926. https://doi.org/10.1242/jeb.041657 Dyer A.G., Neumeyer C., Chittka L. (2005). Honeybee ( Apis mellifera ) vision can discriminate between and recognise images of human faces. Journal of Experimental Biology, 208(24), 4709–4714. https://doi.org/10.1242/jeb.01929 Dyer A.G., Rosa M.G.P., & Reser D.H. (2008). Honeybees can recognise images of complex natural scenes for use as potential landmarks. Journal of Experimental Biology, 211(8), 1180–1186. https://doi.org/10.1242/jeb.016683 Diamond R., & Carey S. (1986). Why faces are and are not special: an effect of expertise. Journal of Experimental Psychology: General, 115, 107–117. https://doi.org/10.1037/0096-3445.115.2.107 Giurfa M., Hammer M., Stach S., Stollhoff N., Müller-Deisig N., Mizyrycki C. (1999). Pattern learning by honeybees: conditioning procedure and recognition strategy. Animal Behaviour, 57(2), 315–324. https://doi.org/10.1006/anbe.1998.0957 Giurfa M. (2003). The amazing mini-brain: lessons from a honey bee. Bee World, 84(1), 5–18. https://doi.org/10.1080/0005772X.2003.11099566 Goulson D. (2000). Are insects flower constant because they use search images to find flowers? Oikos, 88(3), 547–552. https://doi.org/10.1034/j.1600-0706.2000.880311.x Guiraud M., Roper M., & Chittka L. (2018). High-speed videography reveals how honeybees can turn a spatial concept learning task into a simple discrimination task by stereotyped flight movements and sequential inspection of pattern elements. Frontiers in Psychology, 9, 1347. https://doi.org/10.3389/fpsyg.2018.01347 Guiraud M.-G., Roper M., Wolf S., Woodgate J.L., & Chittka L. (2022). Discrimination of edge orientation by bumblebees. PLOS ONE. https://doi.org/10.1371/journal.pone.0263198 Guiraud M.-G., Quinsal-Keel E., Gallo V. & MaBouDi H. (2024). Bumblebee visual learning: simple solutions for complex stimuli. Animal Behaviour. https://doi.org/10.1016/j.anbehav.2024.123070 Guiraud M-G., MaBouDi H., Woodgate J.L., Bates O.K., Ramos Rodriguez O., Gallo V., Barron A.B.B. (2025). How bumblebees manage conflicting information seen on arrival and departure from flowers. Animal Cognition, 28(1), 11. https://doi.org/10.1007/s10071-024-01926-x Kanwisher N., McDermott J., & Chun M.M. (1997). The fusiform face area: A module in human extrastriate cortex specialized for face perception. Journal of Neuroscience, 17(11), 4302–4311. https://doi.org/10.1523/JNEUROSCI.17-11-04302.1997 Henderson J.M., Williams C.C. & Falk R.J. (2005). Eye movements are functional during face learning. Memory & Cognition, 33(1), 98–106. https://doi.org/10.3758/BF03195300 Kanwisher N. (2000). Domain specificity in face perception. Nature Neuroscience, 3(8), 759–763. https://doi.org/10.1038/77664 Lehrer M. (1994). Spatial vision in the honeybee: the use of different cues in different tasks. Vision Research, 34(18), 2363–2385. https://doi.org/10.1016/0042-6989(94)90282-8 Lehrer M., Horridge G.A., Zhang S.W., Gadagkar R. (1995). Shape vision in bees: innate preference for flower-like patterns. Philosophical Transactions of the Royal Society of London B: Biological Sciences, 347(1320), 123–137. https://doi.org/10.1098/rstb.1995.0017 MaBouDi H., Marshall J.A.R., Dearden N., Barron A.B.B. (2023). How honey bees make fast and accurate decisions. eLife, 12, e86176. https://doi.org/10.7554/eLife.86176 a) MaBouDi H., Richter J., Guiraud M.-G., Roper M., Marshall J.A.R. & Chittka, L. (2025). Active vision of bees in a simple pattern discrimination task. eLife. https://doi.org/10.7554/eLife.106332 b) MaBouDi H., Roper M., Guiraud M.-G., Juusola M., Chittka L. & Marshall J.A.R. (2025). A neuromorphic model of active vision shows how spatiotemporal encoding in lobula neurons can aid pattern recognition in bees. eLife. https://doi.org/10.7554/eLife.89929 Maurer D., Le Grand R. & Mondloch C.J. (2002). The many faces of configural processing. Trends in Cognitive Sciences, 6(6), 255–260. https://doi.org/10.1016/S1364-6613(02)01903-4 Menzel R. & Giurfa M. (2001). Cognitive architecture of a mini-brain: The honeybee. Trends in Cognitive Sciences, 5(2), 62–71. https://doi.org/10.1016/S1364-6613(00)01601-6 Pascalis O., Kelly D.J. & Caldara R. (2006). What can bees really tell us about the face processing system in humans? Journal of Experimental Biology, 209(16), 3266. https://doi.org/10.1242/jeb.02411 Peterson M.F., Eckstein M.P. (2013). Individual differences in eye movements during face identification reflect observer-specific optimal points of fixation. Psychological Science, 24(7), 1216–1225. https://doi.org/10.1177/0956797612471684 Perry C.J., Barron A.B.B., Chittka L. (2017). The frontiers of insect cognition. Current Opinion in Behavioural Sciences, 16, 111–118. https://doi.org/10.1016/j.cobeha.2017.05.011 Ravi N., Gabeur V., Hu Y.-T., Hu R., Ryali C., Ma T., Khedr H., Rädle R., Rolland C., Gustafson L., Mintun E., Pan J., Alwala K.V., Carion N., Wu C.-Y., Girshick R., Dollár P. & Feichtenhofer C. (2024). SAM 2: Segment Anything in Images and Videos. arXiv. https://doi.org/10.48550/arXiv.2408.00714 Richler J.J. & Gauthier I. (2014). A meta-analysis and review of holistic face processing. Psychological Bulletin, 140(5), 1281–1302. https://doi.org/10.1037/a0037004 Ruzzante D. & Vaes J. (2024). Configural face processing and its influence on the timeline of mentalization. Current Research in Ecological and Social Psychology, 6, 100184. https://doi.org/10.1016/j.cresp.2024.100184 Serres J.R., Viollet S. (2018). Insect-inspired vision for autonomous vehicles. Current Opinion in Insect Science, 30, 46–51. https://doi.org/10.1016/j.cois.2018.09.005 Srinivasan M.V., Lehrer M. (1988). Spatial acuity of honeybee vision and its spectral properties. Journal of Comparative Physiology A, 162, 159–172. https://doi.org/10.1007/BF00606081 Srinivasan M.V. (2010). Honey bees as a model for vision, perception, and cognition. Annual Review of Entomology , 55, 267–284. https://doi.org/10.1146/annurev.ento.010908.164537 Tsao D.Y. & Livingstone M.S. (2008). Mechanisms of face perception. Annual Review of Neuroscience, 31, 411–437. https://doi.org/10.1146/annurev.neuro.30.051606.094238 Yin R.K. (1969). Looking at upside-down faces. Journal of Experimental Psychology , 81(1), 141–145. https://doi.org/10.1037/h0027474 Yang C-Y., Huang H-W., Chai W., Jiang Z., Hwang J-N. (2024). SAMURAI: Adapting Segment Anything Model for zero-shot visual tracking with motion-aware memory. arXiv preprint, arXiv:2411.11922. https://doi.org/10.48550/arXiv.2411.11922 Zhang S., Srinivasan M.V., Zhu H., Wong J. (2004). Grouping of visual objects by honeybees. Journal of Experimental Biology, 207(19), 3289–3298. https://doi.org/10.1242/jeb.01155 Additional Declarations No competing interests reported. Supplementary Files FacearticleSuppMaterial21.04.2026.docx Cite Share Download PDF Status: Under Review Version 1 posted Reviewers agreed at journal 06 May, 2026 Reviewers invited by journal 23 Apr, 2026 Editor assigned by journal 23 Apr, 2026 Submission checks completed at journal 22 Apr, 2026 First submitted to journal 21 Apr, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-9487302","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":633696935,"identity":"62594976-627c-4a7f-8346-7be4ae3f6da1","order_by":0,"name":"Marie-Genevieve Guiraud","email":"data:image/png;base64,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","orcid":"","institution":"Macquarie University","correspondingAuthor":true,"prefix":"","firstName":"Marie-Genevieve","middleName":"","lastName":"Guiraud","suffix":""},{"id":633696938,"identity":"33cad9ee-63e3-4b1e-9dce-55dd46105e77","order_by":1,"name":"Adam D. Hines","email":"","orcid":"","institution":"Macquarie University","correspondingAuthor":false,"prefix":"","firstName":"Adam","middleName":"D.","lastName":"Hines","suffix":""},{"id":633696940,"identity":"c4b05bb5-d951-46e2-9d9a-3e3592d658b0","order_by":2,"name":"Mengjie Jin","email":"","orcid":"","institution":"Macquarie University","correspondingAuthor":false,"prefix":"","firstName":"Mengjie","middleName":"","lastName":"Jin","suffix":""},{"id":633696941,"identity":"153fee1c-c6db-4fa5-bf68-f567e11bb53d","order_by":3,"name":"Andrew B. Barron","email":"","orcid":"","institution":"Macquarie University","correspondingAuthor":false,"prefix":"","firstName":"Andrew","middleName":"B.","lastName":"Barron","suffix":""}],"badges":[],"createdAt":"2026-04-21 17:08:15","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9487302/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9487302/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":108427494,"identity":"82219ef6-4930-4c81-8aae-b0488f5ad96b","added_by":"auto","created_at":"2026-05-04 13:55:39","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":358388,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSchematic representation of the experimental setup used to present face stimuli to individually flying bees.\u003c/strong\u003e The testing arena consisted of a white open-top box lined with patterned paper to create a homogeneous visual background. High-resolution greyscale human face images were affixed to the internal walls for controlled stimulus presentation. Several smartphones mounted on articulated arms recorded the bees’ flight behaviour. Below each face, a feeder was available to deliver either sucrose solution or quinine to the bee, depending on which type of face was rewarded.\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-9487302/v1/ba85aacdb8ef6a5817b225d6.png"},{"id":108427431,"identity":"a79a3cf0-c341-4124-a27d-b95a2e1954f6","added_by":"auto","created_at":"2026-05-04 13:55:30","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":700362,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePre-training, training and test conditions displayed on the experimental wall.\u003c/strong\u003e \u003cem\u003e\u003cstrong\u003e(a) Pre-training.\u003c/strong\u003e\u003c/em\u003e Stimuli displayed a neutral grey oval. \u003cem\u003e\u003cstrong\u003e(b) Training\u003c/strong\u003e\u003c/em\u003e\u003cem\u003e \u003c/em\u003e\u003cem\u003e\u003cstrong\u003etrials.\u003c/strong\u003e\u003c/em\u003e\u003cem\u003e \u003c/em\u003eStimuli consisted of six greyscale human faces (three exemplars each of two different faces) displayed simultaneously. Depending on the bee’s group, one face was associated with sucrose solution and the other with quinine solution (bitter taste). The spatial position of all faces was pseudorandomised on every trial to prevent the use of odour/gustatory/location cues. \u003cem\u003e\u003cstrong\u003e(c) Learning test (LT).\u003c/strong\u003e\u003c/em\u003e\u003cem\u003e \u003c/em\u003eThe six learned faces were presented in a new randomised configuration, but all feeders presented distilled water. \u003cem\u003e\u003cstrong\u003e(d) Inverted test (IT).\u003c/strong\u003e\u003c/em\u003e The same faces used in training were presented upside-down. Feeders remained unrewarded. \u003cem\u003e\u003cstrong\u003e(e) Generalisation test (GT).\u003c/strong\u003e\u003c/em\u003e\u003cem\u003e \u003c/em\u003eTo evaluate generalisation, bees were shown two new faces – one from each of the two categories used in the CFD – which had not been used in training.\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-9487302/v1/b87c7579a08fa2dae2a82421.png"},{"id":108427492,"identity":"4e5d66bb-6d59-4458-9dbe-f0712a4f2166","added_by":"auto","created_at":"2026-05-04 13:55:39","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":92762,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003e(a)\u003c/strong\u003e \u003cstrong\u003eLearning performance during the first 50 choices vs. the last 50 choices (N=33)\u003c/strong\u003e. GLMM (p-value \u0026lt;0.001). The red dashed line indicates chance level (50%). Error bars represent the standard error of the mean (SEM). \u003cstrong\u003e(b) Individual performance of bees in the three non-rewarded tests: \u003c/strong\u003ethe Learning Test (LT), the Inverted Test (IT), and the Generalisation Test (GT). Bees performed significantly above chance in the LT (Wilcoxon signed-rank test, p \u0026lt; 0.001), did not differ from chance in the Inverted Test (Wilcoxon signed-rank test, n.s.), and performed above chance in the Generalisation Test (Wilcoxon signed-rank test, *p \u0026lt; 0.05). Bars show the mean ± SEM, and the number inside each bar indicates the sample size (N). Each point represents one bee and boxplots show the median, interquartile range and range. The red dashed line indicates chance level (50 %). Asterisks denote significant differences from chance (Wilcoxon signed-rank tests: *** P \u0026lt; 0.001, * P \u0026lt; 0.05; n.s. = non-significant).\u003c/p\u003e","description":"","filename":"3.png","url":"https://assets-eu.researchsquare.com/files/rs-9487302/v1/1a186b852b0cdb016ac9f004.png"},{"id":108427433,"identity":"f4c222a7-67f5-438f-a82a-ba7251c11c50","added_by":"auto","created_at":"2026-05-04 13:55:31","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":79688,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eScanning dynamics across naïve, learning, inversion, and generalisation tests.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003e(a) Mean translational flight speed (px/s) within the region of interest for the Learning Test (LT), Inverted Test (IT), Generalisation Test (GT), and naïve (N) condition.\u003c/strong\u003e Bees flew significantly faster after learning and during IT and GT compared to naïve flights (see Kruskal–Wallis with Dunn’s post hoc tests in text, LT P\u0026lt;0.01**; IT P\u0026lt;0.001*** and GT P\u0026lt;0.05*). Error bars show SEM; Black points represent individual bees. \u003cstrong\u003e(b) Mean angular velocity (radians/s) across conditions.\u003c/strong\u003e Naïve bees displayed the highest and most variable angular velocities. After learning, angular velocity decreased substantially, resulting in straighter, more directed flights (Kruskal–Wallis with Dunn’s post hoc test LT P=0.016, see text). GT and IT maintained angular velocity levels comparable to LT, and with a tendency to be lower than naïve bees.\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-9487302/v1/7a9c835fb5f4e0743f08d07e.png"},{"id":108804364,"identity":"06de6dd3-2890-4083-adfd-4a5855ddce23","added_by":"auto","created_at":"2026-05-08 15:19:46","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1510670,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9487302/v1/3ea01f5f-c798-4f99-a6b8-d75549455559.pdf"},{"id":108427517,"identity":"4b2f32f5-6b50-40ea-8f7d-9f1ad15ac4b1","added_by":"auto","created_at":"2026-05-04 13:55:42","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":996196,"visible":true,"origin":"","legend":"","description":"","filename":"FacearticleSuppMaterial21.04.2026.docx","url":"https://assets-eu.researchsquare.com/files/rs-9487302/v1/3ea18a6f095f9938ecc739e8.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Learning complex stimuli with a honey bee brain: contrasting consequences on performance and flight behaviour from degrading configural versus featural information","fulltext":[{"header":"Introduction","content":"\u003cp\u003eThe honey bee has become an important model for visual recognition and visual cognition (Giurfa \u003cem\u003eet al.\u003c/em\u003e 2003, Srinivasan \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e2010\u003c/span\u003e). Compared to humans and most mammals, honey bees have limited eye resolution and minute brains, yet bees can discriminate shapes, colours, spatial configurations and even human faces under some conditions (Dyer et al. \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2005\u003c/span\u003e, Avargu\u0026egrave;s-Weber et al. \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2010\u003c/span\u003e). This paradox\u0026mdash;high-level visual performance with minimal neural resources\u0026mdash;has fuelled interest in identifying the mechanisms by which bees acquire, encode and use visual information. The face case is especially interesting because face discrimination is a subtle multi-element comparison that, for humans, involves both featural (individual features) and configural information (the spatial relationships among those features; Maurer et al. \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2002\u003c/span\u003e, Richler \u003cem\u003eet al.\u003c/em\u003e 2014) and it is important to note that for bees, learning faces is outside of their cognitive ecology. In this study, we explored how bees learn these complex visual stimuli by testing their sensitivity to manipulations that degraded either the configural or the featural information in faces.\u003c/p\u003e \u003cp\u003eBees show striking abilities in visual learning and categorisation (Giurfa 2013, Chittka \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). Bees can discriminate complex natural scenes (Zhang et al. \u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e2004\u003c/span\u003e), learn abstract relationships such as same/different, and encode global relational properties of visual patterns even when composed of simple geometric elements (Avargu\u0026egrave;s-Weber et al. \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2010\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2011\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2012\u003c/span\u003e). Although insect vision lacks both a retina capable of high spatial resolution and a cortex for higher-order visual object recognition (Kanwisher et al. \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e1997\u003c/span\u003e, Kanwisher \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2000\u003c/span\u003e), bees nonetheless achieve robust visual recognition using compact neural circuits and highly efficient active-vision strategies\u0026mdash;namely, the integration of movement, selective sampling and iterative evidence accumulation (Srinivasan \u0026amp; Lehrer \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e1988\u003c/span\u003e, Lehrer \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e1994\u003c/span\u003e, Baird et al. \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2013\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eDyer et al. (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2005\u003c/span\u003e) first showed that honey bees can discriminate individual human faces when trained on whole-face images. This study raised interest in how bees might be solving complex discriminations such as this. Face perception has been considered a hallmark of primate cognition (Kanwisher et al. \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e1997\u003c/span\u003e, Tsao \u0026amp; Livingstone \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2008\u003c/span\u003e). Humans and other primates possess highly specialized neural systems for detecting, encoding and individuating faces, relying on both featural information (eyes, nose, mouth) and the spatial relations among features\u0026mdash;so-called configural or holistic processing (Kanwisher \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2000\u003c/span\u003e, Maurer et al. \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2002\u003c/span\u003e). The classic face-inversion effect (Yin \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e1969\u003c/span\u003e), in which inversion of faces slows recognition more than inversion of most objects, exemplifies the sensitivity of primates to the orientation-dependent spatial relations of facial features. Extensive developmental, neurophysiological and psychophysical research demonstrates that experience strongly shapes the strategies humans use to extract identity-relevant cues (Carey \u0026amp; Diamond \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e1997\u003c/span\u003e, Peterson \u0026amp; Eckstein \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2013\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eIn Dyer\u0026rsquo;s original study (Dyer et al. \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2005\u003c/span\u003e), five bees were first pre-trained to associate a rewarded human face with sucrose against a schematic distractor, before being trained to discriminate a target face from other similar face images. Following differential conditioning, bees successfully discriminated the rewarded face from novel distractor faces, indicating generalisation across exemplars. However, performance dropped markedly when both target and distractor faces were presented inverted, suggesting that recognition relied on orientation-dependent visual cues. This finding can be interpreted as being consistent with sensitivity to configural information in faces, although inversion effects alone cannot distinguish between configural and feature-based strategies. Accordingly, this interpretation has been questioned (Pascalis et al. \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2006\u003c/span\u003e). Configural processing encompasses different levels of relational encoding, from basic first-order arrangements of features to complex metric and holistic representations, which are typically considered more demanding than simple feature-based learning. In this context, and in the absence of dedicated face-specific neural circuits comparable to those described in primates, it remains unclear to what extent bees rely on genuinely configural representations as opposed to alternative, simpler strategies. A follow-up study by Avargu\u0026egrave;s-Weber \u003cem\u003eet al.\u003c/em\u003e (2013) showed that honeybees trained on minimalistic, schematic \u0026ldquo;face-like\u0026rdquo; stimuli can learn first-order relational (configural) properties between stimulus elements. Bees trained to recognise a canonical arrangement of features were able to generalise this learning to novel stimuli, discriminating face-like configurations from scrambled arrangements composed of the same elements (Avargu\u0026egrave;s-Weber \u003cem\u003eet al.\u003c/em\u003e 2013). This provides strong evidence that bees can encode spatial relationships between features rather than relying solely on local cues. However, these results are limited to first-order relational structure and much remains to be understood about the mechanisms that allow bees to rapidly learn and exploit such structured visual information.\u003c/p\u003e \u003cp\u003eBees\u0026rsquo; movements can yield information on how bees learn. Bees do not passively receive visual input and instead continuously sample a scene through structured flight manoeuvres that shape what information reaches the eye (Srinivasan \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e2010\u003c/span\u003e, Guiraud et al. \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2018\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2024\u003c/span\u003e, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2025\u003c/span\u003e, MaBouDi et al. \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). When approaching complex stimuli, bees exhibit characteristic saccade-like turns, scanning loops, and speed modulations that guide visual processing to facilitate evidence accumulation during decision-making (Braun et al. \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2012\u003c/span\u003e, Lehrer \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e1994\u003c/span\u003e). Bees fly close to specific regions, follow contours, oscillate laterally, or sweep across edges in ways that maximise contrast detection and minimise processing costs (Lehrer et al. \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e1995\u003c/span\u003e; Dittmar et al. \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2010\u003c/span\u003e). These movements generate temporal sequences of visual features that the mushroom bodies can associate with reinforcement (Menzel \u0026amp; Giurfa, \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2001\u003c/span\u003e). Detailed analyses of bee flight during pattern discrimination tasks showed that trained bees shifted from slow, rotational sampling during unfamiliar stimuli to faster and more directed trajectories once they recognise a target (MaBouDi \u003cem\u003eet al.\u003c/em\u003e 2025\u003csup\u003eb\u003c/sup\u003e). Therefore, investigation of bees\u0026rsquo; flight behaviours can reveal whether or not a stimulus is familiar.\u003c/p\u003e \u003cp\u003eHere, we asked whether bees trained on a single pair of human faces selected from different categories within the Chicago Face Database (CFD) would (i) generalise their learning to a test in which both alternatives were novel faces, (ii) be sensitive to inversion (using the same two stimuli used in training but inverted at 180\u0026deg;), which disrupts the trained spatial arrangement of features, and (iii) show corresponding changes in flight behaviour indicative of a learned active-sampling strategy. By combining performance measures with flight kinematics, we aimed to determine whether bees could solve these discriminations, and whether successful recognition in conditions where the featural (i) and configural information (ii) was degraded is accompanied by distinct signatures of learned visual sampling.\u003c/p\u003e"},{"header":"Material and methods","content":"\u003cp\u003eExperiments were conducted over three consecutive summers (2021\u0026ndash;2024). A single hive was placed within a 160 m\u003csup\u003e2\u003c/sup\u003e enclosed shade house for 1-2 months, depending on the hive\u0026rsquo;s health. Foragers from this hive accessed a communal gravity feeder containing 30% (w/w) sucrose solution positioned 1.5 m from the hive.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eIndividual bees were trained from the gravity feeder to fly into a testing arena measuring 60 cm (L) \u0026times; 60 cm (W) \u0026times; 40 cm (H) (Fig. 1). Six vertically mounted Eppendorf tubes (experimental feeders) were affixed to the rear wall beneath six visual stimuli. Each feeder dispensed 15 \u0026mu;l of 50% (w/w) sucrose solution\u0026mdash;below the bee\u0026rsquo;s crop capacity\u0026mdash;to encourage multiple feeder visits within a trial in the testing arena. Stimuli were real human faces (self-reported sex and ethnicity) in shades of grey against white background from the Chicago Face Database (CFD) (https://www.chicagofaces.org). For the training and tests, we used faces from two different categories in the CFD: self-reported women faces from self-rated Asian ethnicity and self-reported men faces from self-rated Latino ethnicity.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eIndividual pre-training protocol\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo recruit single foragers to the testing arena, a returning bee at the gravity feeder was offered a 50% (w/w) sucrose-soaked cotton swab. Once the bee began feeding, the experimenter gently transferred it to one of six Eppendorf-tube feeding stations in the testing arena, simultaneously marking the bee\u0026rsquo;s dorsal abdomen with a unique coloured dot (Posca pens, Uni-Ball, Japan).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eDuring pre-training, the stimuli wall displayed six uniformly grey, oval-shaped discs (similar in size to the faces). Only one bee was trained at a time. This procedure was repeated until the bee learnt to fly directly to the feeders in the testing arena. Typically, a bee needed two or three transfers to the testing arena before she returned of her own volition. The experimenter ensured that the selected individual visited each of the six feeders in the testing arena at least once. This phase familiarised the bee with the spatial layout of the feeders. All feeders contained an identical sucrose reward at this stage. Unmarked bees were removed from the testing arena.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTraining\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eOnce bees consistently visited all six feeder locations during pre-training, we initiated differential conditioning with two different faces drawn from the CFD. In each trial, six faces were displayed \u0026ndash; two distinct faces, one from each of our selected categories from the CFD, three copies of each - on the rear wall in randomised positions. Bees in group A received 15 \u0026mu;L of 50% (w/w) sucrose solution when landing on feeders beneath Asian women (CS+) faces and 15 \u0026mu;L of saturated quinine solution (0.12% w/w) when landing on feeders beneath Latino men (CS\u0026ndash;) faces. Training group B was trained with the reciprocal contingencies. Randomising stimuli positions each trial ensured that bees relied on visual features rather than location. To eliminate residual odours and chemical cues between trials, the arena and stimuli were thoroughly wiped with 70% ethanol. Bees are known to leave pheromone-laden footprints or scent marks during foraging and exploration, which can influence the behaviour of subsequent individuals if not removed (Chittka \u003cem\u003eet al.\u003c/em\u003e 1999, Goulson \u003cem\u003eet al.\u003c/em\u003e 2000). Ethanol is widely used in bee cognition studies as an effective cleaning agent to remove such olfactory traces. To ensure that ethanol itself did not interfere with the bees\u0026rsquo; behaviour, a minimum of 2\u0026ndash;3 minutes was observed after cleaning, allowing sufficient time for complete evaporation before the next bee was introduced (Giurfa \u003cem\u003eet al.\u003c/em\u003e 1999). Each trial encompassed the full duration of a bee\u0026rsquo;s visit to the testing arena, beginning with its entry into the arena and concluding with its departure. During this period, a bee could land on multiple feeders, with each landing recorded as a choice. Individual bees typically made three to five landings \u003cem\u003eper\u003c/em\u003e trial. Training continued until each bee achieved \u0026ge; 80% correct choices (landing on the rewarded faces) across its most recent 20 landings. Bees typically required 25\u0026ndash;35 trials to reach criterion, which would take 1-2 days of training.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTesting\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAfter reaching the training criterion, we conducted three unrewarded test types: learning test (LT), inverted test (IT), and generalisation test (GT; Fig. 2). In all tests, all six feeders offered distilled water. In the learning test, bees were presented with the original six stimuli. This test assessed whether bees had genuinely learned to use the previously rewarded face stimulus to perform the discrimination. In the inverted test, the same six faces were shown inverted (180\u0026deg; rotation) to assess sensitivity to configural changes. A drop in performance compared to the learning test would indicate reliance on canonical face orientation in training. In the generalisation test, a new self-assigned gender and ethnicity Asian woman and Latino man face was presented matched for greyscale and relative size to evaluate whether bees could extend the learned discrimination to new exemplars. Each test lasted 2 minutes, during which we recorded the number of landings on the previously rewarded (correct) \u003cem\u003eversus\u0026nbsp;\u003c/em\u003epreviously punished faces\u003cem\u003e\u0026nbsp;\u003c/em\u003e(incorrect). To sustain motivation, unrewarded tests were interleaved with refreshment trials using the original training contingencies (sucrose on CS+; quinine on CS\u0026ndash;). Bees were required to achieve \u0026ge; 80 % correct in a refreshment trial before attempting the next unrewarded test, typically performing one to three refresher trials between tests. By comparing performance across these tests, we quantified both retention of the learned discrimination and the flexibility of visual generalisation under altered presentation conditions.\u003c/p\u003e\n\u003cp\u003eA total of 33 bees completed at least the training phase. Sample size varied across analyses because not all individuals completed all training blocks and subsequent tests. Sixteen bees completed the training and the first learning test; fourteen completed the training, the first learning test, and the inverted test; and twelve completed the training and all three tests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eVideo analysis: 2D bee tracking\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe recorded flight trajectories and scanning behaviour using an iPhone X camera (1,280 \u0026times; 720 px at 120 fps) opposite the stimulus wall. Recording began as soon as the marked bee entered the arena and continued until departure, generating one video file\u003cem\u003e\u0026nbsp;per\u0026nbsp;\u003c/em\u003etrial. For the bee tracking experiments, we utilised the Segment Anything 2 (SAM2) model and SAMURAI for zero-shot visual tracking and motion-aware memory to track bees over several frames (Yang et al. 2024, Ravi et al. 2024). The SAM2 model used was the SAM2.1 Hiera Large for the highest accuracy segmentation and was fine-tuned on 100 examples of bees in the experimentation arena. Ground truth segmentation masks were annotated manually using ImageJ. The fine-tuned code, model including the images, masks, and bounding boxes used for fine-tuning are available at https://huggingface.co/AdamHines/samurai-beetracker.\u003c/p\u003e\n\u003cp\u003eIndividual frames for segmentation were derived from the experimental recordings at a rate of 24 frames \u003cem\u003eper\u003c/em\u003e second and divided into 1,000 frame chunks for processing to circumvent computational memory limitations (Ravi et al. 2024). Initial frames \u003cem\u003eper\u003c/em\u003e chunk were selected and a bounding box was manually selected around the bee for the starting segmentation. Once selected, the SAMURAI system performed motion-aware tracking across the rest of the frames (Yang et al. 2024). In order to carry segmentation information from one chunk to the next, the bounding box from the last frame of the preceding chunk was used to initialise the segmentation of the first frame of the next chunk. Once all chunks were processed, the entire trajectory was stitched together for analysis.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStatistical analysis\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll statistical analyses were conducted in R (v4.x). Learning performance was analysed using a Generalised Linear Mixed Model (GLMM; package \u003cem\u003elme4\u003c/em\u003e) with a binomial error structure, where the response variable was the number of correct \u003cem\u003eversus\u003c/em\u003e incorrect choices per block of 10 trials (successes/failures). Fixed effects included training block and rewarded stimulus category (stimulus \u0026times; protocol), and bee identity (Bee_ID) was included as a random intercept to account for repeated measures and inter-individual variability. Model assumptions and rank deficiency were evaluated through inspection of residuals and by verifying collinearity among predictors. \u003cem\u003ePost-hoc\u003c/em\u003e comparison of learning rates between women-rewarded and men-rewarded groups was assessed by comparing a reduced (Block + Group) and a full (Block \u0026times; Group) GLMM using likelihood-ratio tests.\u003c/p\u003e\n\u003cp\u003eFor unrewarded tests, bees\u0026rsquo; performance (proportion of correct choices) was compared against chance level (0.5) using Wilcoxon signed-rank tests, as normality assumptions were not met. Differences between treatment groups (female-rewarded vs male-rewarded) were evaluated using Wilcoxon rank-sum tests or Kruskal\u0026ndash;Wallis tests when more than two groups were compared. For each test type (learning, inverted, generalisation), effect sizes, test statistics, and exact \u003cem\u003ep\u003c/em\u003e-values were reported. Relationships between training length (total number of training trials) and test performance were examined using Pearson correlations with pairwise deletion of missing values.\u003c/p\u003e\n\u003cp\u003eFlight-tracking data obtained from the SAM2/SAMURAI segmentation pipeline were analysed separately. We reconstructed flights within a narrow region of interest around each of the face stimuli. This was defined as a rectangle centred on the face and extending beyond it by approx. 40 pixels in each direction. For each flight segment within the region of interest of the face stimulus, the mean translational speed (px/s) and angular velocity (rad/s) were extracted. Differences in speed and angular velocity between experimental conditions (na\u0026iuml;ve, learning test, inversion test, generalisation test) were first analysed using a one-way ANOVA when assumptions of normality and variance homogeneity were met. Otherwise, a Kruskal\u0026ndash;Wallis test was used. Significant omnibus tests were followed by Tukey\u0026rsquo;s HSD (for ANOVA) or Dunn\u0026rsquo;s post-hoc comparisons with Holm correction (for non-parametric tests). Data are reported as means \u0026plusmn; SEM unless stated otherwise. All tests were two-tailed and significance thresholds were set at \u0026alpha; = 0.05.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003e\u003cstrong\u003eBees\u0026rsquo; performance in face recognition task\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll 33 bees included in the training analysis learnt the task to criterion (Fig. 3a). Training of bees stopped when an individual made 80% (or more) correct choices within the last 20 choices, therefore the number of training choices differed for each bee. Bees progressively increased their proportion of correct choices across training, as reflected by a significant positive effect of training block in the GLMM analysis (Table S1; Fig. 3a). The response variable was the number of correct choices within each block of 10 trials, modelled as a binomial response (successes \u003cem\u003eversus\u003c/em\u003e failures). The model revealed a strong effect of training block on number of correct choices per block (estimate = 0.0168 \u0026plusmn; 0.0037, \u003cem\u003ez\u003c/em\u003e = 4.55, \u003cem\u003ep\u003c/em\u003e = 5.34 \u0026times; 10⁻⁶), indicating performance improvement with training. Neither training group (A vs. B) nor its interaction with training block explained additional variance (Table S1). An additional Wilcoxon test showed the significant difference between the two last bouts (Wilcoxon tests: p \u0026lt; 0.01, p \u0026lt; 0.001).\u003c/p\u003e\n\u003cp\u003eBees performed significantly above chance during the unrewarded learning test (Fig. 2c), confirming that they remembered the rewarded stimulus in an unrewarded test (Wilcoxon signed-rank test: V = 325, P = 6.36 \u0026times; 10⁻⁶; Fig. 3b and Table S3). There was no significant difference in LT performance between bees trained in groups A or B (Wilcoxon rank-sum test: W = 60.5, P = 0.435).\u003c/p\u003e\n\u003cp\u003eIn the unrewarded inverted test IT (Fig. 2d) bees did not exhibit a preference for the previously rewarded stimulus (Wilcoxon signed-rank test: V = 101, P = 0.405; Fig. 3c). Performance did not differ between training groups (Kruskal\u0026ndash;Wallis test: N = 14, H = 0.161, P = 0.688; Table S4). These results indicate that bees were unable to recognise the trained face when inverted.\u003c/p\u003e\n\u003cp\u003eIn the unrewarded generalisation test (GT) (Fig. 2e), bees showed a significant preference for the face from the same CFD category they were trained on (Wilcoxon signed-rank test: V = 98, P = 0.015; Fig. 3b). Performance did not differ between training groups (Kruskal\u0026ndash;Wallis test: N = 12, H = 0.0017, P = 0.967; Table S5).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eBees\u0026rsquo; individual differences and correlations\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eBees exhibited marked individual variation (Fig. 3b) in their learning performance and in their ability to generalise facial information, despite the clear group-level patterns observed in the training and test phases. During acquisition, some individuals rapidly reached high accuracy, whereas others required substantially more trials to improve, leading to a broad range of total training lengths (from ~60 to \u0026gt;300 choices; Table S7). These differences in training were not predictive of performance in the Learning Test (LT: r = \u0026ndash;0.33, P = 0.113; Table S6) or in the Inverted Test (IT: r = 0.18, P = 0.415; Table S6), where performance remained tightly clustered around the group means. However, training length was positively correlated with performance in the Generalisation Test (GT: r = 0.56\u0026ndash;0.62, P = 0.012\u0026ndash;0.004; Table S6), indicating that individuals exposed to more training trials developed more robust or more flexible face representations. These behavioural divergences were also reflected in individual outcomes: some bees completed very few choices during trials (\u003cem\u003ee.g.,\u003c/em\u003e Bee_18, 60 choices), while others exceeded 300, and several individuals achieved perfect accuracy during the Learning Test (Bee_10, Bee_21, Bee_25). In contrast, a few bees excelled specifically in the Inverted Test (Bee_8, Bee_33), suggesting differential sensitivity to stimulus rotation. The most pronounced variability emerged in the Generalisation Test, where a subset of \u0026ldquo;high-level generalisers\u0026rdquo; (\u003cem\u003ee.g.,\u003c/em\u003e Bee_3: 100%; Bee_6: 80%; Bee_4 and Bee_24: 75%) displayed exceptionally strong performance, whereas others remained at or near chance (e.g., Bee_13, Bee_14, Bee_16, Bee_33, Bee_9).\u003c/p\u003e\n\u003cp\u003e\u0026nbsp;\u003cstrong\u003eFlight and scanning behaviour\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eTo determine how bees sampled the face stimuli during decision making, we reconstructed their flight trajectories and quantified their movement dynamics across all test conditions (Fig. 4a and b). Flight paths obtained using a semantic-segmentation-based tracking pipeline revealed systematic differences between na\u0026iuml;ve bees and bees that had undergone learning (Fig. 4a-b; see Methods).\u003c/p\u003e\n\u003cp\u003eNa\u0026iuml;ve bees approached the face stimuli slowly with highly rotational flight, exhibiting the lowest speeds (mean = 69.37 px/s, n = 12) and the highest angular velocities (mean = \u0026ndash;2.55 rad/s, n = 12; Fig. 4a\u0026ndash;b) as they scanned the novel stimuli. There is a global effect of the conditions on flight speed (Kruskal-Wallis global test H (3)=13.55, p\u0026lt;0.003) but not for angular velocity (Kruskal-Wallis global test H (3)=6.056, p=0.1094).\u003c/p\u003e\n\u003cp\u003eAfter learning (during the learning test LT), bees showed a marked shift in scanning behaviour. Flight speed increased significantly (mean = 98.6 px/s, n = 12; Kruskal\u0026ndash;Wallis with Dunn\u0026rsquo;s post hoc test LT \u003cem\u003evs\u003c/em\u003e\u003cem\u003e.\u003c/em\u003e N: z= 2.668, p = 0.008), and angular velocity was strongly reduced (mean = \u0026ndash;0.49 rad/s, n = 12; Kruskal\u0026ndash;Wallis with Dunn\u0026rsquo;s post hoc test LT \u003cem\u003evs\u003c/em\u003e\u003cem\u003e.\u0026nbsp;\u003c/em\u003eN: z= 2.403, p = 0.016), indicating straighter, and faster trajectories. In the Inverted Test IT, flight speed increased further (mean = 127.07 px/s), significantly higher than in na\u0026iuml;ve bees (Kruskal\u0026ndash;Wallis with Dunn\u0026rsquo;s post hoc test IT \u003cem\u003evs\u003c/em\u003e\u003cem\u003e.\u003c/em\u003e N: z= 3.479, p = 0.0005), although not significantly higher than the learning test LT (Kruskal\u0026ndash;Wallis with Dunn\u0026rsquo;s post hoc test LT \u003cem\u003evs.\u003c/em\u003e IT: z= 0.794, p = 0.427; Fig. 4b). Angular velocity remained similar to the learning condition (Kruskal\u0026ndash;Wallis with Dunn\u0026rsquo;s post hoc test LT \u003cem\u003evs.\u003c/em\u003e IT, z= -0.802, p = 0.422), but was still lower\u0026mdash;although not significantly\u0026mdash;than in na\u0026iuml;ve bees (Kruskal\u0026ndash;Wallis with Dunn\u0026rsquo;s post hoc test IT \u003cem\u003evs.\u003c/em\u003e N, z= 1.584, p = 0.11), In the Generalisation Test GT, bees flew at speeds comparable to the learning test LT (mean = 97.09 px/s) and significantly faster than na\u0026iuml;ve bees (Kruskal\u0026ndash;Wallis with Dunn\u0026rsquo;s post hoc test GT \u003cem\u003evs.\u003c/em\u003e N: z= 2.342, p = 0.019). Angular velocity in this condition was low and did not differ from the learning test LT or the inverted test IT (Kruskal\u0026ndash;Wallis with Dunn\u0026rsquo;s post hoc test GT vs. LT: z= -0.982, p = 0.325 and GT vs. IT: z= -0.163, p = 0.871; Fig. 4b).\u0026nbsp;\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eOur findings build on a growing body of work showing that bees can perform sophisticated visual discriminations despite small brains and coarse visual resolution (Guiraud et al. \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2018\u003c/span\u003e, \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2022\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2024\u003c/span\u003e, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2025\u003c/span\u003e, MaBouDi et al., \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2023\u003c/span\u003e, 2025\u003csup\u003ea,b\u003c/sup\u003e). Consistent with previous studies showing that bees can recognise complex visual patterns, including faces and face-like stimuli, as holistic configurations (Dyer et al. \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2005\u003c/span\u003e; Avargu\u0026egrave;s-Weber et al. \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2010\u003c/span\u003e), bees rapidly learned to discriminate between two different human faces. As a population, bees performed significantly above chance in the unrewarded learning test (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eb). All tested bees chose the face rewarded in training more than 50% in the test, and some bees chose the rewarded face 100% in test (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eb and Table S7). We also noted clear changes in bees\u0026rsquo; flight behaviour as they learned the visual discrimination task. Na\u0026iuml;ve bees flew slowly and with highly rotational trajectories, consistent with exploratory sampling during unfamiliar visual tasks (Guiraud et al. \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2025\u003c/span\u003e; MaBouDi \u003cem\u003eet al.\u003c/em\u003e 2025\u003csup\u003ea,b\u003c/sup\u003e). After training to an 80% performance criterion, bees exhibited faster, more linear flight (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ea-b, Supp. Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e), possibly reflecting a transition to more directed sampling and stable evidence accumulation. This change mirrors findings by MaBouDi et al. (\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) showing that bees gather diagnostic information through dynamic, efficient scanning routines when familiar with a visual task. The changes in flight behaviour we saw were unlikely to be adaptations to the flight arena because we considered flights within a small region around each face only and not globally within the arena.\u003c/p\u003e \u003cp\u003eOnce trained with just a single pair of faces, bees were able to generalise their learned face information to successfully discriminate between two novel faces (GT, Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eb). Although novel faces differed from those experienced during training, bees significantly preferred the face from the same CFD category as the one rewarded in training (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eb). In our generalisation test, both faces were new to the bee. This differs from Dyer et al. (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2005\u003c/span\u003e) who tested whether bees could distinguish a learned rewarded face from a novel distractor. We believe this is the first test of whether bees can distinguish between two novel faces based on similarity to trained exemplars. Generalisation is a robust feature of bee visual learning and has been documented for naturalistic scenes and visual objects, as well as for abstract and relational pattern learning (Zhang et al. \u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e2004\u003c/span\u003e, Dyer et al. \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2008\u003c/span\u003e, Avargu\u0026egrave;s-Weber et al. \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2011\u003c/span\u003e, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2012\u003c/span\u003e). Because both test stimuli were novel, successful performance in the generalisation test is difficult to explain as mere recall of one rewarded training image. While bees performed significantly above chance in the generalisation test, their performance was less robust than in the learning test (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eb). Generalisation performance varied between individuals, with a subset achieving exceptional accuracy (\u003cem\u003ee.g.\u003c/em\u003e Bee_3, Bee_6, Bee_4, see Table S7). Performance in the generalisation test \u0026mdash; but not in the learning or inverted tests (Fig. S2) \u0026mdash; was positively correlated with the training level, indicating that extended training facilitates generalisation of learned complex stimuli in bees. Flight speed and angular velocity in the generalisation test did not differ from the learning test, indicating that bees used the same learned flight strategy in the learning test and generalisation test (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ea and b). Our generalisation test degraded featural information in the learned face discrimination but did not alter configural information (Maurer et al. \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2002\u003c/span\u003e, Richler \u003cem\u003eet al.\u003c/em\u003e 2014). Even though honey bees were given relatively short training on a single pair of faces and were not trained to generalise learning to a category, it is noteworthy that bees could still generalise face learning to other faces from the same CFD category.\u003c/p\u003e \u003cp\u003eOur inversion test degraded configural information but not featural information (Maurer et al. \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2002\u003c/span\u003e, Richler \u003cem\u003eet al.\u003c/em\u003e 2014), and in this test bees performed at chance levels. Therefore, we conclude that for bees, face inversion abolished discrimination. In humans and other primates, performance in face recognition tasks is slowed and decreases when faces are inverted, but it is not eliminated (Yin \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e1969\u003c/span\u003e, Diamond \u0026amp; Carey \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e1986\u003c/span\u003e). Clearly, changing the spatial configuration of features had an especially strong effect on bees\u0026rsquo; ability to recognise learned faces. Bees have the capacity to learn configurations as it was shown by Avargues- Weber \u003cem\u003eet al.\u003c/em\u003e (2011) who used abstract stimuli to isolate and test the capacity of bees to learn configurations. Our inverted test offered bees intact featural information but despite this the inversion reduced performance to chance levels. Dyer et al. (\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2005\u003c/span\u003e) also reported that inversion eliminated recognition of a learned rewarding face in honey bees. We noted that in the inversion test bees\u0026rsquo; flight speed and angular velocity remained comparable to that seen in the learning test rather than resembling the slower, more rotational flights of na\u0026iuml;ve individuals (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ea). This dissociation suggests that bees retained and deployed a learnt movement strategy for inspecting face stimuli, but that this strategy was no longer effective when the faces were inverted. The most plausible interpretation is that, during training, bees established an active sensorimotor routine that sampled facial information in a particular spatial and temporal sequence. When the stimuli were turned upside down, the same learnt routine would not match the learnt arrangements; preventing successful matching and disrupting the configural structure on which the recognition depended. Inversion would disrupt a learnt sensorimotor-perceptual loop, and we argue this is why inversion, but not generalisation, had such a dramatic effect on bee face recognition.\u003c/p\u003e \u003cp\u003eOur data support theories of active vision: that bees use learned movements to aid learning and recognition of complex stimuli. Our findings reinforce emerging views that bees use efficient sampling strategies and robust neural representations to solve perceptual challenges normally associated with vertebrates (MaBouDi \u003cem\u003eet al.\u003c/em\u003e 2025\u003csup\u003ea,b\u003c/sup\u003e, Avargu\u0026egrave;s-Weber \u0026amp; Giurfa,2013, Guiraud et al. \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2018\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eInterestingly, the inversion condition elicited the fastest flights of all, a pattern previously described in other forms of perceptual dissonance in bees (\u003cem\u003ee.g.\u003c/em\u003e Guiraud et al. \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2025\u003c/span\u003e), potentially reflecting heightened uncertainty or the use of a default \u0026ldquo;rapid pass\u0026rdquo; strategy when trained relational cues fail. In contrast, generalisation flights closely resembled those from the learned condition, suggesting that bees recognised enough structural similarity to maintain an efficient scanning mode.\u003c/p\u003e \u003cp\u003eTogether, these results support the hypothesis that bees rely on active vision \u0026mdash; combining selective scanning motions with efficient neural coding \u0026mdash; to process visually complex stimuli. The fact that bees can learn, differentiate, and generalise human faces despite coarse spatial resolution indicates that high-level visual recognition does not require large brains, but rather efficient computation paired with informative sampling strategies. This aligns with recent work showing that small networks can replicate key aspects of insect visual decision-making (MaBouDi et al. \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2023\u003c/span\u003e, 2025\u003csup\u003ea,b\u003c/sup\u003e), and suggests that simple, energy-efficient architectures may suffice for tasks traditionally associated with higher cognition.\u003c/p\u003e \u003cp\u003eFinally, our findings have significant implications for the comparative study of vision and for bio-inspired artificial intelligence. Bees\u0026rsquo; ability to extract structural features from complex patterns, even across novel identities, offers a natural model for efficient object recognition systems in robotics \u0026mdash; an area where active vision strategies inspired by insects are increasingly influential (Serres \u003cem\u003eet al.\u003c/em\u003e 2018, MaBouDi \u003cem\u003eet al.\u003c/em\u003e 2025\u003csup\u003ea\u003c/sup\u003e, Ruzzante \u003cem\u003eet al.\u003c/em\u003e 2024). These results highlight the potential for integrating insect-derived principles of sampling, generalisation, and face recognition learning into lightweight computational frameworks.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn conclusion, honey bees successfully learnt to discriminate complex human face stimuli, generalised this learning to novel exemplars, but failed when the learnt faces were inverted as a group. Together with the associated changes in flight trajectories, these results suggest that successful recognition depended not only on what bees learnt about the stimuli, but also on how they actively sampled them. Our findings therefore support the view that efficient active vision, rather than high neural complexity alone, can sustain robust recognition of visually complex patterns.\u003c/p\u003e"},{"header":"Statements and Declarations ","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis research was supported by Australian Research Council (DP230100006), Australian Research Council (DP210100740),\u0026nbsp;Templeton World Charity Foundation TWCF-2020-20539 and the European Union’s Horizon Europe research and innovation programme under the Marie Skłodowska-Curie grant agreement No. 101268923 (FACETS).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eStatement of Animal Ethics\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study involved honey bees (\u003cem\u003eApis mellifera\u003c/em\u003e) and consisted exclusively of non-invasive behavioural experiments. No surgical, immobilising, or otherwise invasive procedures were performed. According to the applicable institutional and local regulations governing research on invertebrates at the time and place at which the study was conducted, formal approval by an animal ethics committee was not required for this work. All animals were handled in accordance with standard good practice for behavioural research on insects, and all efforts were made to minimise disturbance and stress during training and testing.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConflict of Interest Statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNo competing interests declared.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData Availability Statement\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe datasets generated and analysed during the current study are available from the corresponding author upon reasonable request. The code and model resources used for bee tracking are available as described in the Methods section.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eMGG and ABB conceived the study. MGG designed the protocol. MGG and MJ acquired the data. MGG curated the data. ADH performed video tracking. MGG analysed and presented the data. MGG drafted the manuscript. MGG and ABB revised the manuscript. All authors contributed to the finalisation of the manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003e\u003cem\u003eAvargu\u0026egrave;s-Weber A., Portelli G., B\u0026eacute;nard J., Dyer A. G., \u0026amp; Giurfa, M. (2010). \u003c/em\u003e\u003cem\u003eConfigural processing enables discrimination and categorisation of face-like stimuli in honeybees. Journal of Experimental Biology, 213, 593\u0026ndash;601. \u003c/em\u003ehttps://doi.org/10.1242/jeb.039263\u003c/li\u003e\n\u003cli\u003eAvargu\u0026egrave;s-Weber A., Dyer A. G., \u0026amp; Giurfa M. (2011). Conceptualization of above and below relationships by an insect. \u003cem\u003eProceedings of the Royal Society B: Biological Sciences\u003c/em\u003e, 278(1707), 2299\u0026ndash;2306. https://doi.org/10.1098/rspb.2010.1891 \u003c/li\u003e\n\u003cli\u003eAvargu\u0026egrave;s-Weber A., Dyer A. G., Combe M., \u0026amp; Giurfa M. (2012). Simultaneous mastering of two abstract concepts by the miniature brain of bees. Proceedings of the National Academy of Sciences, 109(19), 7481\u0026ndash;7486. https://doi.org/10.1073/pnas.1202576109 \u003c/li\u003e\n\u003cli\u003eAvargu\u0026egrave;s-Weber A., \u0026amp; Giurfa M. (2013). Conceptual learning by miniature brains. \u003cem\u003eProceedings of the Royal Society B: Biological Sciences\u003c/em\u003e, 280(1772), 20131907. https://doi.org/10.1098/rspb.2013.1907 \u003c/li\u003e\n\u003cli\u003eBaird E., Boeddeker N., Ibbotson M.R., \u0026amp; Srinivasan M. V. (2013). A universal strategy for visually guided landing. Proceedings of the National Academy of Sciences, 110, 18686\u0026ndash;18691. https://doi.org/10.1073/pnas.1314311110\u003c/li\u003e\n\u003cli\u003eBarron A.B., Halina M., Klein C. (2023). Transitions in cognitive evolution. Proceedings of the Royal Society B: Biological Sciences, 290, 20230671. https://doi.org/10.1098/rspb.2023.0671 \u003c/li\u003e\n\u003cli\u003eBraun E., Dittmar L., Boeddeker N., Egelhaaf M. (2012). Prototypical components of honeybee homing flight behaviour depend on the visual appearance of objects surrounding the goal. Frontiers in Behavioural Neuroscience, 6, 1. https://doi.org/10.3389/fnbeh.2012.00001\u003c/li\u003e\n\u003cli\u003eCarey S., \u0026amp; Diamond R. (1997). Are faces perceived as configurations more by adults than by children? Visual Cognition, 4, 253\u0026ndash;274. https://doi.org/10.1080/13506289408402302 \u003c/li\u003e\n\u003cli\u003eChittka L., Thomson J.D., \u0026amp; Waser N.M. (1999). Flower constancy, insect psychology, and plant evolution. \u003cem\u003eNaturwissenschaften\u003c/em\u003e, 86, 361\u0026ndash;377. https://doi.org/10.1007/s001140050636 \u003c/li\u003e\n\u003cli\u003eChittka L., Dyer A. G., Bock F., \u0026amp; Dornhaus A. (2003). Bees trade off foraging speed for accuracy. Nature. https://doi.org/10.1038/424388a \u003c/li\u003e\n\u003cli\u003eChittka L. (2017). Bee cognition. Current Biology, 27(19), R1049\u0026ndash;R1053. https://doi.org/10.1016/j.cub.2017.08.008 \u003c/li\u003e\n\u003cli\u003eDittmar L., St\u0026uuml;rzl W., Baird E., Boeddeker N., \u0026amp; Egelhaaf M. (2010). Goal seeking in honeybees: Matching optic flow to distance. Journal of Experimental Biology, 213, 2916\u0026ndash;2926. https://doi.org/10.1242/jeb.041657 \u003c/li\u003e\n\u003cli\u003eDyer A.G., Neumeyer C., Chittka L. (2005). Honeybee (\u003cem\u003eApis mellifera\u003c/em\u003e) vision can discriminate between and recognise images of human faces. Journal of Experimental Biology, 208(24), 4709\u0026ndash;4714. https://doi.org/10.1242/jeb.01929\u003c/li\u003e\n\u003cli\u003eDyer A.G., Rosa M.G.P., \u0026amp; Reser D.H. (2008). Honeybees can recognise images of complex natural scenes for use as potential landmarks. Journal of Experimental Biology, 211(8), 1180\u0026ndash;1186. https://doi.org/10.1242/jeb.016683 \u003c/li\u003e\n\u003cli\u003eDiamond R., \u0026amp; Carey S. (1986). Why faces are and are not special: an effect of expertise. Journal of Experimental Psychology: General, 115, 107\u0026ndash;117. https://doi.org/10.1037/0096-3445.115.2.107 \u003c/li\u003e\n\u003cli\u003eGiurfa M., Hammer M., Stach S., Stollhoff N., M\u0026uuml;ller-Deisig N., Mizyrycki C. (1999). Pattern learning by honeybees: conditioning procedure and recognition strategy. Animal Behaviour, 57(2), 315\u0026ndash;324. https://doi.org/10.1006/anbe.1998.0957\u003c/li\u003e\n\u003cli\u003eGiurfa M. (2003). The amazing mini-brain: lessons from a honey bee. Bee World, 84(1), 5\u0026ndash;18. https://doi.org/10.1080/0005772X.2003.11099566\u003c/li\u003e\n\u003cli\u003eGoulson D. (2000). Are insects flower constant because they use search images to find flowers? Oikos, 88(3), 547\u0026ndash;552. https://doi.org/10.1034/j.1600-0706.2000.880311.x\u003c/li\u003e\n\u003cli\u003eGuiraud M., Roper M., \u0026amp; Chittka L. (2018). High-speed videography reveals how honeybees can turn a spatial concept learning task into a simple discrimination task by stereotyped flight movements and sequential inspection of pattern elements. Frontiers in Psychology, 9, 1347. https://doi.org/10.3389/fpsyg.2018.01347\u003c/li\u003e\n\u003cli\u003eGuiraud M.-G., Roper M., Wolf S., Woodgate J.L., \u0026amp; Chittka L. (2022). Discrimination of edge orientation by bumblebees. PLOS ONE. https://doi.org/10.1371/journal.pone.0263198 \u003c/li\u003e\n\u003cli\u003eGuiraud M.-G., Quinsal-Keel E., Gallo V. \u0026amp; MaBouDi H. (2024). Bumblebee visual learning: simple solutions for complex stimuli. Animal Behaviour. https://doi.org/10.1016/j.anbehav.2024.123070\u003c/li\u003e\n\u003cli\u003eGuiraud M-G., MaBouDi H., Woodgate J.L., Bates O.K., Ramos Rodriguez O., Gallo V., Barron A.B.B. (2025). How bumblebees manage conflicting information seen on arrival and departure from flowers. Animal Cognition, 28(1), 11. https://doi.org/10.1007/s10071-024-01926-x\u003c/li\u003e\n\u003cli\u003eKanwisher N., McDermott J., \u0026amp; Chun M.M. (1997). The fusiform face area: A module in human extrastriate cortex specialized for face perception. Journal of Neuroscience, 17(11), 4302\u0026ndash;4311. https://doi.org/10.1523/JNEUROSCI.17-11-04302.1997 \u003c/li\u003e\n\u003cli\u003eHenderson J.M., Williams C.C. \u0026amp; Falk R.J. (2005). Eye movements are functional during face learning. Memory \u0026amp; Cognition, 33(1), 98\u0026ndash;106. https://doi.org/10.3758/BF03195300\u003c/li\u003e\n\u003cli\u003eKanwisher N. (2000). Domain specificity in face perception. Nature Neuroscience, 3(8), 759\u0026ndash;763. https://doi.org/10.1038/77664 \u003c/li\u003e\n\u003cli\u003eLehrer M. (1994). Spatial vision in the honeybee: the use of different cues in different tasks. Vision Research, 34(18), 2363\u0026ndash;2385. https://doi.org/10.1016/0042-6989(94)90282-8\u003c/li\u003e\n\u003cli\u003eLehrer M., Horridge G.A., Zhang S.W., Gadagkar R. (1995). Shape vision in bees: innate preference for flower-like patterns. Philosophical Transactions of the Royal Society of London B: Biological Sciences, 347(1320), 123\u0026ndash;137. https://doi.org/10.1098/rstb.1995.0017 \u003c/li\u003e\n\u003cli\u003eMaBouDi H., Marshall J.A.R., Dearden N., Barron A.B.B. (2023). How honey bees make fast and accurate decisions. eLife, 12, e86176. https://doi.org/10.7554/eLife.86176\u003c/li\u003e\n\u003cli\u003ea) MaBouDi H., Richter J., Guiraud M.-G., Roper M., Marshall J.A.R. \u0026amp; Chittka, L. (2025). Active vision of bees in a simple pattern discrimination task. eLife. https://doi.org/10.7554/eLife.106332\u003c/li\u003e\n\u003cli\u003eb) MaBouDi H., Roper M., Guiraud M.-G., Juusola M., Chittka L. \u0026amp; Marshall J.A.R. (2025). A neuromorphic model of active vision shows how spatiotemporal encoding in lobula neurons can aid pattern recognition in bees. eLife. https://doi.org/10.7554/eLife.89929 \u003c/li\u003e\n\u003cli\u003eMaurer D., Le Grand R. \u0026amp; Mondloch C.J. (2002). The many faces of configural processing. Trends in Cognitive Sciences, 6(6), 255\u0026ndash;260. https://doi.org/10.1016/S1364-6613(02)01903-4\u003c/li\u003e\n\u003cli\u003eMenzel R. \u0026amp; Giurfa M. (2001). Cognitive architecture of a mini-brain: The honeybee. Trends in Cognitive Sciences, 5(2), 62\u0026ndash;71. https://doi.org/10.1016/S1364-6613(00)01601-6 \u003c/li\u003e\n\u003cli\u003ePascalis O., Kelly D.J. \u0026amp; Caldara R. (2006). What can bees really tell us about the face processing system in humans? Journal of Experimental Biology, 209(16), 3266. https://doi.org/10.1242/jeb.02411 \u003c/li\u003e\n\u003cli\u003ePeterson M.F., Eckstein M.P. (2013). Individual differences in eye movements during face identification reflect observer-specific optimal points of fixation. Psychological Science, 24(7), 1216\u0026ndash;1225. https://doi.org/10.1177/0956797612471684 \u003c/li\u003e\n\u003cli\u003ePerry C.J., Barron A.B.B., Chittka L. (2017). The frontiers of insect cognition. Current Opinion in Behavioural Sciences, 16, 111\u0026ndash;118. https://doi.org/10.1016/j.cobeha.2017.05.011 \u003c/li\u003e\n\u003cli\u003eRavi N., Gabeur V., Hu Y.-T., Hu R., Ryali C., Ma T., Khedr H., R\u0026auml;dle R., Rolland C., Gustafson L., Mintun E., Pan J., Alwala K.V., Carion N., Wu C.-Y., Girshick R., Doll\u0026aacute;r P. \u0026amp; Feichtenhofer C. (2024). SAM 2: Segment Anything in Images and Videos. arXiv. https://doi.org/10.48550/arXiv.2408.00714 \u003c/li\u003e\n\u003cli\u003eRichler J.J. \u0026amp; Gauthier I. (2014). A meta-analysis and review of holistic face processing. Psychological Bulletin, 140(5), 1281\u0026ndash;1302. https://doi.org/10.1037/a0037004 \u003c/li\u003e\n\u003cli\u003eRuzzante D. \u0026amp; Vaes J. (2024). Configural face processing and its influence on the timeline of mentalization. Current Research in Ecological and Social Psychology, 6, 100184. https://doi.org/10.1016/j.cresp.2024.100184 \u003c/li\u003e\n\u003cli\u003eSerres J.R., Viollet S. (2018). Insect-inspired vision for autonomous vehicles. Current Opinion in Insect Science, 30, 46\u0026ndash;51. https://doi.org/10.1016/j.cois.2018.09.005\u003c/li\u003e\n\u003cli\u003eSrinivasan M.V., Lehrer M. (1988). Spatial acuity of honeybee vision and its spectral properties. Journal of Comparative Physiology A, 162, 159\u0026ndash;172. https://doi.org/10.1007/BF00606081\u003c/li\u003e\n\u003cli\u003eSrinivasan M.V. (2010). Honey bees as a model for vision, perception, and cognition. \u003cem\u003eAnnual Review of Entomology\u003c/em\u003e, 55, 267\u0026ndash;284. https://doi.org/10.1146/annurev.ento.010908.164537\u003c/li\u003e\n\u003cli\u003eTsao D.Y. \u0026amp; Livingstone M.S. (2008). Mechanisms of face perception. Annual Review of Neuroscience, 31, 411\u0026ndash;437. https://doi.org/10.1146/annurev.neuro.30.051606.094238 \u003c/li\u003e\n\u003cli\u003eYin R.K. (1969). Looking at upside-down faces. \u003cem\u003eJournal of Experimental Psychology\u003c/em\u003e, 81(1), 141\u0026ndash;145. https://doi.org/10.1037/h0027474 \u003c/li\u003e\n\u003cli\u003eYang C-Y., Huang H-W., Chai W., Jiang Z., Hwang J-N. (2024). SAMURAI: Adapting Segment Anything Model for zero-shot visual tracking with motion-aware memory. arXiv preprint, arXiv:2411.11922. https://doi.org/10.48550/arXiv.2411.11922 \u003c/li\u003e\n\u003cli\u003eZhang S., Srinivasan M.V., Zhu H., Wong J. (2004). Grouping of visual objects by honeybees. Journal of Experimental Biology, 207(19), 3289\u0026ndash;3298. https://doi.org/10.1242/jeb.01155\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"animal-cognition","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"anco","sideBox":"Learn more about [Animal Cognition](http://link.springer.com/journal/10071)","snPcode":"10071","submissionUrl":"https://submission.nature.com/new-submission/10071/3","title":"Animal Cognition","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"Apis mellifera, active vision, visual cognition, insect cognition, 2D tracking, face recognition","lastPublishedDoi":"10.21203/rs.3.rs-9487302/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9487302/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eHoneybees (\u003cem\u003eApis mellifera\u003c/em\u003e) impress with their visual learning abilities despite their miniature brains and relatively coarse vision. A striking demonstration of bee cognition shows that they can rapidly learn to discriminate human faces, despite these stimuli being ecologically irrelevant. To probe the mechanisms underlying this ability, we assessed bees\u0026rsquo; capacity to learn and categorise faces selected from the Chicago Face Database (CFD). Once trained, bees were further tested with stimuli that degraded either the featural or configural information in the face stimuli. The generalisation test examined whether trained bees could successfully discriminate novel faces from the two trained categories: degrading featural information did not degrade configural information in faces. The inverted test offered bees the two faces used in training but rotated by 180\u0026deg;: degrading configural information but not featural information. In all tests, we recorded bees\u0026rsquo; performance and flight path. Bees successfully generalised their learning to novel faces, and performance in the generalisation test correlated with the amount of training. Most trained bees failed the inverted test, however. In all tests, bees\u0026rsquo; flights were faster and more linear than na\u0026iuml;ve bees. We propose that during training bees learned a flight strategy that aided their discrimination of faces, but as a consequence, bee learning of complex stimuli is more sensitive to configural changes that changed the spatial ordering of features than featural changes.\u003c/p\u003e","manuscriptTitle":"Learning complex stimuli with a honey bee brain: contrasting consequences on performance and flight behaviour from degrading configural versus featural information","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-05-04 13:54:43","doi":"10.21203/rs.3.rs-9487302/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"reviewerAgreed","content":"243746207178233754143535841761486991647","date":"2026-05-06T13:24:20+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-04-23T18:27:59+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-04-23T18:15:21+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-04-22T13:24:22+00:00","index":"","fulltext":""},{"type":"submitted","content":"Animal Cognition","date":"2026-04-21T17:00:30+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"animal-cognition","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"anco","sideBox":"Learn more about [Animal Cognition](http://link.springer.com/journal/10071)","snPcode":"10071","submissionUrl":"https://submission.nature.com/new-submission/10071/3","title":"Animal Cognition","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"em","reportingPortfolio":"Springer Hybrid","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"6d9e00a1-aa81-4b4b-ba26-c226c609f8c2","owner":[],"postedDate":"May 4th, 2026","published":true,"recentEditorialEvents":[{"type":"reviewerAgreed","content":"243746207178233754143535841761486991647","date":"2026-05-06T13:24:20+00:00","index":11,"fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2026-05-04T13:54:43+00:00","versionOfRecord":[],"versionCreatedAt":"2026-05-04 13:54:43","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9487302","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9487302","identity":"rs-9487302","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}